Showing preview only (209K chars total). Download the full file or copy to clipboard to get everything.
Repository: facebookresearch/detr
Branch: main
Commit: 29901c51d7fe
Files: 42
Total size: 197.3 KB
Directory structure:
gitextract_f347uvq4/
├── .github/
│ ├── CODE_OF_CONDUCT.md
│ ├── CONTRIBUTING.md
│ └── ISSUE_TEMPLATE/
│ ├── bugs.md
│ ├── questions-help-support.md
│ └── unexpected-problems-bugs.md
├── .gitignore
├── Dockerfile
├── LICENSE
├── README.md
├── d2/
│ ├── README.md
│ ├── configs/
│ │ ├── detr_256_6_6_torchvision.yaml
│ │ └── detr_segm_256_6_6_torchvision.yaml
│ ├── converter.py
│ ├── detr/
│ │ ├── __init__.py
│ │ ├── config.py
│ │ ├── dataset_mapper.py
│ │ └── detr.py
│ └── train_net.py
├── datasets/
│ ├── __init__.py
│ ├── coco.py
│ ├── coco_eval.py
│ ├── coco_panoptic.py
│ ├── panoptic_eval.py
│ └── transforms.py
├── engine.py
├── hubconf.py
├── main.py
├── models/
│ ├── __init__.py
│ ├── backbone.py
│ ├── detr.py
│ ├── matcher.py
│ ├── position_encoding.py
│ ├── segmentation.py
│ └── transformer.py
├── requirements.txt
├── run_with_submitit.py
├── test_all.py
├── tox.ini
└── util/
├── __init__.py
├── box_ops.py
├── misc.py
└── plot_utils.py
================================================
FILE CONTENTS
================================================
================================================
FILE: .github/CODE_OF_CONDUCT.md
================================================
# Code of Conduct
Facebook has adopted a Code of Conduct that we expect project participants to adhere to.
Please read the [full text](https://code.fb.com/codeofconduct/)
so that you can understand what actions will and will not be tolerated.
================================================
FILE: .github/CONTRIBUTING.md
================================================
# Contributing to DETR
We want to make contributing to this project as easy and transparent as
possible.
## Our Development Process
Minor changes and improvements will be released on an ongoing basis. Larger changes (e.g., changesets implementing a new paper) will be released on a more periodic basis.
## Pull Requests
We actively welcome your pull requests.
1. Fork the repo and create your branch from `master`.
2. If you've added code that should be tested, add tests.
3. If you've changed APIs, update the documentation.
4. Ensure the test suite passes.
5. Make sure your code lints.
6. If you haven't already, complete the Contributor License Agreement ("CLA").
## Contributor License Agreement ("CLA")
In order to accept your pull request, we need you to submit a CLA. You only need
to do this once to work on any of Facebook's open source projects.
Complete your CLA here: <https://code.facebook.com/cla>
## Issues
We use GitHub issues to track public bugs. Please ensure your description is
clear and has sufficient instructions to be able to reproduce the issue.
Facebook has a [bounty program](https://www.facebook.com/whitehat/) for the safe
disclosure of security bugs. In those cases, please go through the process
outlined on that page and do not file a public issue.
## Coding Style
* 4 spaces for indentation rather than tabs
* 80 character line length
* PEP8 formatting following [Black](https://black.readthedocs.io/en/stable/)
## License
By contributing to DETR, you agree that your contributions will be licensed
under the LICENSE file in the root directory of this source tree.
================================================
FILE: .github/ISSUE_TEMPLATE/bugs.md
================================================
---
name: "🐛 Bugs"
about: Report bugs in DETR
title: Please read & provide the following
---
## Instructions To Reproduce the 🐛 Bug:
1. what changes you made (`git diff`) or what code you wrote
```
<put diff or code here>
```
2. what exact command you run:
3. what you observed (including __full logs__):
```
<put logs here>
```
4. please simplify the steps as much as possible so they do not require additional resources to
run, such as a private dataset.
## Expected behavior:
If there are no obvious error in "what you observed" provided above,
please tell us the expected behavior.
## Environment:
Provide your environment information using the following command:
```
python -m torch.utils.collect_env
```
================================================
FILE: .github/ISSUE_TEMPLATE/questions-help-support.md
================================================
---
name: "How to do something❓"
about: How to do something using DETR?
---
## ❓ How to do something using DETR
Describe what you want to do, including:
1. what inputs you will provide, if any:
2. what outputs you are expecting:
NOTE:
1. Only general answers are provided.
If you want to ask about "why X did not work", please use the
[Unexpected behaviors](https://github.com/facebookresearch/detr/issues/new/choose) issue template.
2. About how to implement new models / new dataloader / new training logic, etc., check documentation first.
3. We do not answer general machine learning / computer vision questions that are not specific to DETR, such as how a model works, how to improve your training/make it converge, or what algorithm/methods can be used to achieve X.
================================================
FILE: .github/ISSUE_TEMPLATE/unexpected-problems-bugs.md
================================================
---
name: "Unexpected behaviors"
about: Run into unexpected behaviors when using DETR
title: Please read & provide the following
---
If you do not know the root cause of the problem, and wish someone to help you, please
post according to this template:
## Instructions To Reproduce the Issue:
1. what changes you made (`git diff`) or what code you wrote
```
<put diff or code here>
```
2. what exact command you run:
3. what you observed (including __full logs__):
```
<put logs here>
```
4. please simplify the steps as much as possible so they do not require additional resources to
run, such as a private dataset.
## Expected behavior:
If there are no obvious error in "what you observed" provided above,
please tell us the expected behavior.
If you expect the model to converge / work better, note that we do not give suggestions
on how to train a new model.
Only in one of the two conditions we will help with it:
(1) You're unable to reproduce the results in DETR model zoo.
(2) It indicates a DETR bug.
## Environment:
Provide your environment information using the following command:
```
python -m torch.utils.collect_env
```
================================================
FILE: .gitignore
================================================
.nfs*
*.ipynb
*.pyc
.dumbo.json
.DS_Store
.*.swp
*.pth
**/__pycache__/**
.ipynb_checkpoints/
datasets/data/
experiment-*
*.tmp
*.pkl
**/.mypy_cache/*
.mypy_cache/*
not_tracked_dir/
.vscode
================================================
FILE: Dockerfile
================================================
FROM pytorch/pytorch:1.5-cuda10.1-cudnn7-runtime
ENV DEBIAN_FRONTEND=noninteractive
RUN apt-get update -qq && \
apt-get install -y git vim libgtk2.0-dev && \
rm -rf /var/cache/apk/*
RUN pip --no-cache-dir install Cython
COPY requirements.txt /workspace
RUN pip --no-cache-dir install -r /workspace/requirements.txt
================================================
FILE: LICENSE
================================================
Apache License
Version 2.0, January 2004
http://www.apache.org/licenses/
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
1. Definitions.
"License" shall mean the terms and conditions for use, reproduction,
and distribution as defined by Sections 1 through 9 of this document.
"Licensor" shall mean the copyright owner or entity authorized by
the copyright owner that is granting the License.
"Legal Entity" shall mean the union of the acting entity and all
other entities that control, are controlled by, or are under common
control with that entity. For the purposes of this definition,
"control" means (i) the power, direct or indirect, to cause the
direction or management of such entity, whether by contract or
otherwise, or (ii) ownership of fifty percent (50%) or more of the
outstanding shares, or (iii) beneficial ownership of such entity.
"You" (or "Your") shall mean an individual or Legal Entity
exercising permissions granted by this License.
"Source" form shall mean the preferred form for making modifications,
including but not limited to software source code, documentation
source, and configuration files.
"Object" form shall mean any form resulting from mechanical
transformation or translation of a Source form, including but
not limited to compiled object code, generated documentation,
and conversions to other media types.
"Work" shall mean the work of authorship, whether in Source or
Object form, made available under the License, as indicated by a
copyright notice that is included in or attached to the work
(an example is provided in the Appendix below).
"Derivative Works" shall mean any work, whether in Source or Object
form, that is based on (or derived from) the Work and for which the
editorial revisions, annotations, elaborations, or other modifications
represent, as a whole, an original work of authorship. For the purposes
of this License, Derivative Works shall not include works that remain
separable from, or merely link (or bind by name) to the interfaces of,
the Work and Derivative Works thereof.
"Contribution" shall mean any work of authorship, including
the original version of the Work and any modifications or additions
to that Work or Derivative Works thereof, that is intentionally
submitted to Licensor for inclusion in the Work by the copyright owner
or by an individual or Legal Entity authorized to submit on behalf of
the copyright owner. For the purposes of this definition, "submitted"
means any form of electronic, verbal, or written communication sent
to the Licensor or its representatives, including but not limited to
communication on electronic mailing lists, source code control systems,
and issue tracking systems that are managed by, or on behalf of, the
Licensor for the purpose of discussing and improving the Work, but
excluding communication that is conspicuously marked or otherwise
designated in writing by the copyright owner as "Not a Contribution."
"Contributor" shall mean Licensor and any individual or Legal Entity
on behalf of whom a Contribution has been received by Licensor and
subsequently incorporated within the Work.
2. Grant of Copyright License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
copyright license to reproduce, prepare Derivative Works of,
publicly display, publicly perform, sublicense, and distribute the
Work and such Derivative Works in Source or Object form.
3. Grant of Patent License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
(except as stated in this section) patent license to make, have made,
use, offer to sell, sell, import, and otherwise transfer the Work,
where such license applies only to those patent claims licensable
by such Contributor that are necessarily infringed by their
Contribution(s) alone or by combination of their Contribution(s)
with the Work to which such Contribution(s) was submitted. If You
institute patent litigation against any entity (including a
cross-claim or counterclaim in a lawsuit) alleging that the Work
or a Contribution incorporated within the Work constitutes direct
or contributory patent infringement, then any patent licenses
granted to You under this License for that Work shall terminate
as of the date such litigation is filed.
4. Redistribution. You may reproduce and distribute copies of the
Work or Derivative Works thereof in any medium, with or without
modifications, and in Source or Object form, provided that You
meet the following conditions:
(a) You must give any other recipients of the Work or
Derivative Works a copy of this License; and
(b) You must cause any modified files to carry prominent notices
stating that You changed the files; and
(c) You must retain, in the Source form of any Derivative Works
that You distribute, all copyright, patent, trademark, and
attribution notices from the Source form of the Work,
excluding those notices that do not pertain to any part of
the Derivative Works; and
(d) If the Work includes a "NOTICE" text file as part of its
distribution, then any Derivative Works that You distribute must
include a readable copy of the attribution notices contained
within such NOTICE file, excluding those notices that do not
pertain to any part of the Derivative Works, in at least one
of the following places: within a NOTICE text file distributed
as part of the Derivative Works; within the Source form or
documentation, if provided along with the Derivative Works; or,
within a display generated by the Derivative Works, if and
wherever such third-party notices normally appear. The contents
of the NOTICE file are for informational purposes only and
do not modify the License. You may add Your own attribution
notices within Derivative Works that You distribute, alongside
or as an addendum to the NOTICE text from the Work, provided
that such additional attribution notices cannot be construed
as modifying the License.
You may add Your own copyright statement to Your modifications and
may provide additional or different license terms and conditions
for use, reproduction, or distribution of Your modifications, or
for any such Derivative Works as a whole, provided Your use,
reproduction, and distribution of the Work otherwise complies with
the conditions stated in this License.
5. Submission of Contributions. Unless You explicitly state otherwise,
any Contribution intentionally submitted for inclusion in the Work
by You to the Licensor shall be under the terms and conditions of
this License, without any additional terms or conditions.
Notwithstanding the above, nothing herein shall supersede or modify
the terms of any separate license agreement you may have executed
with Licensor regarding such Contributions.
6. Trademarks. This License does not grant permission to use the trade
names, trademarks, service marks, or product names of the Licensor,
except as required for reasonable and customary use in describing the
origin of the Work and reproducing the content of the NOTICE file.
7. Disclaimer of Warranty. Unless required by applicable law or
agreed to in writing, Licensor provides the Work (and each
Contributor provides its Contributions) on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
implied, including, without limitation, any warranties or conditions
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
PARTICULAR PURPOSE. You are solely responsible for determining the
appropriateness of using or redistributing the Work and assume any
risks associated with Your exercise of permissions under this License.
8. Limitation of Liability. In no event and under no legal theory,
whether in tort (including negligence), contract, or otherwise,
unless required by applicable law (such as deliberate and grossly
negligent acts) or agreed to in writing, shall any Contributor be
liable to You for damages, including any direct, indirect, special,
incidental, or consequential damages of any character arising as a
result of this License or out of the use or inability to use the
Work (including but not limited to damages for loss of goodwill,
work stoppage, computer failure or malfunction, or any and all
other commercial damages or losses), even if such Contributor
has been advised of the possibility of such damages.
9. Accepting Warranty or Additional Liability. While redistributing
the Work or Derivative Works thereof, You may choose to offer,
and charge a fee for, acceptance of support, warranty, indemnity,
or other liability obligations and/or rights consistent with this
License. However, in accepting such obligations, You may act only
on Your own behalf and on Your sole responsibility, not on behalf
of any other Contributor, and only if You agree to indemnify,
defend, and hold each Contributor harmless for any liability
incurred by, or claims asserted against, such Contributor by reason
of your accepting any such warranty or additional liability.
END OF TERMS AND CONDITIONS
APPENDIX: How to apply the Apache License to your work.
To apply the Apache License to your work, attach the following
boilerplate notice, with the fields enclosed by brackets "[]"
replaced with your own identifying information. (Don't include
the brackets!) The text should be enclosed in the appropriate
comment syntax for the file format. We also recommend that a
file or class name and description of purpose be included on the
same "printed page" as the copyright notice for easier
identification within third-party archives.
Copyright 2020 - present, Facebook, Inc
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
================================================
FILE: README.md
================================================
**DE⫶TR**: End-to-End Object Detection with Transformers
========
[](https://opensource.fb.com/support-ukraine)
PyTorch training code and pretrained models for **DETR** (**DE**tection **TR**ansformer).
We replace the full complex hand-crafted object detection pipeline with a Transformer, and match Faster R-CNN with a ResNet-50, obtaining **42 AP** on COCO using half the computation power (FLOPs) and the same number of parameters. Inference in 50 lines of PyTorch.
**What it is**. Unlike traditional computer vision techniques, DETR approaches object detection as a direct set prediction problem. It consists of a set-based global loss, which forces unique predictions via bipartite matching, and a Transformer encoder-decoder architecture.
Given a fixed small set of learned object queries, DETR reasons about the relations of the objects and the global image context to directly output the final set of predictions in parallel. Due to this parallel nature, DETR is very fast and efficient.
**About the code**. We believe that object detection should not be more difficult than classification,
and should not require complex libraries for training and inference.
DETR is very simple to implement and experiment with, and we provide a
[standalone Colab Notebook](https://colab.research.google.com/github/facebookresearch/detr/blob/colab/notebooks/detr_demo.ipynb)
showing how to do inference with DETR in only a few lines of PyTorch code.
Training code follows this idea - it is not a library,
but simply a [main.py](main.py) importing model and criterion
definitions with standard training loops.
Additionnally, we provide a Detectron2 wrapper in the d2/ folder. See the readme there for more information.
For details see [End-to-End Object Detection with Transformers](https://ai.facebook.com/research/publications/end-to-end-object-detection-with-transformers) by Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko.
See our [blog post](https://ai.facebook.com/blog/end-to-end-object-detection-with-transformers/) to learn more about end to end object detection with transformers.
# Model Zoo
We provide baseline DETR and DETR-DC5 models, and plan to include more in future.
AP is computed on COCO 2017 val5k, and inference time is over the first 100 val5k COCO images,
with torchscript transformer.
<table>
<thead>
<tr style="text-align: right;">
<th></th>
<th>name</th>
<th>backbone</th>
<th>schedule</th>
<th>inf_time</th>
<th>box AP</th>
<th>url</th>
<th>size</th>
</tr>
</thead>
<tbody>
<tr>
<th>0</th>
<td>DETR</td>
<td>R50</td>
<td>500</td>
<td>0.036</td>
<td>42.0</td>
<td><a href="https://dl.fbaipublicfiles.com/detr/detr-r50-e632da11.pth">model</a> | <a href="https://dl.fbaipublicfiles.com/detr/logs/detr-r50_log.txt">logs</a></td>
<td>159Mb</td>
</tr>
<tr>
<th>1</th>
<td>DETR-DC5</td>
<td>R50</td>
<td>500</td>
<td>0.083</td>
<td>43.3</td>
<td><a href="https://dl.fbaipublicfiles.com/detr/detr-r50-dc5-f0fb7ef5.pth">model</a> | <a href="https://dl.fbaipublicfiles.com/detr/logs/detr-r50-dc5_log.txt">logs</a></td>
<td>159Mb</td>
</tr>
<tr>
<th>2</th>
<td>DETR</td>
<td>R101</td>
<td>500</td>
<td>0.050</td>
<td>43.5</td>
<td><a href="https://dl.fbaipublicfiles.com/detr/detr-r101-2c7b67e5.pth">model</a> | <a href="https://dl.fbaipublicfiles.com/detr/logs/detr-r101_log.txt">logs</a></td>
<td>232Mb</td>
</tr>
<tr>
<th>3</th>
<td>DETR-DC5</td>
<td>R101</td>
<td>500</td>
<td>0.097</td>
<td>44.9</td>
<td><a href="https://dl.fbaipublicfiles.com/detr/detr-r101-dc5-a2e86def.pth">model</a> | <a href="https://dl.fbaipublicfiles.com/detr/logs/detr-r101-dc5_log.txt">logs</a></td>
<td>232Mb</td>
</tr>
</tbody>
</table>
COCO val5k evaluation results can be found in this [gist](https://gist.github.com/szagoruyko/9c9ebb8455610958f7deaa27845d7918).
The models are also available via torch hub,
to load DETR R50 with pretrained weights simply do:
```python
model = torch.hub.load('facebookresearch/detr:main', 'detr_resnet50', pretrained=True)
```
COCO panoptic val5k models:
<table>
<thead>
<tr style="text-align: right;">
<th></th>
<th>name</th>
<th>backbone</th>
<th>box AP</th>
<th>segm AP</th>
<th>PQ</th>
<th>url</th>
<th>size</th>
</tr>
</thead>
<tbody>
<tr>
<th>0</th>
<td>DETR</td>
<td>R50</td>
<td>38.8</td>
<td>31.1</td>
<td>43.4</td>
<td><a href="https://dl.fbaipublicfiles.com/detr/detr-r50-panoptic-00ce5173.pth">download</a></td>
<td>165Mb</td>
</tr>
<tr>
<th>1</th>
<td>DETR-DC5</td>
<td>R50</td>
<td>40.2</td>
<td>31.9</td>
<td>44.6</td>
<td><a href="https://dl.fbaipublicfiles.com/detr/detr-r50-dc5-panoptic-da08f1b1.pth">download</a></td>
<td>165Mb</td>
</tr>
<tr>
<th>2</th>
<td>DETR</td>
<td>R101</td>
<td>40.1</td>
<td>33</td>
<td>45.1</td>
<td><a href="https://dl.fbaipublicfiles.com/detr/detr-r101-panoptic-40021d53.pth">download</a></td>
<td>237Mb</td>
</tr>
</tbody>
</table>
Checkout our [panoptic colab](https://colab.research.google.com/github/facebookresearch/detr/blob/colab/notebooks/DETR_panoptic.ipynb)
to see how to use and visualize DETR's panoptic segmentation prediction.
# Notebooks
We provide a few notebooks in colab to help you get a grasp on DETR:
* [DETR's hands on Colab Notebook](https://colab.research.google.com/github/facebookresearch/detr/blob/colab/notebooks/detr_attention.ipynb): Shows how to load a model from hub, generate predictions, then visualize the attention of the model (similar to the figures of the paper)
* [Standalone Colab Notebook](https://colab.research.google.com/github/facebookresearch/detr/blob/colab/notebooks/detr_demo.ipynb): In this notebook, we demonstrate how to implement a simplified version of DETR from the grounds up in 50 lines of Python, then visualize the predictions. It is a good starting point if you want to gain better understanding the architecture and poke around before diving in the codebase.
* [Panoptic Colab Notebook](https://colab.research.google.com/github/facebookresearch/detr/blob/colab/notebooks/DETR_panoptic.ipynb): Demonstrates how to use DETR for panoptic segmentation and plot the predictions.
# Usage - Object detection
There are no extra compiled components in DETR and package dependencies are minimal,
so the code is very simple to use. We provide instructions how to install dependencies via conda.
First, clone the repository locally:
```
git clone https://github.com/facebookresearch/detr.git
```
Then, install PyTorch 1.5+ and torchvision 0.6+:
```
conda install -c pytorch pytorch torchvision
```
Install pycocotools (for evaluation on COCO) and scipy (for training):
```
conda install cython scipy
pip install -U 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'
```
That's it, should be good to train and evaluate detection models.
(optional) to work with panoptic install panopticapi:
```
pip install git+https://github.com/cocodataset/panopticapi.git
```
## Data preparation
Download and extract COCO 2017 train and val images with annotations from
[http://cocodataset.org](http://cocodataset.org/#download).
We expect the directory structure to be the following:
```
path/to/coco/
annotations/ # annotation json files
train2017/ # train images
val2017/ # val images
```
## Training
To train baseline DETR on a single node with 8 gpus for 300 epochs run:
```
python -m torch.distributed.launch --nproc_per_node=8 --use_env main.py --coco_path /path/to/coco
```
A single epoch takes 28 minutes, so 300 epoch training
takes around 6 days on a single machine with 8 V100 cards.
To ease reproduction of our results we provide
[results and training logs](https://gist.github.com/szagoruyko/b4c3b2c3627294fc369b899987385a3f)
for 150 epoch schedule (3 days on a single machine), achieving 39.5/60.3 AP/AP50.
We train DETR with AdamW setting learning rate in the transformer to 1e-4 and 1e-5 in the backbone.
Horizontal flips, scales and crops are used for augmentation.
Images are rescaled to have min size 800 and max size 1333.
The transformer is trained with dropout of 0.1, and the whole model is trained with grad clip of 0.1.
## Evaluation
To evaluate DETR R50 on COCO val5k with a single GPU run:
```
python main.py --batch_size 2 --no_aux_loss --eval --resume https://dl.fbaipublicfiles.com/detr/detr-r50-e632da11.pth --coco_path /path/to/coco
```
We provide results for all DETR detection models in this
[gist](https://gist.github.com/szagoruyko/9c9ebb8455610958f7deaa27845d7918).
Note that numbers vary depending on batch size (number of images) per GPU.
Non-DC5 models were trained with batch size 2, and DC5 with 1,
so DC5 models show a significant drop in AP if evaluated with more
than 1 image per GPU.
## Multinode training
Distributed training is available via Slurm and [submitit](https://github.com/facebookincubator/submitit):
```
pip install submitit
```
Train baseline DETR-6-6 model on 4 nodes for 300 epochs:
```
python run_with_submitit.py --timeout 3000 --coco_path /path/to/coco
```
# Usage - Segmentation
We show that it is relatively straightforward to extend DETR to predict segmentation masks. We mainly demonstrate strong panoptic segmentation results.
## Data preparation
For panoptic segmentation, you need the panoptic annotations additionally to the coco dataset (see above for the coco dataset). You need to download and extract the [annotations](http://images.cocodataset.org/annotations/panoptic_annotations_trainval2017.zip).
We expect the directory structure to be the following:
```
path/to/coco_panoptic/
annotations/ # annotation json files
panoptic_train2017/ # train panoptic annotations
panoptic_val2017/ # val panoptic annotations
```
## Training
We recommend training segmentation in two stages: first train DETR to detect all the boxes, and then train the segmentation head.
For panoptic segmentation, DETR must learn to detect boxes for both stuff and things classes. You can train it on a single node with 8 gpus for 300 epochs with:
```
python -m torch.distributed.launch --nproc_per_node=8 --use_env main.py --coco_path /path/to/coco --coco_panoptic_path /path/to/coco_panoptic --dataset_file coco_panoptic --output_dir /output/path/box_model
```
For instance segmentation, you can simply train a normal box model (or used a pre-trained one we provide).
Once you have a box model checkpoint, you need to freeze it, and train the segmentation head in isolation.
For panoptic segmentation you can train on a single node with 8 gpus for 25 epochs:
```
python -m torch.distributed.launch --nproc_per_node=8 --use_env main.py --masks --epochs 25 --lr_drop 15 --coco_path /path/to/coco --coco_panoptic_path /path/to/coco_panoptic --dataset_file coco_panoptic --frozen_weights /output/path/box_model/checkpoint.pth --output_dir /output/path/segm_model
```
For instance segmentation only, simply remove the `dataset_file` and `coco_panoptic_path` arguments from the above command line.
# License
DETR is released under the Apache 2.0 license. Please see the [LICENSE](LICENSE) file for more information.
# Contributing
We actively welcome your pull requests! Please see [CONTRIBUTING.md](.github/CONTRIBUTING.md) and [CODE_OF_CONDUCT.md](.github/CODE_OF_CONDUCT.md) for more info.
================================================
FILE: d2/README.md
================================================
Detectron2 wrapper for DETR
=======
We provide a Detectron2 wrapper for DETR, thus providing a way to better integrate it in the existing detection ecosystem. It can be used for example to easily leverage datasets or backbones provided in Detectron2.
This wrapper currently supports only box detection, and is intended to be as close as possible to the original implementation, and we checked that it indeed match the results. Some notable facts and caveats:
- The data augmentation matches DETR's original data augmentation. This required patching the RandomCrop augmentation from Detectron2, so you'll need a version from the master branch from June 24th 2020 or more recent.
- To match DETR's original backbone initialization, we use the weights of a ResNet50 trained on imagenet using torchvision. This network uses a different pixel mean and std than most of the backbones available in Detectron2 by default, so extra care must be taken when switching to another one. Note that no other torchvision models are available in Detectron2 as of now, though it may change in the future.
- The gradient clipping mode is "full_model", which is not the default in Detectron2.
# Usage
To install Detectron2, please follow the [official installation instructions](https://github.com/facebookresearch/detectron2/blob/master/INSTALL.md).
## Evaluating a model
For convenience, we provide a conversion script to convert models trained by the main DETR training loop into the format of this wrapper. To download and convert the main Resnet50 model, simply do:
```
python converter.py --source_model https://dl.fbaipublicfiles.com/detr/detr-r50-e632da11.pth --output_model converted_model.pth
```
You can then evaluate it using:
```
python train_net.py --eval-only --config configs/detr_256_6_6_torchvision.yaml MODEL.WEIGHTS "converted_model.pth"
```
## Training
To train DETR on a single node with 8 gpus, simply use:
```
python train_net.py --config configs/detr_256_6_6_torchvision.yaml --num-gpus 8
```
To fine-tune DETR for instance segmentation on a single node with 8 gpus, simply use:
```
python train_net.py --config configs/detr_segm_256_6_6_torchvision.yaml --num-gpus 8 MODEL.DETR.FROZEN_WEIGHTS <model_path>
```
================================================
FILE: d2/configs/detr_256_6_6_torchvision.yaml
================================================
MODEL:
META_ARCHITECTURE: "Detr"
WEIGHTS: "detectron2://ImageNetPretrained/torchvision/R-50.pkl"
PIXEL_MEAN: [123.675, 116.280, 103.530]
PIXEL_STD: [58.395, 57.120, 57.375]
MASK_ON: False
RESNETS:
DEPTH: 50
STRIDE_IN_1X1: False
OUT_FEATURES: ["res2", "res3", "res4", "res5"]
DETR:
GIOU_WEIGHT: 2.0
L1_WEIGHT: 5.0
NUM_OBJECT_QUERIES: 100
DATASETS:
TRAIN: ("coco_2017_train",)
TEST: ("coco_2017_val",)
SOLVER:
IMS_PER_BATCH: 64
BASE_LR: 0.0001
STEPS: (369600,)
MAX_ITER: 554400
WARMUP_FACTOR: 1.0
WARMUP_ITERS: 10
WEIGHT_DECAY: 0.0001
OPTIMIZER: "ADAMW"
BACKBONE_MULTIPLIER: 0.1
CLIP_GRADIENTS:
ENABLED: True
CLIP_TYPE: "full_model"
CLIP_VALUE: 0.01
NORM_TYPE: 2.0
INPUT:
MIN_SIZE_TRAIN: (480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800)
CROP:
ENABLED: True
TYPE: "absolute_range"
SIZE: (384, 600)
FORMAT: "RGB"
TEST:
EVAL_PERIOD: 4000
DATALOADER:
FILTER_EMPTY_ANNOTATIONS: False
NUM_WORKERS: 4
VERSION: 2
================================================
FILE: d2/configs/detr_segm_256_6_6_torchvision.yaml
================================================
MODEL:
META_ARCHITECTURE: "Detr"
# WEIGHTS: "detectron2://ImageNetPretrained/torchvision/R-50.pkl"
PIXEL_MEAN: [123.675, 116.280, 103.530]
PIXEL_STD: [58.395, 57.120, 57.375]
MASK_ON: True
RESNETS:
DEPTH: 50
STRIDE_IN_1X1: False
OUT_FEATURES: ["res2", "res3", "res4", "res5"]
DETR:
GIOU_WEIGHT: 2.0
L1_WEIGHT: 5.0
NUM_OBJECT_QUERIES: 100
FROZEN_WEIGHTS: ''
DATASETS:
TRAIN: ("coco_2017_train",)
TEST: ("coco_2017_val",)
SOLVER:
IMS_PER_BATCH: 64
BASE_LR: 0.0001
STEPS: (55440,)
MAX_ITER: 92400
WARMUP_FACTOR: 1.0
WARMUP_ITERS: 10
WEIGHT_DECAY: 0.0001
OPTIMIZER: "ADAMW"
BACKBONE_MULTIPLIER: 0.1
CLIP_GRADIENTS:
ENABLED: True
CLIP_TYPE: "full_model"
CLIP_VALUE: 0.01
NORM_TYPE: 2.0
INPUT:
MIN_SIZE_TRAIN: (480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800)
CROP:
ENABLED: True
TYPE: "absolute_range"
SIZE: (384, 600)
FORMAT: "RGB"
TEST:
EVAL_PERIOD: 4000
DATALOADER:
FILTER_EMPTY_ANNOTATIONS: False
NUM_WORKERS: 4
VERSION: 2
================================================
FILE: d2/converter.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
"""
Helper script to convert models trained with the main version of DETR to be used with the Detectron2 version.
"""
import json
import argparse
import numpy as np
import torch
def parse_args():
parser = argparse.ArgumentParser("D2 model converter")
parser.add_argument("--source_model", default="", type=str, help="Path or url to the DETR model to convert")
parser.add_argument("--output_model", default="", type=str, help="Path where to save the converted model")
return parser.parse_args()
def main():
args = parse_args()
# D2 expects contiguous classes, so we need to remap the 92 classes from DETR
# fmt: off
coco_idx = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25,
27, 28, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 46, 47, 48, 49, 50, 51,
52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 67, 70, 72, 73, 74, 75, 76, 77,
78, 79, 80, 81, 82, 84, 85, 86, 87, 88, 89, 90, 91]
# fmt: on
coco_idx = np.array(coco_idx)
if args.source_model.startswith("https"):
checkpoint = torch.hub.load_state_dict_from_url(args.source_model, map_location="cpu", check_hash=True)
else:
checkpoint = torch.load(args.source_model, map_location="cpu")
model_to_convert = checkpoint["model"]
model_converted = {}
for k in model_to_convert.keys():
old_k = k
if "backbone" in k:
k = k.replace("backbone.0.body.", "")
if "layer" not in k:
k = "stem." + k
for t in [1, 2, 3, 4]:
k = k.replace(f"layer{t}", f"res{t + 1}")
for t in [1, 2, 3]:
k = k.replace(f"bn{t}", f"conv{t}.norm")
k = k.replace("downsample.0", "shortcut")
k = k.replace("downsample.1", "shortcut.norm")
k = "backbone.0.backbone." + k
k = "detr." + k
print(old_k, "->", k)
if "class_embed" in old_k:
v = model_to_convert[old_k].detach()
if v.shape[0] == 92:
shape_old = v.shape
model_converted[k] = v[coco_idx]
print("Head conversion: changing shape from {} to {}".format(shape_old, model_converted[k].shape))
continue
model_converted[k] = model_to_convert[old_k].detach()
model_to_save = {"model": model_converted}
torch.save(model_to_save, args.output_model)
if __name__ == "__main__":
main()
================================================
FILE: d2/detr/__init__.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
from .config import add_detr_config
from .detr import Detr
from .dataset_mapper import DetrDatasetMapper
================================================
FILE: d2/detr/config.py
================================================
# -*- coding: utf-8 -*-
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
from detectron2.config import CfgNode as CN
def add_detr_config(cfg):
"""
Add config for DETR.
"""
cfg.MODEL.DETR = CN()
cfg.MODEL.DETR.NUM_CLASSES = 80
# For Segmentation
cfg.MODEL.DETR.FROZEN_WEIGHTS = ''
# LOSS
cfg.MODEL.DETR.GIOU_WEIGHT = 2.0
cfg.MODEL.DETR.L1_WEIGHT = 5.0
cfg.MODEL.DETR.DEEP_SUPERVISION = True
cfg.MODEL.DETR.NO_OBJECT_WEIGHT = 0.1
# TRANSFORMER
cfg.MODEL.DETR.NHEADS = 8
cfg.MODEL.DETR.DROPOUT = 0.1
cfg.MODEL.DETR.DIM_FEEDFORWARD = 2048
cfg.MODEL.DETR.ENC_LAYERS = 6
cfg.MODEL.DETR.DEC_LAYERS = 6
cfg.MODEL.DETR.PRE_NORM = False
cfg.MODEL.DETR.HIDDEN_DIM = 256
cfg.MODEL.DETR.NUM_OBJECT_QUERIES = 100
cfg.SOLVER.OPTIMIZER = "ADAMW"
cfg.SOLVER.BACKBONE_MULTIPLIER = 0.1
================================================
FILE: d2/detr/dataset_mapper.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
import copy
import logging
import numpy as np
import torch
from detectron2.data import detection_utils as utils
from detectron2.data import transforms as T
from detectron2.data.transforms import TransformGen
__all__ = ["DetrDatasetMapper"]
def build_transform_gen(cfg, is_train):
"""
Create a list of :class:`TransformGen` from config.
Returns:
list[TransformGen]
"""
if is_train:
min_size = cfg.INPUT.MIN_SIZE_TRAIN
max_size = cfg.INPUT.MAX_SIZE_TRAIN
sample_style = cfg.INPUT.MIN_SIZE_TRAIN_SAMPLING
else:
min_size = cfg.INPUT.MIN_SIZE_TEST
max_size = cfg.INPUT.MAX_SIZE_TEST
sample_style = "choice"
if sample_style == "range":
assert len(min_size) == 2, "more than 2 ({}) min_size(s) are provided for ranges".format(len(min_size))
logger = logging.getLogger(__name__)
tfm_gens = []
if is_train:
tfm_gens.append(T.RandomFlip())
tfm_gens.append(T.ResizeShortestEdge(min_size, max_size, sample_style))
if is_train:
logger.info("TransformGens used in training: " + str(tfm_gens))
return tfm_gens
class DetrDatasetMapper:
"""
A callable which takes a dataset dict in Detectron2 Dataset format,
and map it into a format used by DETR.
The callable currently does the following:
1. Read the image from "file_name"
2. Applies geometric transforms to the image and annotation
3. Find and applies suitable cropping to the image and annotation
4. Prepare image and annotation to Tensors
"""
def __init__(self, cfg, is_train=True):
if cfg.INPUT.CROP.ENABLED and is_train:
self.crop_gen = [
T.ResizeShortestEdge([400, 500, 600], sample_style="choice"),
T.RandomCrop(cfg.INPUT.CROP.TYPE, cfg.INPUT.CROP.SIZE),
]
else:
self.crop_gen = None
self.mask_on = cfg.MODEL.MASK_ON
self.tfm_gens = build_transform_gen(cfg, is_train)
logging.getLogger(__name__).info(
"Full TransformGens used in training: {}, crop: {}".format(str(self.tfm_gens), str(self.crop_gen))
)
self.img_format = cfg.INPUT.FORMAT
self.is_train = is_train
def __call__(self, dataset_dict):
"""
Args:
dataset_dict (dict): Metadata of one image, in Detectron2 Dataset format.
Returns:
dict: a format that builtin models in detectron2 accept
"""
dataset_dict = copy.deepcopy(dataset_dict) # it will be modified by code below
image = utils.read_image(dataset_dict["file_name"], format=self.img_format)
utils.check_image_size(dataset_dict, image)
if self.crop_gen is None:
image, transforms = T.apply_transform_gens(self.tfm_gens, image)
else:
if np.random.rand() > 0.5:
image, transforms = T.apply_transform_gens(self.tfm_gens, image)
else:
image, transforms = T.apply_transform_gens(
self.tfm_gens[:-1] + self.crop_gen + self.tfm_gens[-1:], image
)
image_shape = image.shape[:2] # h, w
# Pytorch's dataloader is efficient on torch.Tensor due to shared-memory,
# but not efficient on large generic data structures due to the use of pickle & mp.Queue.
# Therefore it's important to use torch.Tensor.
dataset_dict["image"] = torch.as_tensor(np.ascontiguousarray(image.transpose(2, 0, 1)))
if not self.is_train:
# USER: Modify this if you want to keep them for some reason.
dataset_dict.pop("annotations", None)
return dataset_dict
if "annotations" in dataset_dict:
# USER: Modify this if you want to keep them for some reason.
for anno in dataset_dict["annotations"]:
if not self.mask_on:
anno.pop("segmentation", None)
anno.pop("keypoints", None)
# USER: Implement additional transformations if you have other types of data
annos = [
utils.transform_instance_annotations(obj, transforms, image_shape)
for obj in dataset_dict.pop("annotations")
if obj.get("iscrowd", 0) == 0
]
instances = utils.annotations_to_instances(annos, image_shape)
dataset_dict["instances"] = utils.filter_empty_instances(instances)
return dataset_dict
================================================
FILE: d2/detr/detr.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
import logging
import math
from typing import List
import numpy as np
import torch
import torch.distributed as dist
import torch.nn.functional as F
from scipy.optimize import linear_sum_assignment
from torch import nn
from detectron2.layers import ShapeSpec
from detectron2.modeling import META_ARCH_REGISTRY, build_backbone, detector_postprocess
from detectron2.structures import Boxes, ImageList, Instances, BitMasks, PolygonMasks
from detectron2.utils.logger import log_first_n
from fvcore.nn import giou_loss, smooth_l1_loss
from models.backbone import Joiner
from models.detr import DETR, SetCriterion
from models.matcher import HungarianMatcher
from models.position_encoding import PositionEmbeddingSine
from models.transformer import Transformer
from models.segmentation import DETRsegm, PostProcessPanoptic, PostProcessSegm
from util.box_ops import box_cxcywh_to_xyxy, box_xyxy_to_cxcywh
from util.misc import NestedTensor
from datasets.coco import convert_coco_poly_to_mask
__all__ = ["Detr"]
class MaskedBackbone(nn.Module):
""" This is a thin wrapper around D2's backbone to provide padding masking"""
def __init__(self, cfg):
super().__init__()
self.backbone = build_backbone(cfg)
backbone_shape = self.backbone.output_shape()
self.feature_strides = [backbone_shape[f].stride for f in backbone_shape.keys()]
self.num_channels = backbone_shape[list(backbone_shape.keys())[-1]].channels
def forward(self, images):
features = self.backbone(images.tensor)
masks = self.mask_out_padding(
[features_per_level.shape for features_per_level in features.values()],
images.image_sizes,
images.tensor.device,
)
assert len(features) == len(masks)
for i, k in enumerate(features.keys()):
features[k] = NestedTensor(features[k], masks[i])
return features
def mask_out_padding(self, feature_shapes, image_sizes, device):
masks = []
assert len(feature_shapes) == len(self.feature_strides)
for idx, shape in enumerate(feature_shapes):
N, _, H, W = shape
masks_per_feature_level = torch.ones((N, H, W), dtype=torch.bool, device=device)
for img_idx, (h, w) in enumerate(image_sizes):
masks_per_feature_level[
img_idx,
: int(np.ceil(float(h) / self.feature_strides[idx])),
: int(np.ceil(float(w) / self.feature_strides[idx])),
] = 0
masks.append(masks_per_feature_level)
return masks
@META_ARCH_REGISTRY.register()
class Detr(nn.Module):
"""
Implement Detr
"""
def __init__(self, cfg):
super().__init__()
self.device = torch.device(cfg.MODEL.DEVICE)
self.num_classes = cfg.MODEL.DETR.NUM_CLASSES
self.mask_on = cfg.MODEL.MASK_ON
hidden_dim = cfg.MODEL.DETR.HIDDEN_DIM
num_queries = cfg.MODEL.DETR.NUM_OBJECT_QUERIES
# Transformer parameters:
nheads = cfg.MODEL.DETR.NHEADS
dropout = cfg.MODEL.DETR.DROPOUT
dim_feedforward = cfg.MODEL.DETR.DIM_FEEDFORWARD
enc_layers = cfg.MODEL.DETR.ENC_LAYERS
dec_layers = cfg.MODEL.DETR.DEC_LAYERS
pre_norm = cfg.MODEL.DETR.PRE_NORM
# Loss parameters:
giou_weight = cfg.MODEL.DETR.GIOU_WEIGHT
l1_weight = cfg.MODEL.DETR.L1_WEIGHT
deep_supervision = cfg.MODEL.DETR.DEEP_SUPERVISION
no_object_weight = cfg.MODEL.DETR.NO_OBJECT_WEIGHT
N_steps = hidden_dim // 2
d2_backbone = MaskedBackbone(cfg)
backbone = Joiner(d2_backbone, PositionEmbeddingSine(N_steps, normalize=True))
backbone.num_channels = d2_backbone.num_channels
transformer = Transformer(
d_model=hidden_dim,
dropout=dropout,
nhead=nheads,
dim_feedforward=dim_feedforward,
num_encoder_layers=enc_layers,
num_decoder_layers=dec_layers,
normalize_before=pre_norm,
return_intermediate_dec=deep_supervision,
)
self.detr = DETR(
backbone, transformer, num_classes=self.num_classes, num_queries=num_queries, aux_loss=deep_supervision
)
if self.mask_on:
frozen_weights = cfg.MODEL.DETR.FROZEN_WEIGHTS
if frozen_weights != '':
print("LOAD pre-trained weights")
weight = torch.load(frozen_weights, map_location=lambda storage, loc: storage)['model']
new_weight = {}
for k, v in weight.items():
if 'detr.' in k:
new_weight[k.replace('detr.', '')] = v
else:
print(f"Skipping loading weight {k} from frozen model")
del weight
self.detr.load_state_dict(new_weight)
del new_weight
self.detr = DETRsegm(self.detr, freeze_detr=(frozen_weights != ''))
self.seg_postprocess = PostProcessSegm
self.detr.to(self.device)
# building criterion
matcher = HungarianMatcher(cost_class=1, cost_bbox=l1_weight, cost_giou=giou_weight)
weight_dict = {"loss_ce": 1, "loss_bbox": l1_weight}
weight_dict["loss_giou"] = giou_weight
if deep_supervision:
aux_weight_dict = {}
for i in range(dec_layers - 1):
aux_weight_dict.update({k + f"_{i}": v for k, v in weight_dict.items()})
weight_dict.update(aux_weight_dict)
losses = ["labels", "boxes", "cardinality"]
if self.mask_on:
losses += ["masks"]
self.criterion = SetCriterion(
self.num_classes, matcher=matcher, weight_dict=weight_dict, eos_coef=no_object_weight, losses=losses,
)
self.criterion.to(self.device)
pixel_mean = torch.Tensor(cfg.MODEL.PIXEL_MEAN).to(self.device).view(3, 1, 1)
pixel_std = torch.Tensor(cfg.MODEL.PIXEL_STD).to(self.device).view(3, 1, 1)
self.normalizer = lambda x: (x - pixel_mean) / pixel_std
self.to(self.device)
def forward(self, batched_inputs):
"""
Args:
batched_inputs: a list, batched outputs of :class:`DatasetMapper` .
Each item in the list contains the inputs for one image.
For now, each item in the list is a dict that contains:
* image: Tensor, image in (C, H, W) format.
* instances: Instances
Other information that's included in the original dicts, such as:
* "height", "width" (int): the output resolution of the model, used in inference.
See :meth:`postprocess` for details.
Returns:
dict[str: Tensor]:
mapping from a named loss to a tensor storing the loss. Used during training only.
"""
images = self.preprocess_image(batched_inputs)
output = self.detr(images)
if self.training:
gt_instances = [x["instances"].to(self.device) for x in batched_inputs]
targets = self.prepare_targets(gt_instances)
loss_dict = self.criterion(output, targets)
weight_dict = self.criterion.weight_dict
for k in loss_dict.keys():
if k in weight_dict:
loss_dict[k] *= weight_dict[k]
return loss_dict
else:
box_cls = output["pred_logits"]
box_pred = output["pred_boxes"]
mask_pred = output["pred_masks"] if self.mask_on else None
results = self.inference(box_cls, box_pred, mask_pred, images.image_sizes)
processed_results = []
for results_per_image, input_per_image, image_size in zip(results, batched_inputs, images.image_sizes):
height = input_per_image.get("height", image_size[0])
width = input_per_image.get("width", image_size[1])
r = detector_postprocess(results_per_image, height, width)
processed_results.append({"instances": r})
return processed_results
def prepare_targets(self, targets):
new_targets = []
for targets_per_image in targets:
h, w = targets_per_image.image_size
image_size_xyxy = torch.as_tensor([w, h, w, h], dtype=torch.float, device=self.device)
gt_classes = targets_per_image.gt_classes
gt_boxes = targets_per_image.gt_boxes.tensor / image_size_xyxy
gt_boxes = box_xyxy_to_cxcywh(gt_boxes)
new_targets.append({"labels": gt_classes, "boxes": gt_boxes})
if self.mask_on and hasattr(targets_per_image, 'gt_masks'):
gt_masks = targets_per_image.gt_masks
gt_masks = convert_coco_poly_to_mask(gt_masks.polygons, h, w)
new_targets[-1].update({'masks': gt_masks})
return new_targets
def inference(self, box_cls, box_pred, mask_pred, image_sizes):
"""
Arguments:
box_cls (Tensor): tensor of shape (batch_size, num_queries, K).
The tensor predicts the classification probability for each query.
box_pred (Tensor): tensors of shape (batch_size, num_queries, 4).
The tensor predicts 4-vector (x,y,w,h) box
regression values for every queryx
image_sizes (List[torch.Size]): the input image sizes
Returns:
results (List[Instances]): a list of #images elements.
"""
assert len(box_cls) == len(image_sizes)
results = []
# For each box we assign the best class or the second best if the best on is `no_object`.
scores, labels = F.softmax(box_cls, dim=-1)[:, :, :-1].max(-1)
for i, (scores_per_image, labels_per_image, box_pred_per_image, image_size) in enumerate(zip(
scores, labels, box_pred, image_sizes
)):
result = Instances(image_size)
result.pred_boxes = Boxes(box_cxcywh_to_xyxy(box_pred_per_image))
result.pred_boxes.scale(scale_x=image_size[1], scale_y=image_size[0])
if self.mask_on:
mask = F.interpolate(mask_pred[i].unsqueeze(0), size=image_size, mode='bilinear', align_corners=False)
mask = mask[0].sigmoid() > 0.5
B, N, H, W = mask_pred.shape
mask = BitMasks(mask.cpu()).crop_and_resize(result.pred_boxes.tensor.cpu(), 32)
result.pred_masks = mask.unsqueeze(1).to(mask_pred[0].device)
result.scores = scores_per_image
result.pred_classes = labels_per_image
results.append(result)
return results
def preprocess_image(self, batched_inputs):
"""
Normalize, pad and batch the input images.
"""
images = [self.normalizer(x["image"].to(self.device)) for x in batched_inputs]
images = ImageList.from_tensors(images)
return images
================================================
FILE: d2/train_net.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
"""
DETR Training Script.
This script is a simplified version of the training script in detectron2/tools.
"""
import os
import sys
import itertools
# fmt: off
sys.path.insert(1, os.path.join(sys.path[0], '..'))
# fmt: on
import time
from typing import Any, Dict, List, Set
import torch
import detectron2.utils.comm as comm
from d2.detr import DetrDatasetMapper, add_detr_config
from detectron2.checkpoint import DetectionCheckpointer
from detectron2.config import get_cfg
from detectron2.data import MetadataCatalog, build_detection_train_loader
from detectron2.engine import DefaultTrainer, default_argument_parser, default_setup, launch
from detectron2.evaluation import COCOEvaluator, verify_results
from detectron2.solver.build import maybe_add_gradient_clipping
class Trainer(DefaultTrainer):
"""
Extension of the Trainer class adapted to DETR.
"""
@classmethod
def build_evaluator(cls, cfg, dataset_name, output_folder=None):
"""
Create evaluator(s) for a given dataset.
This uses the special metadata "evaluator_type" associated with each builtin dataset.
For your own dataset, you can simply create an evaluator manually in your
script and do not have to worry about the hacky if-else logic here.
"""
if output_folder is None:
output_folder = os.path.join(cfg.OUTPUT_DIR, "inference")
return COCOEvaluator(dataset_name, cfg, True, output_folder)
@classmethod
def build_train_loader(cls, cfg):
if "Detr" == cfg.MODEL.META_ARCHITECTURE:
mapper = DetrDatasetMapper(cfg, True)
else:
mapper = None
return build_detection_train_loader(cfg, mapper=mapper)
@classmethod
def build_optimizer(cls, cfg, model):
params: List[Dict[str, Any]] = []
memo: Set[torch.nn.parameter.Parameter] = set()
for key, value in model.named_parameters(recurse=True):
if not value.requires_grad:
continue
# Avoid duplicating parameters
if value in memo:
continue
memo.add(value)
lr = cfg.SOLVER.BASE_LR
weight_decay = cfg.SOLVER.WEIGHT_DECAY
if "backbone" in key:
lr = lr * cfg.SOLVER.BACKBONE_MULTIPLIER
params += [{"params": [value], "lr": lr, "weight_decay": weight_decay}]
def maybe_add_full_model_gradient_clipping(optim): # optim: the optimizer class
# detectron2 doesn't have full model gradient clipping now
clip_norm_val = cfg.SOLVER.CLIP_GRADIENTS.CLIP_VALUE
enable = (
cfg.SOLVER.CLIP_GRADIENTS.ENABLED
and cfg.SOLVER.CLIP_GRADIENTS.CLIP_TYPE == "full_model"
and clip_norm_val > 0.0
)
class FullModelGradientClippingOptimizer(optim):
def step(self, closure=None):
all_params = itertools.chain(*[x["params"] for x in self.param_groups])
torch.nn.utils.clip_grad_norm_(all_params, clip_norm_val)
super().step(closure=closure)
return FullModelGradientClippingOptimizer if enable else optim
optimizer_type = cfg.SOLVER.OPTIMIZER
if optimizer_type == "SGD":
optimizer = maybe_add_full_model_gradient_clipping(torch.optim.SGD)(
params, cfg.SOLVER.BASE_LR, momentum=cfg.SOLVER.MOMENTUM
)
elif optimizer_type == "ADAMW":
optimizer = maybe_add_full_model_gradient_clipping(torch.optim.AdamW)(
params, cfg.SOLVER.BASE_LR
)
else:
raise NotImplementedError(f"no optimizer type {optimizer_type}")
if not cfg.SOLVER.CLIP_GRADIENTS.CLIP_TYPE == "full_model":
optimizer = maybe_add_gradient_clipping(cfg, optimizer)
return optimizer
def setup(args):
"""
Create configs and perform basic setups.
"""
cfg = get_cfg()
add_detr_config(cfg)
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.freeze()
default_setup(cfg, args)
return cfg
def main(args):
cfg = setup(args)
if args.eval_only:
model = Trainer.build_model(cfg)
DetectionCheckpointer(model, save_dir=cfg.OUTPUT_DIR).resume_or_load(cfg.MODEL.WEIGHTS, resume=args.resume)
res = Trainer.test(cfg, model)
if comm.is_main_process():
verify_results(cfg, res)
return res
trainer = Trainer(cfg)
trainer.resume_or_load(resume=args.resume)
return trainer.train()
if __name__ == "__main__":
args = default_argument_parser().parse_args()
print("Command Line Args:", args)
launch(
main,
args.num_gpus,
num_machines=args.num_machines,
machine_rank=args.machine_rank,
dist_url=args.dist_url,
args=(args,),
)
================================================
FILE: datasets/__init__.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
import torch.utils.data
import torchvision
from .coco import build as build_coco
def get_coco_api_from_dataset(dataset):
for _ in range(10):
# if isinstance(dataset, torchvision.datasets.CocoDetection):
# break
if isinstance(dataset, torch.utils.data.Subset):
dataset = dataset.dataset
if isinstance(dataset, torchvision.datasets.CocoDetection):
return dataset.coco
def build_dataset(image_set, args):
if args.dataset_file == 'coco':
return build_coco(image_set, args)
if args.dataset_file == 'coco_panoptic':
# to avoid making panopticapi required for coco
from .coco_panoptic import build as build_coco_panoptic
return build_coco_panoptic(image_set, args)
raise ValueError(f'dataset {args.dataset_file} not supported')
================================================
FILE: datasets/coco.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
"""
COCO dataset which returns image_id for evaluation.
Mostly copy-paste from https://github.com/pytorch/vision/blob/13b35ff/references/detection/coco_utils.py
"""
from pathlib import Path
import torch
import torch.utils.data
import torchvision
from pycocotools import mask as coco_mask
import datasets.transforms as T
class CocoDetection(torchvision.datasets.CocoDetection):
def __init__(self, img_folder, ann_file, transforms, return_masks):
super(CocoDetection, self).__init__(img_folder, ann_file)
self._transforms = transforms
self.prepare = ConvertCocoPolysToMask(return_masks)
def __getitem__(self, idx):
img, target = super(CocoDetection, self).__getitem__(idx)
image_id = self.ids[idx]
target = {'image_id': image_id, 'annotations': target}
img, target = self.prepare(img, target)
if self._transforms is not None:
img, target = self._transforms(img, target)
return img, target
def convert_coco_poly_to_mask(segmentations, height, width):
masks = []
for polygons in segmentations:
rles = coco_mask.frPyObjects(polygons, height, width)
mask = coco_mask.decode(rles)
if len(mask.shape) < 3:
mask = mask[..., None]
mask = torch.as_tensor(mask, dtype=torch.uint8)
mask = mask.any(dim=2)
masks.append(mask)
if masks:
masks = torch.stack(masks, dim=0)
else:
masks = torch.zeros((0, height, width), dtype=torch.uint8)
return masks
class ConvertCocoPolysToMask(object):
def __init__(self, return_masks=False):
self.return_masks = return_masks
def __call__(self, image, target):
w, h = image.size
image_id = target["image_id"]
image_id = torch.tensor([image_id])
anno = target["annotations"]
anno = [obj for obj in anno if 'iscrowd' not in obj or obj['iscrowd'] == 0]
boxes = [obj["bbox"] for obj in anno]
# guard against no boxes via resizing
boxes = torch.as_tensor(boxes, dtype=torch.float32).reshape(-1, 4)
boxes[:, 2:] += boxes[:, :2]
boxes[:, 0::2].clamp_(min=0, max=w)
boxes[:, 1::2].clamp_(min=0, max=h)
classes = [obj["category_id"] for obj in anno]
classes = torch.tensor(classes, dtype=torch.int64)
if self.return_masks:
segmentations = [obj["segmentation"] for obj in anno]
masks = convert_coco_poly_to_mask(segmentations, h, w)
keypoints = None
if anno and "keypoints" in anno[0]:
keypoints = [obj["keypoints"] for obj in anno]
keypoints = torch.as_tensor(keypoints, dtype=torch.float32)
num_keypoints = keypoints.shape[0]
if num_keypoints:
keypoints = keypoints.view(num_keypoints, -1, 3)
keep = (boxes[:, 3] > boxes[:, 1]) & (boxes[:, 2] > boxes[:, 0])
boxes = boxes[keep]
classes = classes[keep]
if self.return_masks:
masks = masks[keep]
if keypoints is not None:
keypoints = keypoints[keep]
target = {}
target["boxes"] = boxes
target["labels"] = classes
if self.return_masks:
target["masks"] = masks
target["image_id"] = image_id
if keypoints is not None:
target["keypoints"] = keypoints
# for conversion to coco api
area = torch.tensor([obj["area"] for obj in anno])
iscrowd = torch.tensor([obj["iscrowd"] if "iscrowd" in obj else 0 for obj in anno])
target["area"] = area[keep]
target["iscrowd"] = iscrowd[keep]
target["orig_size"] = torch.as_tensor([int(h), int(w)])
target["size"] = torch.as_tensor([int(h), int(w)])
return image, target
def make_coco_transforms(image_set):
normalize = T.Compose([
T.ToTensor(),
T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
scales = [480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800]
if image_set == 'train':
return T.Compose([
T.RandomHorizontalFlip(),
T.RandomSelect(
T.RandomResize(scales, max_size=1333),
T.Compose([
T.RandomResize([400, 500, 600]),
T.RandomSizeCrop(384, 600),
T.RandomResize(scales, max_size=1333),
])
),
normalize,
])
if image_set == 'val':
return T.Compose([
T.RandomResize([800], max_size=1333),
normalize,
])
raise ValueError(f'unknown {image_set}')
def build(image_set, args):
root = Path(args.coco_path)
assert root.exists(), f'provided COCO path {root} does not exist'
mode = 'instances'
PATHS = {
"train": (root / "train2017", root / "annotations" / f'{mode}_train2017.json'),
"val": (root / "val2017", root / "annotations" / f'{mode}_val2017.json'),
}
img_folder, ann_file = PATHS[image_set]
dataset = CocoDetection(img_folder, ann_file, transforms=make_coco_transforms(image_set), return_masks=args.masks)
return dataset
================================================
FILE: datasets/coco_eval.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
"""
COCO evaluator that works in distributed mode.
Mostly copy-paste from https://github.com/pytorch/vision/blob/edfd5a7/references/detection/coco_eval.py
The difference is that there is less copy-pasting from pycocotools
in the end of the file, as python3 can suppress prints with contextlib
"""
import os
import contextlib
import copy
import numpy as np
import torch
from pycocotools.cocoeval import COCOeval
from pycocotools.coco import COCO
import pycocotools.mask as mask_util
from util.misc import all_gather
class CocoEvaluator(object):
def __init__(self, coco_gt, iou_types):
assert isinstance(iou_types, (list, tuple))
coco_gt = copy.deepcopy(coco_gt)
self.coco_gt = coco_gt
self.iou_types = iou_types
self.coco_eval = {}
for iou_type in iou_types:
self.coco_eval[iou_type] = COCOeval(coco_gt, iouType=iou_type)
self.img_ids = []
self.eval_imgs = {k: [] for k in iou_types}
def update(self, predictions):
img_ids = list(np.unique(list(predictions.keys())))
self.img_ids.extend(img_ids)
for iou_type in self.iou_types:
results = self.prepare(predictions, iou_type)
# suppress pycocotools prints
with open(os.devnull, 'w') as devnull:
with contextlib.redirect_stdout(devnull):
coco_dt = COCO.loadRes(self.coco_gt, results) if results else COCO()
coco_eval = self.coco_eval[iou_type]
coco_eval.cocoDt = coco_dt
coco_eval.params.imgIds = list(img_ids)
img_ids, eval_imgs = evaluate(coco_eval)
self.eval_imgs[iou_type].append(eval_imgs)
def synchronize_between_processes(self):
for iou_type in self.iou_types:
self.eval_imgs[iou_type] = np.concatenate(self.eval_imgs[iou_type], 2)
create_common_coco_eval(self.coco_eval[iou_type], self.img_ids, self.eval_imgs[iou_type])
def accumulate(self):
for coco_eval in self.coco_eval.values():
coco_eval.accumulate()
def summarize(self):
for iou_type, coco_eval in self.coco_eval.items():
print("IoU metric: {}".format(iou_type))
coco_eval.summarize()
def prepare(self, predictions, iou_type):
if iou_type == "bbox":
return self.prepare_for_coco_detection(predictions)
elif iou_type == "segm":
return self.prepare_for_coco_segmentation(predictions)
elif iou_type == "keypoints":
return self.prepare_for_coco_keypoint(predictions)
else:
raise ValueError("Unknown iou type {}".format(iou_type))
def prepare_for_coco_detection(self, predictions):
coco_results = []
for original_id, prediction in predictions.items():
if len(prediction) == 0:
continue
boxes = prediction["boxes"]
boxes = convert_to_xywh(boxes).tolist()
scores = prediction["scores"].tolist()
labels = prediction["labels"].tolist()
coco_results.extend(
[
{
"image_id": original_id,
"category_id": labels[k],
"bbox": box,
"score": scores[k],
}
for k, box in enumerate(boxes)
]
)
return coco_results
def prepare_for_coco_segmentation(self, predictions):
coco_results = []
for original_id, prediction in predictions.items():
if len(prediction) == 0:
continue
scores = prediction["scores"]
labels = prediction["labels"]
masks = prediction["masks"]
masks = masks > 0.5
scores = prediction["scores"].tolist()
labels = prediction["labels"].tolist()
rles = [
mask_util.encode(np.array(mask[0, :, :, np.newaxis], dtype=np.uint8, order="F"))[0]
for mask in masks
]
for rle in rles:
rle["counts"] = rle["counts"].decode("utf-8")
coco_results.extend(
[
{
"image_id": original_id,
"category_id": labels[k],
"segmentation": rle,
"score": scores[k],
}
for k, rle in enumerate(rles)
]
)
return coco_results
def prepare_for_coco_keypoint(self, predictions):
coco_results = []
for original_id, prediction in predictions.items():
if len(prediction) == 0:
continue
boxes = prediction["boxes"]
boxes = convert_to_xywh(boxes).tolist()
scores = prediction["scores"].tolist()
labels = prediction["labels"].tolist()
keypoints = prediction["keypoints"]
keypoints = keypoints.flatten(start_dim=1).tolist()
coco_results.extend(
[
{
"image_id": original_id,
"category_id": labels[k],
'keypoints': keypoint,
"score": scores[k],
}
for k, keypoint in enumerate(keypoints)
]
)
return coco_results
def convert_to_xywh(boxes):
xmin, ymin, xmax, ymax = boxes.unbind(1)
return torch.stack((xmin, ymin, xmax - xmin, ymax - ymin), dim=1)
def merge(img_ids, eval_imgs):
all_img_ids = all_gather(img_ids)
all_eval_imgs = all_gather(eval_imgs)
merged_img_ids = []
for p in all_img_ids:
merged_img_ids.extend(p)
merged_eval_imgs = []
for p in all_eval_imgs:
merged_eval_imgs.append(p)
merged_img_ids = np.array(merged_img_ids)
merged_eval_imgs = np.concatenate(merged_eval_imgs, 2)
# keep only unique (and in sorted order) images
merged_img_ids, idx = np.unique(merged_img_ids, return_index=True)
merged_eval_imgs = merged_eval_imgs[..., idx]
return merged_img_ids, merged_eval_imgs
def create_common_coco_eval(coco_eval, img_ids, eval_imgs):
img_ids, eval_imgs = merge(img_ids, eval_imgs)
img_ids = list(img_ids)
eval_imgs = list(eval_imgs.flatten())
coco_eval.evalImgs = eval_imgs
coco_eval.params.imgIds = img_ids
coco_eval._paramsEval = copy.deepcopy(coco_eval.params)
#################################################################
# From pycocotools, just removed the prints and fixed
# a Python3 bug about unicode not defined
#################################################################
def evaluate(self):
'''
Run per image evaluation on given images and store results (a list of dict) in self.evalImgs
:return: None
'''
# tic = time.time()
# print('Running per image evaluation...')
p = self.params
# add backward compatibility if useSegm is specified in params
if p.useSegm is not None:
p.iouType = 'segm' if p.useSegm == 1 else 'bbox'
print('useSegm (deprecated) is not None. Running {} evaluation'.format(p.iouType))
# print('Evaluate annotation type *{}*'.format(p.iouType))
p.imgIds = list(np.unique(p.imgIds))
if p.useCats:
p.catIds = list(np.unique(p.catIds))
p.maxDets = sorted(p.maxDets)
self.params = p
self._prepare()
# loop through images, area range, max detection number
catIds = p.catIds if p.useCats else [-1]
if p.iouType == 'segm' or p.iouType == 'bbox':
computeIoU = self.computeIoU
elif p.iouType == 'keypoints':
computeIoU = self.computeOks
self.ious = {
(imgId, catId): computeIoU(imgId, catId)
for imgId in p.imgIds
for catId in catIds}
evaluateImg = self.evaluateImg
maxDet = p.maxDets[-1]
evalImgs = [
evaluateImg(imgId, catId, areaRng, maxDet)
for catId in catIds
for areaRng in p.areaRng
for imgId in p.imgIds
]
# this is NOT in the pycocotools code, but could be done outside
evalImgs = np.asarray(evalImgs).reshape(len(catIds), len(p.areaRng), len(p.imgIds))
self._paramsEval = copy.deepcopy(self.params)
# toc = time.time()
# print('DONE (t={:0.2f}s).'.format(toc-tic))
return p.imgIds, evalImgs
#################################################################
# end of straight copy from pycocotools, just removing the prints
#################################################################
================================================
FILE: datasets/coco_panoptic.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
import json
from pathlib import Path
import numpy as np
import torch
from PIL import Image
from panopticapi.utils import rgb2id
from util.box_ops import masks_to_boxes
from .coco import make_coco_transforms
class CocoPanoptic:
def __init__(self, img_folder, ann_folder, ann_file, transforms=None, return_masks=True):
with open(ann_file, 'r') as f:
self.coco = json.load(f)
# sort 'images' field so that they are aligned with 'annotations'
# i.e., in alphabetical order
self.coco['images'] = sorted(self.coco['images'], key=lambda x: x['id'])
# sanity check
if "annotations" in self.coco:
for img, ann in zip(self.coco['images'], self.coco['annotations']):
assert img['file_name'][:-4] == ann['file_name'][:-4]
self.img_folder = img_folder
self.ann_folder = ann_folder
self.ann_file = ann_file
self.transforms = transforms
self.return_masks = return_masks
def __getitem__(self, idx):
ann_info = self.coco['annotations'][idx] if "annotations" in self.coco else self.coco['images'][idx]
img_path = Path(self.img_folder) / ann_info['file_name'].replace('.png', '.jpg')
ann_path = Path(self.ann_folder) / ann_info['file_name']
img = Image.open(img_path).convert('RGB')
w, h = img.size
if "segments_info" in ann_info:
masks = np.asarray(Image.open(ann_path), dtype=np.uint32)
masks = rgb2id(masks)
ids = np.array([ann['id'] for ann in ann_info['segments_info']])
masks = masks == ids[:, None, None]
masks = torch.as_tensor(masks, dtype=torch.uint8)
labels = torch.tensor([ann['category_id'] for ann in ann_info['segments_info']], dtype=torch.int64)
target = {}
target['image_id'] = torch.tensor([ann_info['image_id'] if "image_id" in ann_info else ann_info["id"]])
if self.return_masks:
target['masks'] = masks
target['labels'] = labels
target["boxes"] = masks_to_boxes(masks)
target['size'] = torch.as_tensor([int(h), int(w)])
target['orig_size'] = torch.as_tensor([int(h), int(w)])
if "segments_info" in ann_info:
for name in ['iscrowd', 'area']:
target[name] = torch.tensor([ann[name] for ann in ann_info['segments_info']])
if self.transforms is not None:
img, target = self.transforms(img, target)
return img, target
def __len__(self):
return len(self.coco['images'])
def get_height_and_width(self, idx):
img_info = self.coco['images'][idx]
height = img_info['height']
width = img_info['width']
return height, width
def build(image_set, args):
img_folder_root = Path(args.coco_path)
ann_folder_root = Path(args.coco_panoptic_path)
assert img_folder_root.exists(), f'provided COCO path {img_folder_root} does not exist'
assert ann_folder_root.exists(), f'provided COCO path {ann_folder_root} does not exist'
mode = 'panoptic'
PATHS = {
"train": ("train2017", Path("annotations") / f'{mode}_train2017.json'),
"val": ("val2017", Path("annotations") / f'{mode}_val2017.json'),
}
img_folder, ann_file = PATHS[image_set]
img_folder_path = img_folder_root / img_folder
ann_folder = ann_folder_root / f'{mode}_{img_folder}'
ann_file = ann_folder_root / ann_file
dataset = CocoPanoptic(img_folder_path, ann_folder, ann_file,
transforms=make_coco_transforms(image_set), return_masks=args.masks)
return dataset
================================================
FILE: datasets/panoptic_eval.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
import json
import os
import util.misc as utils
try:
from panopticapi.evaluation import pq_compute
except ImportError:
pass
class PanopticEvaluator(object):
def __init__(self, ann_file, ann_folder, output_dir="panoptic_eval"):
self.gt_json = ann_file
self.gt_folder = ann_folder
if utils.is_main_process():
if not os.path.exists(output_dir):
os.mkdir(output_dir)
self.output_dir = output_dir
self.predictions = []
def update(self, predictions):
for p in predictions:
with open(os.path.join(self.output_dir, p["file_name"]), "wb") as f:
f.write(p.pop("png_string"))
self.predictions += predictions
def synchronize_between_processes(self):
all_predictions = utils.all_gather(self.predictions)
merged_predictions = []
for p in all_predictions:
merged_predictions += p
self.predictions = merged_predictions
def summarize(self):
if utils.is_main_process():
json_data = {"annotations": self.predictions}
predictions_json = os.path.join(self.output_dir, "predictions.json")
with open(predictions_json, "w") as f:
f.write(json.dumps(json_data))
return pq_compute(self.gt_json, predictions_json, gt_folder=self.gt_folder, pred_folder=self.output_dir)
return None
================================================
FILE: datasets/transforms.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
"""
Transforms and data augmentation for both image + bbox.
"""
import random
import PIL
import torch
import torchvision.transforms as T
import torchvision.transforms.functional as F
from util.box_ops import box_xyxy_to_cxcywh
from util.misc import interpolate
def crop(image, target, region):
cropped_image = F.crop(image, *region)
target = target.copy()
i, j, h, w = region
# should we do something wrt the original size?
target["size"] = torch.tensor([h, w])
fields = ["labels", "area", "iscrowd"]
if "boxes" in target:
boxes = target["boxes"]
max_size = torch.as_tensor([w, h], dtype=torch.float32)
cropped_boxes = boxes - torch.as_tensor([j, i, j, i])
cropped_boxes = torch.min(cropped_boxes.reshape(-1, 2, 2), max_size)
cropped_boxes = cropped_boxes.clamp(min=0)
area = (cropped_boxes[:, 1, :] - cropped_boxes[:, 0, :]).prod(dim=1)
target["boxes"] = cropped_boxes.reshape(-1, 4)
target["area"] = area
fields.append("boxes")
if "masks" in target:
# FIXME should we update the area here if there are no boxes?
target['masks'] = target['masks'][:, i:i + h, j:j + w]
fields.append("masks")
# remove elements for which the boxes or masks that have zero area
if "boxes" in target or "masks" in target:
# favor boxes selection when defining which elements to keep
# this is compatible with previous implementation
if "boxes" in target:
cropped_boxes = target['boxes'].reshape(-1, 2, 2)
keep = torch.all(cropped_boxes[:, 1, :] > cropped_boxes[:, 0, :], dim=1)
else:
keep = target['masks'].flatten(1).any(1)
for field in fields:
target[field] = target[field][keep]
return cropped_image, target
def hflip(image, target):
flipped_image = F.hflip(image)
w, h = image.size
target = target.copy()
if "boxes" in target:
boxes = target["boxes"]
boxes = boxes[:, [2, 1, 0, 3]] * torch.as_tensor([-1, 1, -1, 1]) + torch.as_tensor([w, 0, w, 0])
target["boxes"] = boxes
if "masks" in target:
target['masks'] = target['masks'].flip(-1)
return flipped_image, target
def resize(image, target, size, max_size=None):
# size can be min_size (scalar) or (w, h) tuple
def get_size_with_aspect_ratio(image_size, size, max_size=None):
w, h = image_size
if max_size is not None:
min_original_size = float(min((w, h)))
max_original_size = float(max((w, h)))
if max_original_size / min_original_size * size > max_size:
size = int(round(max_size * min_original_size / max_original_size))
if (w <= h and w == size) or (h <= w and h == size):
return (h, w)
if w < h:
ow = size
oh = int(size * h / w)
else:
oh = size
ow = int(size * w / h)
return (oh, ow)
def get_size(image_size, size, max_size=None):
if isinstance(size, (list, tuple)):
return size[::-1]
else:
return get_size_with_aspect_ratio(image_size, size, max_size)
size = get_size(image.size, size, max_size)
rescaled_image = F.resize(image, size)
if target is None:
return rescaled_image, None
ratios = tuple(float(s) / float(s_orig) for s, s_orig in zip(rescaled_image.size, image.size))
ratio_width, ratio_height = ratios
target = target.copy()
if "boxes" in target:
boxes = target["boxes"]
scaled_boxes = boxes * torch.as_tensor([ratio_width, ratio_height, ratio_width, ratio_height])
target["boxes"] = scaled_boxes
if "area" in target:
area = target["area"]
scaled_area = area * (ratio_width * ratio_height)
target["area"] = scaled_area
h, w = size
target["size"] = torch.tensor([h, w])
if "masks" in target:
target['masks'] = interpolate(
target['masks'][:, None].float(), size, mode="nearest")[:, 0] > 0.5
return rescaled_image, target
def pad(image, target, padding):
# assumes that we only pad on the bottom right corners
padded_image = F.pad(image, (0, 0, padding[0], padding[1]))
if target is None:
return padded_image, None
target = target.copy()
# should we do something wrt the original size?
target["size"] = torch.tensor(padded_image.size[::-1])
if "masks" in target:
target['masks'] = torch.nn.functional.pad(target['masks'], (0, padding[0], 0, padding[1]))
return padded_image, target
class RandomCrop(object):
def __init__(self, size):
self.size = size
def __call__(self, img, target):
region = T.RandomCrop.get_params(img, self.size)
return crop(img, target, region)
class RandomSizeCrop(object):
def __init__(self, min_size: int, max_size: int):
self.min_size = min_size
self.max_size = max_size
def __call__(self, img: PIL.Image.Image, target: dict):
w = random.randint(self.min_size, min(img.width, self.max_size))
h = random.randint(self.min_size, min(img.height, self.max_size))
region = T.RandomCrop.get_params(img, [h, w])
return crop(img, target, region)
class CenterCrop(object):
def __init__(self, size):
self.size = size
def __call__(self, img, target):
image_width, image_height = img.size
crop_height, crop_width = self.size
crop_top = int(round((image_height - crop_height) / 2.))
crop_left = int(round((image_width - crop_width) / 2.))
return crop(img, target, (crop_top, crop_left, crop_height, crop_width))
class RandomHorizontalFlip(object):
def __init__(self, p=0.5):
self.p = p
def __call__(self, img, target):
if random.random() < self.p:
return hflip(img, target)
return img, target
class RandomResize(object):
def __init__(self, sizes, max_size=None):
assert isinstance(sizes, (list, tuple))
self.sizes = sizes
self.max_size = max_size
def __call__(self, img, target=None):
size = random.choice(self.sizes)
return resize(img, target, size, self.max_size)
class RandomPad(object):
def __init__(self, max_pad):
self.max_pad = max_pad
def __call__(self, img, target):
pad_x = random.randint(0, self.max_pad)
pad_y = random.randint(0, self.max_pad)
return pad(img, target, (pad_x, pad_y))
class RandomSelect(object):
"""
Randomly selects between transforms1 and transforms2,
with probability p for transforms1 and (1 - p) for transforms2
"""
def __init__(self, transforms1, transforms2, p=0.5):
self.transforms1 = transforms1
self.transforms2 = transforms2
self.p = p
def __call__(self, img, target):
if random.random() < self.p:
return self.transforms1(img, target)
return self.transforms2(img, target)
class ToTensor(object):
def __call__(self, img, target):
return F.to_tensor(img), target
class RandomErasing(object):
def __init__(self, *args, **kwargs):
self.eraser = T.RandomErasing(*args, **kwargs)
def __call__(self, img, target):
return self.eraser(img), target
class Normalize(object):
def __init__(self, mean, std):
self.mean = mean
self.std = std
def __call__(self, image, target=None):
image = F.normalize(image, mean=self.mean, std=self.std)
if target is None:
return image, None
target = target.copy()
h, w = image.shape[-2:]
if "boxes" in target:
boxes = target["boxes"]
boxes = box_xyxy_to_cxcywh(boxes)
boxes = boxes / torch.tensor([w, h, w, h], dtype=torch.float32)
target["boxes"] = boxes
return image, target
class Compose(object):
def __init__(self, transforms):
self.transforms = transforms
def __call__(self, image, target):
for t in self.transforms:
image, target = t(image, target)
return image, target
def __repr__(self):
format_string = self.__class__.__name__ + "("
for t in self.transforms:
format_string += "\n"
format_string += " {0}".format(t)
format_string += "\n)"
return format_string
================================================
FILE: engine.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
"""
Train and eval functions used in main.py
"""
import math
import os
import sys
from typing import Iterable
import torch
import util.misc as utils
from datasets.coco_eval import CocoEvaluator
from datasets.panoptic_eval import PanopticEvaluator
def train_one_epoch(model: torch.nn.Module, criterion: torch.nn.Module,
data_loader: Iterable, optimizer: torch.optim.Optimizer,
device: torch.device, epoch: int, max_norm: float = 0):
model.train()
criterion.train()
metric_logger = utils.MetricLogger(delimiter=" ")
metric_logger.add_meter('lr', utils.SmoothedValue(window_size=1, fmt='{value:.6f}'))
metric_logger.add_meter('class_error', utils.SmoothedValue(window_size=1, fmt='{value:.2f}'))
header = 'Epoch: [{}]'.format(epoch)
print_freq = 10
for samples, targets in metric_logger.log_every(data_loader, print_freq, header):
samples = samples.to(device)
targets = [{k: v.to(device) for k, v in t.items()} for t in targets]
outputs = model(samples)
loss_dict = criterion(outputs, targets)
weight_dict = criterion.weight_dict
losses = sum(loss_dict[k] * weight_dict[k] for k in loss_dict.keys() if k in weight_dict)
# reduce losses over all GPUs for logging purposes
loss_dict_reduced = utils.reduce_dict(loss_dict)
loss_dict_reduced_unscaled = {f'{k}_unscaled': v
for k, v in loss_dict_reduced.items()}
loss_dict_reduced_scaled = {k: v * weight_dict[k]
for k, v in loss_dict_reduced.items() if k in weight_dict}
losses_reduced_scaled = sum(loss_dict_reduced_scaled.values())
loss_value = losses_reduced_scaled.item()
if not math.isfinite(loss_value):
print("Loss is {}, stopping training".format(loss_value))
print(loss_dict_reduced)
sys.exit(1)
optimizer.zero_grad()
losses.backward()
if max_norm > 0:
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm)
optimizer.step()
metric_logger.update(loss=loss_value, **loss_dict_reduced_scaled, **loss_dict_reduced_unscaled)
metric_logger.update(class_error=loss_dict_reduced['class_error'])
metric_logger.update(lr=optimizer.param_groups[0]["lr"])
# gather the stats from all processes
metric_logger.synchronize_between_processes()
print("Averaged stats:", metric_logger)
return {k: meter.global_avg for k, meter in metric_logger.meters.items()}
@torch.no_grad()
def evaluate(model, criterion, postprocessors, data_loader, base_ds, device, output_dir):
model.eval()
criterion.eval()
metric_logger = utils.MetricLogger(delimiter=" ")
metric_logger.add_meter('class_error', utils.SmoothedValue(window_size=1, fmt='{value:.2f}'))
header = 'Test:'
iou_types = tuple(k for k in ('segm', 'bbox') if k in postprocessors.keys())
coco_evaluator = CocoEvaluator(base_ds, iou_types)
# coco_evaluator.coco_eval[iou_types[0]].params.iouThrs = [0, 0.1, 0.5, 0.75]
panoptic_evaluator = None
if 'panoptic' in postprocessors.keys():
panoptic_evaluator = PanopticEvaluator(
data_loader.dataset.ann_file,
data_loader.dataset.ann_folder,
output_dir=os.path.join(output_dir, "panoptic_eval"),
)
for samples, targets in metric_logger.log_every(data_loader, 10, header):
samples = samples.to(device)
targets = [{k: v.to(device) for k, v in t.items()} for t in targets]
outputs = model(samples)
loss_dict = criterion(outputs, targets)
weight_dict = criterion.weight_dict
# reduce losses over all GPUs for logging purposes
loss_dict_reduced = utils.reduce_dict(loss_dict)
loss_dict_reduced_scaled = {k: v * weight_dict[k]
for k, v in loss_dict_reduced.items() if k in weight_dict}
loss_dict_reduced_unscaled = {f'{k}_unscaled': v
for k, v in loss_dict_reduced.items()}
metric_logger.update(loss=sum(loss_dict_reduced_scaled.values()),
**loss_dict_reduced_scaled,
**loss_dict_reduced_unscaled)
metric_logger.update(class_error=loss_dict_reduced['class_error'])
orig_target_sizes = torch.stack([t["orig_size"] for t in targets], dim=0)
results = postprocessors['bbox'](outputs, orig_target_sizes)
if 'segm' in postprocessors.keys():
target_sizes = torch.stack([t["size"] for t in targets], dim=0)
results = postprocessors['segm'](results, outputs, orig_target_sizes, target_sizes)
res = {target['image_id'].item(): output for target, output in zip(targets, results)}
if coco_evaluator is not None:
coco_evaluator.update(res)
if panoptic_evaluator is not None:
res_pano = postprocessors["panoptic"](outputs, target_sizes, orig_target_sizes)
for i, target in enumerate(targets):
image_id = target["image_id"].item()
file_name = f"{image_id:012d}.png"
res_pano[i]["image_id"] = image_id
res_pano[i]["file_name"] = file_name
panoptic_evaluator.update(res_pano)
# gather the stats from all processes
metric_logger.synchronize_between_processes()
print("Averaged stats:", metric_logger)
if coco_evaluator is not None:
coco_evaluator.synchronize_between_processes()
if panoptic_evaluator is not None:
panoptic_evaluator.synchronize_between_processes()
# accumulate predictions from all images
if coco_evaluator is not None:
coco_evaluator.accumulate()
coco_evaluator.summarize()
panoptic_res = None
if panoptic_evaluator is not None:
panoptic_res = panoptic_evaluator.summarize()
stats = {k: meter.global_avg for k, meter in metric_logger.meters.items()}
if coco_evaluator is not None:
if 'bbox' in postprocessors.keys():
stats['coco_eval_bbox'] = coco_evaluator.coco_eval['bbox'].stats.tolist()
if 'segm' in postprocessors.keys():
stats['coco_eval_masks'] = coco_evaluator.coco_eval['segm'].stats.tolist()
if panoptic_res is not None:
stats['PQ_all'] = panoptic_res["All"]
stats['PQ_th'] = panoptic_res["Things"]
stats['PQ_st'] = panoptic_res["Stuff"]
return stats, coco_evaluator
================================================
FILE: hubconf.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
import torch
from models.backbone import Backbone, Joiner
from models.detr import DETR, PostProcess
from models.position_encoding import PositionEmbeddingSine
from models.segmentation import DETRsegm, PostProcessPanoptic
from models.transformer import Transformer
dependencies = ["torch", "torchvision"]
def _make_detr(backbone_name: str, dilation=False, num_classes=91, mask=False):
hidden_dim = 256
backbone = Backbone(backbone_name, train_backbone=True, return_interm_layers=mask, dilation=dilation)
pos_enc = PositionEmbeddingSine(hidden_dim // 2, normalize=True)
backbone_with_pos_enc = Joiner(backbone, pos_enc)
backbone_with_pos_enc.num_channels = backbone.num_channels
transformer = Transformer(d_model=hidden_dim, return_intermediate_dec=True)
detr = DETR(backbone_with_pos_enc, transformer, num_classes=num_classes, num_queries=100)
if mask:
return DETRsegm(detr)
return detr
def detr_resnet50(pretrained=False, num_classes=91, return_postprocessor=False):
"""
DETR R50 with 6 encoder and 6 decoder layers.
Achieves 42/62.4 AP/AP50 on COCO val5k.
"""
model = _make_detr("resnet50", dilation=False, num_classes=num_classes)
if pretrained:
checkpoint = torch.hub.load_state_dict_from_url(
url="https://dl.fbaipublicfiles.com/detr/detr-r50-e632da11.pth", map_location="cpu", check_hash=True
)
model.load_state_dict(checkpoint["model"])
if return_postprocessor:
return model, PostProcess()
return model
def detr_resnet50_dc5(pretrained=False, num_classes=91, return_postprocessor=False):
"""
DETR-DC5 R50 with 6 encoder and 6 decoder layers.
The last block of ResNet-50 has dilation to increase
output resolution.
Achieves 43.3/63.1 AP/AP50 on COCO val5k.
"""
model = _make_detr("resnet50", dilation=True, num_classes=num_classes)
if pretrained:
checkpoint = torch.hub.load_state_dict_from_url(
url="https://dl.fbaipublicfiles.com/detr/detr-r50-dc5-f0fb7ef5.pth", map_location="cpu", check_hash=True
)
model.load_state_dict(checkpoint["model"])
if return_postprocessor:
return model, PostProcess()
return model
def detr_resnet101(pretrained=False, num_classes=91, return_postprocessor=False):
"""
DETR-DC5 R101 with 6 encoder and 6 decoder layers.
Achieves 43.5/63.8 AP/AP50 on COCO val5k.
"""
model = _make_detr("resnet101", dilation=False, num_classes=num_classes)
if pretrained:
checkpoint = torch.hub.load_state_dict_from_url(
url="https://dl.fbaipublicfiles.com/detr/detr-r101-2c7b67e5.pth", map_location="cpu", check_hash=True
)
model.load_state_dict(checkpoint["model"])
if return_postprocessor:
return model, PostProcess()
return model
def detr_resnet101_dc5(pretrained=False, num_classes=91, return_postprocessor=False):
"""
DETR-DC5 R101 with 6 encoder and 6 decoder layers.
The last block of ResNet-101 has dilation to increase
output resolution.
Achieves 44.9/64.7 AP/AP50 on COCO val5k.
"""
model = _make_detr("resnet101", dilation=True, num_classes=num_classes)
if pretrained:
checkpoint = torch.hub.load_state_dict_from_url(
url="https://dl.fbaipublicfiles.com/detr/detr-r101-dc5-a2e86def.pth", map_location="cpu", check_hash=True
)
model.load_state_dict(checkpoint["model"])
if return_postprocessor:
return model, PostProcess()
return model
def detr_resnet50_panoptic(
pretrained=False, num_classes=250, threshold=0.85, return_postprocessor=False
):
"""
DETR R50 with 6 encoder and 6 decoder layers.
Achieves 43.4 PQ on COCO val5k.
threshold is the minimum confidence required for keeping segments in the prediction
"""
model = _make_detr("resnet50", dilation=False, num_classes=num_classes, mask=True)
is_thing_map = {i: i <= 90 for i in range(250)}
if pretrained:
checkpoint = torch.hub.load_state_dict_from_url(
url="https://dl.fbaipublicfiles.com/detr/detr-r50-panoptic-00ce5173.pth",
map_location="cpu",
check_hash=True,
)
model.load_state_dict(checkpoint["model"])
if return_postprocessor:
return model, PostProcessPanoptic(is_thing_map, threshold=threshold)
return model
def detr_resnet50_dc5_panoptic(
pretrained=False, num_classes=250, threshold=0.85, return_postprocessor=False
):
"""
DETR-DC5 R50 with 6 encoder and 6 decoder layers.
The last block of ResNet-50 has dilation to increase
output resolution.
Achieves 44.6 on COCO val5k.
threshold is the minimum confidence required for keeping segments in the prediction
"""
model = _make_detr("resnet50", dilation=True, num_classes=num_classes, mask=True)
is_thing_map = {i: i <= 90 for i in range(250)}
if pretrained:
checkpoint = torch.hub.load_state_dict_from_url(
url="https://dl.fbaipublicfiles.com/detr/detr-r50-dc5-panoptic-da08f1b1.pth",
map_location="cpu",
check_hash=True,
)
model.load_state_dict(checkpoint["model"])
if return_postprocessor:
return model, PostProcessPanoptic(is_thing_map, threshold=threshold)
return model
def detr_resnet101_panoptic(
pretrained=False, num_classes=250, threshold=0.85, return_postprocessor=False
):
"""
DETR-DC5 R101 with 6 encoder and 6 decoder layers.
Achieves 45.1 PQ on COCO val5k.
threshold is the minimum confidence required for keeping segments in the prediction
"""
model = _make_detr("resnet101", dilation=False, num_classes=num_classes, mask=True)
is_thing_map = {i: i <= 90 for i in range(250)}
if pretrained:
checkpoint = torch.hub.load_state_dict_from_url(
url="https://dl.fbaipublicfiles.com/detr/detr-r101-panoptic-40021d53.pth",
map_location="cpu",
check_hash=True,
)
model.load_state_dict(checkpoint["model"])
if return_postprocessor:
return model, PostProcessPanoptic(is_thing_map, threshold=threshold)
return model
================================================
FILE: main.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
import argparse
import datetime
import json
import random
import time
from pathlib import Path
import numpy as np
import torch
from torch.utils.data import DataLoader, DistributedSampler
import datasets
import util.misc as utils
from datasets import build_dataset, get_coco_api_from_dataset
from engine import evaluate, train_one_epoch
from models import build_model
def get_args_parser():
parser = argparse.ArgumentParser('Set transformer detector', add_help=False)
parser.add_argument('--lr', default=1e-4, type=float)
parser.add_argument('--lr_backbone', default=1e-5, type=float)
parser.add_argument('--batch_size', default=2, type=int)
parser.add_argument('--weight_decay', default=1e-4, type=float)
parser.add_argument('--epochs', default=300, type=int)
parser.add_argument('--lr_drop', default=200, type=int)
parser.add_argument('--clip_max_norm', default=0.1, type=float,
help='gradient clipping max norm')
# Model parameters
parser.add_argument('--frozen_weights', type=str, default=None,
help="Path to the pretrained model. If set, only the mask head will be trained")
# * Backbone
parser.add_argument('--backbone', default='resnet50', type=str,
help="Name of the convolutional backbone to use")
parser.add_argument('--dilation', action='store_true',
help="If true, we replace stride with dilation in the last convolutional block (DC5)")
parser.add_argument('--position_embedding', default='sine', type=str, choices=('sine', 'learned'),
help="Type of positional embedding to use on top of the image features")
# * Transformer
parser.add_argument('--enc_layers', default=6, type=int,
help="Number of encoding layers in the transformer")
parser.add_argument('--dec_layers', default=6, type=int,
help="Number of decoding layers in the transformer")
parser.add_argument('--dim_feedforward', default=2048, type=int,
help="Intermediate size of the feedforward layers in the transformer blocks")
parser.add_argument('--hidden_dim', default=256, type=int,
help="Size of the embeddings (dimension of the transformer)")
parser.add_argument('--dropout', default=0.1, type=float,
help="Dropout applied in the transformer")
parser.add_argument('--nheads', default=8, type=int,
help="Number of attention heads inside the transformer's attentions")
parser.add_argument('--num_queries', default=100, type=int,
help="Number of query slots")
parser.add_argument('--pre_norm', action='store_true')
# * Segmentation
parser.add_argument('--masks', action='store_true',
help="Train segmentation head if the flag is provided")
# Loss
parser.add_argument('--no_aux_loss', dest='aux_loss', action='store_false',
help="Disables auxiliary decoding losses (loss at each layer)")
# * Matcher
parser.add_argument('--set_cost_class', default=1, type=float,
help="Class coefficient in the matching cost")
parser.add_argument('--set_cost_bbox', default=5, type=float,
help="L1 box coefficient in the matching cost")
parser.add_argument('--set_cost_giou', default=2, type=float,
help="giou box coefficient in the matching cost")
# * Loss coefficients
parser.add_argument('--mask_loss_coef', default=1, type=float)
parser.add_argument('--dice_loss_coef', default=1, type=float)
parser.add_argument('--bbox_loss_coef', default=5, type=float)
parser.add_argument('--giou_loss_coef', default=2, type=float)
parser.add_argument('--eos_coef', default=0.1, type=float,
help="Relative classification weight of the no-object class")
# dataset parameters
parser.add_argument('--dataset_file', default='coco')
parser.add_argument('--coco_path', type=str)
parser.add_argument('--coco_panoptic_path', type=str)
parser.add_argument('--remove_difficult', action='store_true')
parser.add_argument('--output_dir', default='',
help='path where to save, empty for no saving')
parser.add_argument('--device', default='cuda',
help='device to use for training / testing')
parser.add_argument('--seed', default=42, type=int)
parser.add_argument('--resume', default='', help='resume from checkpoint')
parser.add_argument('--start_epoch', default=0, type=int, metavar='N',
help='start epoch')
parser.add_argument('--eval', action='store_true')
parser.add_argument('--num_workers', default=2, type=int)
# distributed training parameters
parser.add_argument('--world_size', default=1, type=int,
help='number of distributed processes')
parser.add_argument('--dist_url', default='env://', help='url used to set up distributed training')
return parser
def main(args):
utils.init_distributed_mode(args)
print("git:\n {}\n".format(utils.get_sha()))
if args.frozen_weights is not None:
assert args.masks, "Frozen training is meant for segmentation only"
print(args)
device = torch.device(args.device)
# fix the seed for reproducibility
seed = args.seed + utils.get_rank()
torch.manual_seed(seed)
np.random.seed(seed)
random.seed(seed)
model, criterion, postprocessors = build_model(args)
model.to(device)
model_without_ddp = model
if args.distributed:
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu])
model_without_ddp = model.module
n_parameters = sum(p.numel() for p in model.parameters() if p.requires_grad)
print('number of params:', n_parameters)
param_dicts = [
{"params": [p for n, p in model_without_ddp.named_parameters() if "backbone" not in n and p.requires_grad]},
{
"params": [p for n, p in model_without_ddp.named_parameters() if "backbone" in n and p.requires_grad],
"lr": args.lr_backbone,
},
]
optimizer = torch.optim.AdamW(param_dicts, lr=args.lr,
weight_decay=args.weight_decay)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, args.lr_drop)
dataset_train = build_dataset(image_set='train', args=args)
dataset_val = build_dataset(image_set='val', args=args)
if args.distributed:
sampler_train = DistributedSampler(dataset_train)
sampler_val = DistributedSampler(dataset_val, shuffle=False)
else:
sampler_train = torch.utils.data.RandomSampler(dataset_train)
sampler_val = torch.utils.data.SequentialSampler(dataset_val)
batch_sampler_train = torch.utils.data.BatchSampler(
sampler_train, args.batch_size, drop_last=True)
data_loader_train = DataLoader(dataset_train, batch_sampler=batch_sampler_train,
collate_fn=utils.collate_fn, num_workers=args.num_workers)
data_loader_val = DataLoader(dataset_val, args.batch_size, sampler=sampler_val,
drop_last=False, collate_fn=utils.collate_fn, num_workers=args.num_workers)
if args.dataset_file == "coco_panoptic":
# We also evaluate AP during panoptic training, on original coco DS
coco_val = datasets.coco.build("val", args)
base_ds = get_coco_api_from_dataset(coco_val)
else:
base_ds = get_coco_api_from_dataset(dataset_val)
if args.frozen_weights is not None:
checkpoint = torch.load(args.frozen_weights, map_location='cpu')
model_without_ddp.detr.load_state_dict(checkpoint['model'])
output_dir = Path(args.output_dir)
if args.resume:
if args.resume.startswith('https'):
checkpoint = torch.hub.load_state_dict_from_url(
args.resume, map_location='cpu', check_hash=True)
else:
checkpoint = torch.load(args.resume, map_location='cpu')
model_without_ddp.load_state_dict(checkpoint['model'])
if not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint:
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
args.start_epoch = checkpoint['epoch'] + 1
if args.eval:
test_stats, coco_evaluator = evaluate(model, criterion, postprocessors,
data_loader_val, base_ds, device, args.output_dir)
if args.output_dir:
utils.save_on_master(coco_evaluator.coco_eval["bbox"].eval, output_dir / "eval.pth")
return
print("Start training")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
sampler_train.set_epoch(epoch)
train_stats = train_one_epoch(
model, criterion, data_loader_train, optimizer, device, epoch,
args.clip_max_norm)
lr_scheduler.step()
if args.output_dir:
checkpoint_paths = [output_dir / 'checkpoint.pth']
# extra checkpoint before LR drop and every 100 epochs
if (epoch + 1) % args.lr_drop == 0 or (epoch + 1) % 100 == 0:
checkpoint_paths.append(output_dir / f'checkpoint{epoch:04}.pth')
for checkpoint_path in checkpoint_paths:
utils.save_on_master({
'model': model_without_ddp.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch,
'args': args,
}, checkpoint_path)
test_stats, coco_evaluator = evaluate(
model, criterion, postprocessors, data_loader_val, base_ds, device, args.output_dir
)
log_stats = {**{f'train_{k}': v for k, v in train_stats.items()},
**{f'test_{k}': v for k, v in test_stats.items()},
'epoch': epoch,
'n_parameters': n_parameters}
if args.output_dir and utils.is_main_process():
with (output_dir / "log.txt").open("a") as f:
f.write(json.dumps(log_stats) + "\n")
# for evaluation logs
if coco_evaluator is not None:
(output_dir / 'eval').mkdir(exist_ok=True)
if "bbox" in coco_evaluator.coco_eval:
filenames = ['latest.pth']
if epoch % 50 == 0:
filenames.append(f'{epoch:03}.pth')
for name in filenames:
torch.save(coco_evaluator.coco_eval["bbox"].eval,
output_dir / "eval" / name)
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
if __name__ == '__main__':
parser = argparse.ArgumentParser('DETR training and evaluation script', parents=[get_args_parser()])
args = parser.parse_args()
if args.output_dir:
Path(args.output_dir).mkdir(parents=True, exist_ok=True)
main(args)
================================================
FILE: models/__init__.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
from .detr import build
def build_model(args):
return build(args)
================================================
FILE: models/backbone.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
"""
Backbone modules.
"""
from collections import OrderedDict
import torch
import torch.nn.functional as F
import torchvision
from torch import nn
from torchvision.models._utils import IntermediateLayerGetter
from typing import Dict, List
from util.misc import NestedTensor, is_main_process
from .position_encoding import build_position_encoding
class FrozenBatchNorm2d(torch.nn.Module):
"""
BatchNorm2d where the batch statistics and the affine parameters are fixed.
Copy-paste from torchvision.misc.ops with added eps before rqsrt,
without which any other models than torchvision.models.resnet[18,34,50,101]
produce nans.
"""
def __init__(self, n):
super(FrozenBatchNorm2d, self).__init__()
self.register_buffer("weight", torch.ones(n))
self.register_buffer("bias", torch.zeros(n))
self.register_buffer("running_mean", torch.zeros(n))
self.register_buffer("running_var", torch.ones(n))
def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict,
missing_keys, unexpected_keys, error_msgs):
num_batches_tracked_key = prefix + 'num_batches_tracked'
if num_batches_tracked_key in state_dict:
del state_dict[num_batches_tracked_key]
super(FrozenBatchNorm2d, self)._load_from_state_dict(
state_dict, prefix, local_metadata, strict,
missing_keys, unexpected_keys, error_msgs)
def forward(self, x):
# move reshapes to the beginning
# to make it fuser-friendly
w = self.weight.reshape(1, -1, 1, 1)
b = self.bias.reshape(1, -1, 1, 1)
rv = self.running_var.reshape(1, -1, 1, 1)
rm = self.running_mean.reshape(1, -1, 1, 1)
eps = 1e-5
scale = w * (rv + eps).rsqrt()
bias = b - rm * scale
return x * scale + bias
class BackboneBase(nn.Module):
def __init__(self, backbone: nn.Module, train_backbone: bool, num_channels: int, return_interm_layers: bool):
super().__init__()
for name, parameter in backbone.named_parameters():
if not train_backbone or 'layer2' not in name and 'layer3' not in name and 'layer4' not in name:
parameter.requires_grad_(False)
if return_interm_layers:
return_layers = {"layer1": "0", "layer2": "1", "layer3": "2", "layer4": "3"}
else:
return_layers = {'layer4': "0"}
self.body = IntermediateLayerGetter(backbone, return_layers=return_layers)
self.num_channels = num_channels
def forward(self, tensor_list: NestedTensor):
xs = self.body(tensor_list.tensors)
out: Dict[str, NestedTensor] = {}
for name, x in xs.items():
m = tensor_list.mask
assert m is not None
mask = F.interpolate(m[None].float(), size=x.shape[-2:]).to(torch.bool)[0]
out[name] = NestedTensor(x, mask)
return out
class Backbone(BackboneBase):
"""ResNet backbone with frozen BatchNorm."""
def __init__(self, name: str,
train_backbone: bool,
return_interm_layers: bool,
dilation: bool):
backbone = getattr(torchvision.models, name)(
replace_stride_with_dilation=[False, False, dilation],
pretrained=is_main_process(), norm_layer=FrozenBatchNorm2d)
num_channels = 512 if name in ('resnet18', 'resnet34') else 2048
super().__init__(backbone, train_backbone, num_channels, return_interm_layers)
class Joiner(nn.Sequential):
def __init__(self, backbone, position_embedding):
super().__init__(backbone, position_embedding)
def forward(self, tensor_list: NestedTensor):
xs = self[0](tensor_list)
out: List[NestedTensor] = []
pos = []
for name, x in xs.items():
out.append(x)
# position encoding
pos.append(self[1](x).to(x.tensors.dtype))
return out, pos
def build_backbone(args):
position_embedding = build_position_encoding(args)
train_backbone = args.lr_backbone > 0
return_interm_layers = args.masks
backbone = Backbone(args.backbone, train_backbone, return_interm_layers, args.dilation)
model = Joiner(backbone, position_embedding)
model.num_channels = backbone.num_channels
return model
================================================
FILE: models/detr.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
"""
DETR model and criterion classes.
"""
import torch
import torch.nn.functional as F
from torch import nn
from util import box_ops
from util.misc import (NestedTensor, nested_tensor_from_tensor_list,
accuracy, get_world_size, interpolate,
is_dist_avail_and_initialized)
from .backbone import build_backbone
from .matcher import build_matcher
from .segmentation import (DETRsegm, PostProcessPanoptic, PostProcessSegm,
dice_loss, sigmoid_focal_loss)
from .transformer import build_transformer
class DETR(nn.Module):
""" This is the DETR module that performs object detection """
def __init__(self, backbone, transformer, num_classes, num_queries, aux_loss=False):
""" Initializes the model.
Parameters:
backbone: torch module of the backbone to be used. See backbone.py
transformer: torch module of the transformer architecture. See transformer.py
num_classes: number of object classes
num_queries: number of object queries, ie detection slot. This is the maximal number of objects
DETR can detect in a single image. For COCO, we recommend 100 queries.
aux_loss: True if auxiliary decoding losses (loss at each decoder layer) are to be used.
"""
super().__init__()
self.num_queries = num_queries
self.transformer = transformer
hidden_dim = transformer.d_model
self.class_embed = nn.Linear(hidden_dim, num_classes + 1)
self.bbox_embed = MLP(hidden_dim, hidden_dim, 4, 3)
self.query_embed = nn.Embedding(num_queries, hidden_dim)
self.input_proj = nn.Conv2d(backbone.num_channels, hidden_dim, kernel_size=1)
self.backbone = backbone
self.aux_loss = aux_loss
def forward(self, samples: NestedTensor):
""" The forward expects a NestedTensor, which consists of:
- samples.tensor: batched images, of shape [batch_size x 3 x H x W]
- samples.mask: a binary mask of shape [batch_size x H x W], containing 1 on padded pixels
It returns a dict with the following elements:
- "pred_logits": the classification logits (including no-object) for all queries.
Shape= [batch_size x num_queries x (num_classes + 1)]
- "pred_boxes": The normalized boxes coordinates for all queries, represented as
(center_x, center_y, height, width). These values are normalized in [0, 1],
relative to the size of each individual image (disregarding possible padding).
See PostProcess for information on how to retrieve the unnormalized bounding box.
- "aux_outputs": Optional, only returned when auxilary losses are activated. It is a list of
dictionnaries containing the two above keys for each decoder layer.
"""
if isinstance(samples, (list, torch.Tensor)):
samples = nested_tensor_from_tensor_list(samples)
features, pos = self.backbone(samples)
src, mask = features[-1].decompose()
assert mask is not None
hs = self.transformer(self.input_proj(src), mask, self.query_embed.weight, pos[-1])[0]
outputs_class = self.class_embed(hs)
outputs_coord = self.bbox_embed(hs).sigmoid()
out = {'pred_logits': outputs_class[-1], 'pred_boxes': outputs_coord[-1]}
if self.aux_loss:
out['aux_outputs'] = self._set_aux_loss(outputs_class, outputs_coord)
return out
@torch.jit.unused
def _set_aux_loss(self, outputs_class, outputs_coord):
# this is a workaround to make torchscript happy, as torchscript
# doesn't support dictionary with non-homogeneous values, such
# as a dict having both a Tensor and a list.
return [{'pred_logits': a, 'pred_boxes': b}
for a, b in zip(outputs_class[:-1], outputs_coord[:-1])]
class SetCriterion(nn.Module):
""" This class computes the loss for DETR.
The process happens in two steps:
1) we compute hungarian assignment between ground truth boxes and the outputs of the model
2) we supervise each pair of matched ground-truth / prediction (supervise class and box)
"""
def __init__(self, num_classes, matcher, weight_dict, eos_coef, losses):
""" Create the criterion.
Parameters:
num_classes: number of object categories, omitting the special no-object category
matcher: module able to compute a matching between targets and proposals
weight_dict: dict containing as key the names of the losses and as values their relative weight.
eos_coef: relative classification weight applied to the no-object category
losses: list of all the losses to be applied. See get_loss for list of available losses.
"""
super().__init__()
self.num_classes = num_classes
self.matcher = matcher
self.weight_dict = weight_dict
self.eos_coef = eos_coef
self.losses = losses
empty_weight = torch.ones(self.num_classes + 1)
empty_weight[-1] = self.eos_coef
self.register_buffer('empty_weight', empty_weight)
def loss_labels(self, outputs, targets, indices, num_boxes, log=True):
"""Classification loss (NLL)
targets dicts must contain the key "labels" containing a tensor of dim [nb_target_boxes]
"""
assert 'pred_logits' in outputs
src_logits = outputs['pred_logits']
idx = self._get_src_permutation_idx(indices)
target_classes_o = torch.cat([t["labels"][J] for t, (_, J) in zip(targets, indices)])
target_classes = torch.full(src_logits.shape[:2], self.num_classes,
dtype=torch.int64, device=src_logits.device)
target_classes[idx] = target_classes_o
loss_ce = F.cross_entropy(src_logits.transpose(1, 2), target_classes, self.empty_weight)
losses = {'loss_ce': loss_ce}
if log:
# TODO this should probably be a separate loss, not hacked in this one here
losses['class_error'] = 100 - accuracy(src_logits[idx], target_classes_o)[0]
return losses
@torch.no_grad()
def loss_cardinality(self, outputs, targets, indices, num_boxes):
""" Compute the cardinality error, ie the absolute error in the number of predicted non-empty boxes
This is not really a loss, it is intended for logging purposes only. It doesn't propagate gradients
"""
pred_logits = outputs['pred_logits']
device = pred_logits.device
tgt_lengths = torch.as_tensor([len(v["labels"]) for v in targets], device=device)
# Count the number of predictions that are NOT "no-object" (which is the last class)
card_pred = (pred_logits.argmax(-1) != pred_logits.shape[-1] - 1).sum(1)
card_err = F.l1_loss(card_pred.float(), tgt_lengths.float())
losses = {'cardinality_error': card_err}
return losses
def loss_boxes(self, outputs, targets, indices, num_boxes):
"""Compute the losses related to the bounding boxes, the L1 regression loss and the GIoU loss
targets dicts must contain the key "boxes" containing a tensor of dim [nb_target_boxes, 4]
The target boxes are expected in format (center_x, center_y, w, h), normalized by the image size.
"""
assert 'pred_boxes' in outputs
idx = self._get_src_permutation_idx(indices)
src_boxes = outputs['pred_boxes'][idx]
target_boxes = torch.cat([t['boxes'][i] for t, (_, i) in zip(targets, indices)], dim=0)
loss_bbox = F.l1_loss(src_boxes, target_boxes, reduction='none')
losses = {}
losses['loss_bbox'] = loss_bbox.sum() / num_boxes
loss_giou = 1 - torch.diag(box_ops.generalized_box_iou(
box_ops.box_cxcywh_to_xyxy(src_boxes),
box_ops.box_cxcywh_to_xyxy(target_boxes)))
losses['loss_giou'] = loss_giou.sum() / num_boxes
return losses
def loss_masks(self, outputs, targets, indices, num_boxes):
"""Compute the losses related to the masks: the focal loss and the dice loss.
targets dicts must contain the key "masks" containing a tensor of dim [nb_target_boxes, h, w]
"""
assert "pred_masks" in outputs
src_idx = self._get_src_permutation_idx(indices)
tgt_idx = self._get_tgt_permutation_idx(indices)
src_masks = outputs["pred_masks"]
src_masks = src_masks[src_idx]
masks = [t["masks"] for t in targets]
# TODO use valid to mask invalid areas due to padding in loss
target_masks, valid = nested_tensor_from_tensor_list(masks).decompose()
target_masks = target_masks.to(src_masks)
target_masks = target_masks[tgt_idx]
# upsample predictions to the target size
src_masks = interpolate(src_masks[:, None], size=target_masks.shape[-2:],
mode="bilinear", align_corners=False)
src_masks = src_masks[:, 0].flatten(1)
target_masks = target_masks.flatten(1)
target_masks = target_masks.view(src_masks.shape)
losses = {
"loss_mask": sigmoid_focal_loss(src_masks, target_masks, num_boxes),
"loss_dice": dice_loss(src_masks, target_masks, num_boxes),
}
return losses
def _get_src_permutation_idx(self, indices):
# permute predictions following indices
batch_idx = torch.cat([torch.full_like(src, i) for i, (src, _) in enumerate(indices)])
src_idx = torch.cat([src for (src, _) in indices])
return batch_idx, src_idx
def _get_tgt_permutation_idx(self, indices):
# permute targets following indices
batch_idx = torch.cat([torch.full_like(tgt, i) for i, (_, tgt) in enumerate(indices)])
tgt_idx = torch.cat([tgt for (_, tgt) in indices])
return batch_idx, tgt_idx
def get_loss(self, loss, outputs, targets, indices, num_boxes, **kwargs):
loss_map = {
'labels': self.loss_labels,
'cardinality': self.loss_cardinality,
'boxes': self.loss_boxes,
'masks': self.loss_masks
}
assert loss in loss_map, f'do you really want to compute {loss} loss?'
return loss_map[loss](outputs, targets, indices, num_boxes, **kwargs)
def forward(self, outputs, targets):
""" This performs the loss computation.
Parameters:
outputs: dict of tensors, see the output specification of the model for the format
targets: list of dicts, such that len(targets) == batch_size.
The expected keys in each dict depends on the losses applied, see each loss' doc
"""
outputs_without_aux = {k: v for k, v in outputs.items() if k != 'aux_outputs'}
# Retrieve the matching between the outputs of the last layer and the targets
indices = self.matcher(outputs_without_aux, targets)
# Compute the average number of target boxes accross all nodes, for normalization purposes
num_boxes = sum(len(t["labels"]) for t in targets)
num_boxes = torch.as_tensor([num_boxes], dtype=torch.float, device=next(iter(outputs.values())).device)
if is_dist_avail_and_initialized():
torch.distributed.all_reduce(num_boxes)
num_boxes = torch.clamp(num_boxes / get_world_size(), min=1).item()
# Compute all the requested losses
losses = {}
for loss in self.losses:
losses.update(self.get_loss(loss, outputs, targets, indices, num_boxes))
# In case of auxiliary losses, we repeat this process with the output of each intermediate layer.
if 'aux_outputs' in outputs:
for i, aux_outputs in enumerate(outputs['aux_outputs']):
indices = self.matcher(aux_outputs, targets)
for loss in self.losses:
if loss == 'masks':
# Intermediate masks losses are too costly to compute, we ignore them.
continue
kwargs = {}
if loss == 'labels':
# Logging is enabled only for the last layer
kwargs = {'log': False}
l_dict = self.get_loss(loss, aux_outputs, targets, indices, num_boxes, **kwargs)
l_dict = {k + f'_{i}': v for k, v in l_dict.items()}
losses.update(l_dict)
return losses
class PostProcess(nn.Module):
""" This module converts the model's output into the format expected by the coco api"""
@torch.no_grad()
def forward(self, outputs, target_sizes):
""" Perform the computation
Parameters:
outputs: raw outputs of the model
target_sizes: tensor of dimension [batch_size x 2] containing the size of each images of the batch
For evaluation, this must be the original image size (before any data augmentation)
For visualization, this should be the image size after data augment, but before padding
"""
out_logits, out_bbox = outputs['pred_logits'], outputs['pred_boxes']
assert len(out_logits) == len(target_sizes)
assert target_sizes.shape[1] == 2
prob = F.softmax(out_logits, -1)
scores, labels = prob[..., :-1].max(-1)
# convert to [x0, y0, x1, y1] format
boxes = box_ops.box_cxcywh_to_xyxy(out_bbox)
# and from relative [0, 1] to absolute [0, height] coordinates
img_h, img_w = target_sizes.unbind(1)
scale_fct = torch.stack([img_w, img_h, img_w, img_h], dim=1)
boxes = boxes * scale_fct[:, None, :]
results = [{'scores': s, 'labels': l, 'boxes': b} for s, l, b in zip(scores, labels, boxes)]
return results
class MLP(nn.Module):
""" Very simple multi-layer perceptron (also called FFN)"""
def __init__(self, input_dim, hidden_dim, output_dim, num_layers):
super().__init__()
self.num_layers = num_layers
h = [hidden_dim] * (num_layers - 1)
self.layers = nn.ModuleList(nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim]))
def forward(self, x):
for i, layer in enumerate(self.layers):
x = F.relu(layer(x)) if i < self.num_layers - 1 else layer(x)
return x
def build(args):
# the `num_classes` naming here is somewhat misleading.
# it indeed corresponds to `max_obj_id + 1`, where max_obj_id
# is the maximum id for a class in your dataset. For example,
# COCO has a max_obj_id of 90, so we pass `num_classes` to be 91.
# As another example, for a dataset that has a single class with id 1,
# you should pass `num_classes` to be 2 (max_obj_id + 1).
# For more details on this, check the following discussion
# https://github.com/facebookresearch/detr/issues/108#issuecomment-650269223
num_classes = 20 if args.dataset_file != 'coco' else 91
if args.dataset_file == "coco_panoptic":
# for panoptic, we just add a num_classes that is large enough to hold
# max_obj_id + 1, but the exact value doesn't really matter
num_classes = 250
device = torch.device(args.device)
backbone = build_backbone(args)
transformer = build_transformer(args)
model = DETR(
backbone,
transformer,
num_classes=num_classes,
num_queries=args.num_queries,
aux_loss=args.aux_loss,
)
if args.masks:
model = DETRsegm(model, freeze_detr=(args.frozen_weights is not None))
matcher = build_matcher(args)
weight_dict = {'loss_ce': 1, 'loss_bbox': args.bbox_loss_coef}
weight_dict['loss_giou'] = args.giou_loss_coef
if args.masks:
weight_dict["loss_mask"] = args.mask_loss_coef
weight_dict["loss_dice"] = args.dice_loss_coef
# TODO this is a hack
if args.aux_loss:
aux_weight_dict = {}
for i in range(args.dec_layers - 1):
aux_weight_dict.update({k + f'_{i}': v for k, v in weight_dict.items()})
weight_dict.update(aux_weight_dict)
losses = ['labels', 'boxes', 'cardinality']
if args.masks:
losses += ["masks"]
criterion = SetCriterion(num_classes, matcher=matcher, weight_dict=weight_dict,
eos_coef=args.eos_coef, losses=losses)
criterion.to(device)
postprocessors = {'bbox': PostProcess()}
if args.masks:
postprocessors['segm'] = PostProcessSegm()
if args.dataset_file == "coco_panoptic":
is_thing_map = {i: i <= 90 for i in range(201)}
postprocessors["panoptic"] = PostProcessPanoptic(is_thing_map, threshold=0.85)
return model, criterion, postprocessors
================================================
FILE: models/matcher.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
"""
Modules to compute the matching cost and solve the corresponding LSAP.
"""
import torch
from scipy.optimize import linear_sum_assignment
from torch import nn
from util.box_ops import box_cxcywh_to_xyxy, generalized_box_iou
class HungarianMatcher(nn.Module):
"""This class computes an assignment between the targets and the predictions of the network
For efficiency reasons, the targets don't include the no_object. Because of this, in general,
there are more predictions than targets. In this case, we do a 1-to-1 matching of the best predictions,
while the others are un-matched (and thus treated as non-objects).
"""
def __init__(self, cost_class: float = 1, cost_bbox: float = 1, cost_giou: float = 1):
"""Creates the matcher
Params:
cost_class: This is the relative weight of the classification error in the matching cost
cost_bbox: This is the relative weight of the L1 error of the bounding box coordinates in the matching cost
cost_giou: This is the relative weight of the giou loss of the bounding box in the matching cost
"""
super().__init__()
self.cost_class = cost_class
self.cost_bbox = cost_bbox
self.cost_giou = cost_giou
assert cost_class != 0 or cost_bbox != 0 or cost_giou != 0, "all costs cant be 0"
@torch.no_grad()
def forward(self, outputs, targets):
""" Performs the matching
Params:
outputs: This is a dict that contains at least these entries:
"pred_logits": Tensor of dim [batch_size, num_queries, num_classes] with the classification logits
"pred_boxes": Tensor of dim [batch_size, num_queries, 4] with the predicted box coordinates
targets: This is a list of targets (len(targets) = batch_size), where each target is a dict containing:
"labels": Tensor of dim [num_target_boxes] (where num_target_boxes is the number of ground-truth
objects in the target) containing the class labels
"boxes": Tensor of dim [num_target_boxes, 4] containing the target box coordinates
Returns:
A list of size batch_size, containing tuples of (index_i, index_j) where:
- index_i is the indices of the selected predictions (in order)
- index_j is the indices of the corresponding selected targets (in order)
For each batch element, it holds:
len(index_i) = len(index_j) = min(num_queries, num_target_boxes)
"""
bs, num_queries = outputs["pred_logits"].shape[:2]
# We flatten to compute the cost matrices in a batch
out_prob = outputs["pred_logits"].flatten(0, 1).softmax(-1) # [batch_size * num_queries, num_classes]
out_bbox = outputs["pred_boxes"].flatten(0, 1) # [batch_size * num_queries, 4]
# Also concat the target labels and boxes
tgt_ids = torch.cat([v["labels"] for v in targets])
tgt_bbox = torch.cat([v["boxes"] for v in targets])
# Compute the classification cost. Contrary to the loss, we don't use the NLL,
# but approximate it in 1 - proba[target class].
# The 1 is a constant that doesn't change the matching, it can be ommitted.
cost_class = -out_prob[:, tgt_ids]
# Compute the L1 cost between boxes
cost_bbox = torch.cdist(out_bbox, tgt_bbox, p=1)
# Compute the giou cost betwen boxes
cost_giou = -generalized_box_iou(box_cxcywh_to_xyxy(out_bbox), box_cxcywh_to_xyxy(tgt_bbox))
# Final cost matrix
C = self.cost_bbox * cost_bbox + self.cost_class * cost_class + self.cost_giou * cost_giou
C = C.view(bs, num_queries, -1).cpu()
sizes = [len(v["boxes"]) for v in targets]
indices = [linear_sum_assignment(c[i]) for i, c in enumerate(C.split(sizes, -1))]
return [(torch.as_tensor(i, dtype=torch.int64), torch.as_tensor(j, dtype=torch.int64)) for i, j in indices]
def build_matcher(args):
return HungarianMatcher(cost_class=args.set_cost_class, cost_bbox=args.set_cost_bbox, cost_giou=args.set_cost_giou)
================================================
FILE: models/position_encoding.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
"""
Various positional encodings for the transformer.
"""
import math
import torch
from torch import nn
from util.misc import NestedTensor
class PositionEmbeddingSine(nn.Module):
"""
This is a more standard version of the position embedding, very similar to the one
used by the Attention is all you need paper, generalized to work on images.
"""
def __init__(self, num_pos_feats=64, temperature=10000, normalize=False, scale=None):
super().__init__()
self.num_pos_feats = num_pos_feats
self.temperature = temperature
self.normalize = normalize
if scale is not None and normalize is False:
raise ValueError("normalize should be True if scale is passed")
if scale is None:
scale = 2 * math.pi
self.scale = scale
def forward(self, tensor_list: NestedTensor):
x = tensor_list.tensors
mask = tensor_list.mask
assert mask is not None
not_mask = ~mask
y_embed = not_mask.cumsum(1, dtype=torch.float32)
x_embed = not_mask.cumsum(2, dtype=torch.float32)
if self.normalize:
eps = 1e-6
y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale
x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale
dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)
pos_x = x_embed[:, :, :, None] / dim_t
pos_y = y_embed[:, :, :, None] / dim_t
pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4).flatten(3)
pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4).flatten(3)
pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
return pos
class PositionEmbeddingLearned(nn.Module):
"""
Absolute pos embedding, learned.
"""
def __init__(self, num_pos_feats=256):
super().__init__()
self.row_embed = nn.Embedding(50, num_pos_feats)
self.col_embed = nn.Embedding(50, num_pos_feats)
self.reset_parameters()
def reset_parameters(self):
nn.init.uniform_(self.row_embed.weight)
nn.init.uniform_(self.col_embed.weight)
def forward(self, tensor_list: NestedTensor):
x = tensor_list.tensors
h, w = x.shape[-2:]
i = torch.arange(w, device=x.device)
j = torch.arange(h, device=x.device)
x_emb = self.col_embed(i)
y_emb = self.row_embed(j)
pos = torch.cat([
x_emb.unsqueeze(0).repeat(h, 1, 1),
y_emb.unsqueeze(1).repeat(1, w, 1),
], dim=-1).permute(2, 0, 1).unsqueeze(0).repeat(x.shape[0], 1, 1, 1)
return pos
def build_position_encoding(args):
N_steps = args.hidden_dim // 2
if args.position_embedding in ('v2', 'sine'):
# TODO find a better way of exposing other arguments
position_embedding = PositionEmbeddingSine(N_steps, normalize=True)
elif args.position_embedding in ('v3', 'learned'):
position_embedding = PositionEmbeddingLearned(N_steps)
else:
raise ValueError(f"not supported {args.position_embedding}")
return position_embedding
================================================
FILE: models/segmentation.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
"""
This file provides the definition of the convolutional heads used to predict masks, as well as the losses
"""
import io
from collections import defaultdict
from typing import List, Optional
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch import Tensor
from PIL import Image
import util.box_ops as box_ops
from util.misc import NestedTensor, interpolate, nested_tensor_from_tensor_list
try:
from panopticapi.utils import id2rgb, rgb2id
except ImportError:
pass
class DETRsegm(nn.Module):
def __init__(self, detr, freeze_detr=False):
super().__init__()
self.detr = detr
if freeze_detr:
for p in self.parameters():
p.requires_grad_(False)
hidden_dim, nheads = detr.transformer.d_model, detr.transformer.nhead
self.bbox_attention = MHAttentionMap(hidden_dim, hidden_dim, nheads, dropout=0.0)
self.mask_head = MaskHeadSmallConv(hidden_dim + nheads, [1024, 512, 256], hidden_dim)
def forward(self, samples: NestedTensor):
if isinstance(samples, (list, torch.Tensor)):
samples = nested_tensor_from_tensor_list(samples)
features, pos = self.detr.backbone(samples)
bs = features[-1].tensors.shape[0]
src, mask = features[-1].decompose()
assert mask is not None
src_proj = self.detr.input_proj(src)
hs, memory = self.detr.transformer(src_proj, mask, self.detr.query_embed.weight, pos[-1])
outputs_class = self.detr.class_embed(hs)
outputs_coord = self.detr.bbox_embed(hs).sigmoid()
out = {"pred_logits": outputs_class[-1], "pred_boxes": outputs_coord[-1]}
if self.detr.aux_loss:
out['aux_outputs'] = self.detr._set_aux_loss(outputs_class, outputs_coord)
# FIXME h_boxes takes the last one computed, keep this in mind
bbox_mask = self.bbox_attention(hs[-1], memory, mask=mask)
seg_masks = self.mask_head(src_proj, bbox_mask, [features[2].tensors, features[1].tensors, features[0].tensors])
outputs_seg_masks = seg_masks.view(bs, self.detr.num_queries, seg_masks.shape[-2], seg_masks.shape[-1])
out["pred_masks"] = outputs_seg_masks
return out
def _expand(tensor, length: int):
return tensor.unsqueeze(1).repeat(1, int(length), 1, 1, 1).flatten(0, 1)
class MaskHeadSmallConv(nn.Module):
"""
Simple convolutional head, using group norm.
Upsampling is done using a FPN approach
"""
def __init__(self, dim, fpn_dims, context_dim):
super().__init__()
inter_dims = [dim, context_dim // 2, context_dim // 4, context_dim // 8, context_dim // 16, context_dim // 64]
self.lay1 = torch.nn.Conv2d(dim, dim, 3, padding=1)
self.gn1 = torch.nn.GroupNorm(8, dim)
self.lay2 = torch.nn.Conv2d(dim, inter_dims[1], 3, padding=1)
self.gn2 = torch.nn.GroupNorm(8, inter_dims[1])
self.lay3 = torch.nn.Conv2d(inter_dims[1], inter_dims[2], 3, padding=1)
self.gn3 = torch.nn.GroupNorm(8, inter_dims[2])
self.lay4 = torch.nn.Conv2d(inter_dims[2], inter_dims[3], 3, padding=1)
self.gn4 = torch.nn.GroupNorm(8, inter_dims[3])
self.lay5 = torch.nn.Conv2d(inter_dims[3], inter_dims[4], 3, padding=1)
self.gn5 = torch.nn.GroupNorm(8, inter_dims[4])
self.out_lay = torch.nn.Conv2d(inter_dims[4], 1, 3, padding=1)
self.dim = dim
self.adapter1 = torch.nn.Conv2d(fpn_dims[0], inter_dims[1], 1)
self.adapter2 = torch.nn.Conv2d(fpn_dims[1], inter_dims[2], 1)
self.adapter3 = torch.nn.Conv2d(fpn_dims[2], inter_dims[3], 1)
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_uniform_(m.weight, a=1)
nn.init.constant_(m.bias, 0)
def forward(self, x: Tensor, bbox_mask: Tensor, fpns: List[Tensor]):
x = torch.cat([_expand(x, bbox_mask.shape[1]), bbox_mask.flatten(0, 1)], 1)
x = self.lay1(x)
x = self.gn1(x)
x = F.relu(x)
x = self.lay2(x)
x = self.gn2(x)
x = F.relu(x)
cur_fpn = self.adapter1(fpns[0])
if cur_fpn.size(0) != x.size(0):
cur_fpn = _expand(cur_fpn, x.size(0) // cur_fpn.size(0))
x = cur_fpn + F.interpolate(x, size=cur_fpn.shape[-2:], mode="nearest")
x = self.lay3(x)
x = self.gn3(x)
x = F.relu(x)
cur_fpn = self.adapter2(fpns[1])
if cur_fpn.size(0) != x.size(0):
cur_fpn = _expand(cur_fpn, x.size(0) // cur_fpn.size(0))
x = cur_fpn + F.interpolate(x, size=cur_fpn.shape[-2:], mode="nearest")
x = self.lay4(x)
x = self.gn4(x)
x = F.relu(x)
cur_fpn = self.adapter3(fpns[2])
if cur_fpn.size(0) != x.size(0):
cur_fpn = _expand(cur_fpn, x.size(0) // cur_fpn.size(0))
x = cur_fpn + F.interpolate(x, size=cur_fpn.shape[-2:], mode="nearest")
x = self.lay5(x)
x = self.gn5(x)
x = F.relu(x)
x = self.out_lay(x)
return x
class MHAttentionMap(nn.Module):
"""This is a 2D attention module, which only returns the attention softmax (no multiplication by value)"""
def __init__(self, query_dim, hidden_dim, num_heads, dropout=0.0, bias=True):
super().__init__()
self.num_heads = num_heads
self.hidden_dim = hidden_dim
self.dropout = nn.Dropout(dropout)
self.q_linear = nn.Linear(query_dim, hidden_dim, bias=bias)
self.k_linear = nn.Linear(query_dim, hidden_dim, bias=bias)
nn.init.zeros_(self.k_linear.bias)
nn.init.zeros_(self.q_linear.bias)
nn.init.xavier_uniform_(self.k_linear.weight)
nn.init.xavier_uniform_(self.q_linear.weight)
self.normalize_fact = float(hidden_dim / self.num_heads) ** -0.5
def forward(self, q, k, mask: Optional[Tensor] = None):
q = self.q_linear(q)
k = F.conv2d(k, self.k_linear.weight.unsqueeze(-1).unsqueeze(-1), self.k_linear.bias)
qh = q.view(q.shape[0], q.shape[1], self.num_heads, self.hidden_dim // self.num_heads)
kh = k.view(k.shape[0], self.num_heads, self.hidden_dim // self.num_heads, k.shape[-2], k.shape[-1])
weights = torch.einsum("bqnc,bnchw->bqnhw", qh * self.normalize_fact, kh)
if mask is not None:
weights.masked_fill_(mask.unsqueeze(1).unsqueeze(1), float("-inf"))
weights = F.softmax(weights.flatten(2), dim=-1).view(weights.size())
weights = self.dropout(weights)
return weights
def dice_loss(inputs, targets, num_boxes):
"""
Compute the DICE loss, similar to generalized IOU for masks
Args:
inputs: A float tensor of arbitrary shape.
The predictions for each example.
targets: A float tensor with the same shape as inputs. Stores the binary
classification label for each element in inputs
(0 for the negative class and 1 for the positive class).
"""
inputs = inputs.sigmoid()
inputs = inputs.flatten(1)
numerator = 2 * (inputs * targets).sum(1)
denominator = inputs.sum(-1) + targets.sum(-1)
loss = 1 - (numerator + 1) / (denominator + 1)
return loss.sum() / num_boxes
def sigmoid_focal_loss(inputs, targets, num_boxes, alpha: float = 0.25, gamma: float = 2):
"""
Loss used in RetinaNet for dense detection: https://arxiv.org/abs/1708.02002.
Args:
inputs: A float tensor of arbitrary shape.
The predictions for each example.
targets: A float tensor with the same shape as inputs. Stores the binary
classification label for each element in inputs
(0 for the negative class and 1 for the positive class).
alpha: (optional) Weighting factor in range (0,1) to balance
positive vs negative examples. Default = -1 (no weighting).
gamma: Exponent of the modulating factor (1 - p_t) to
balance easy vs hard examples.
Returns:
Loss tensor
"""
prob = inputs.sigmoid()
ce_loss = F.binary_cross_entropy_with_logits(inputs, targets, reduction="none")
p_t = prob * targets + (1 - prob) * (1 - targets)
loss = ce_loss * ((1 - p_t) ** gamma)
if alpha >= 0:
alpha_t = alpha * targets + (1 - alpha) * (1 - targets)
loss = alpha_t * loss
return loss.mean(1).sum() / num_boxes
class PostProcessSegm(nn.Module):
def __init__(self, threshold=0.5):
super().__init__()
self.threshold = threshold
@torch.no_grad()
def forward(self, results, outputs, orig_target_sizes, max_target_sizes):
assert len(orig_target_sizes) == len(max_target_sizes)
max_h, max_w = max_target_sizes.max(0)[0].tolist()
outputs_masks = outputs["pred_masks"].squeeze(2)
outputs_masks = F.interpolate(outputs_masks, size=(max_h, max_w), mode="bilinear", align_corners=False)
outputs_masks = (outputs_masks.sigmoid() > self.threshold).cpu()
for i, (cur_mask, t, tt) in enumerate(zip(outputs_masks, max_target_sizes, orig_target_sizes)):
img_h, img_w = t[0], t[1]
results[i]["masks"] = cur_mask[:, :img_h, :img_w].unsqueeze(1)
results[i]["masks"] = F.interpolate(
results[i]["masks"].float(), size=tuple(tt.tolist()), mode="nearest"
).byte()
return results
class PostProcessPanoptic(nn.Module):
"""This class converts the output of the model to the final panoptic result, in the format expected by the
coco panoptic API """
def __init__(self, is_thing_map, threshold=0.85):
"""
Parameters:
is_thing_map: This is a whose keys are the class ids, and the values a boolean indicating whether
the class is a thing (True) or a stuff (False) class
threshold: confidence threshold: segments with confidence lower than this will be deleted
"""
super().__init__()
self.threshold = threshold
self.is_thing_map = is_thing_map
def forward(self, outputs, processed_sizes, target_sizes=None):
""" This function computes the panoptic prediction from the model's predictions.
Parameters:
outputs: This is a dict coming directly from the model. See the model doc for the content.
processed_sizes: This is a list of tuples (or torch tensors) of sizes of the images that were passed to the
model, ie the size after data augmentation but before batching.
target_sizes: This is a list of tuples (or torch tensors) corresponding to the requested final size
of each prediction. If left to None, it will default to the processed_sizes
"""
if target_sizes is None:
target_sizes = processed_sizes
assert len(processed_sizes) == len(target_sizes)
out_logits, raw_masks, raw_boxes = outputs["pred_logits"], outputs["pred_masks"], outputs["pred_boxes"]
assert len(out_logits) == len(raw_masks) == len(target_sizes)
preds = []
def to_tuple(tup):
if isinstance(tup, tuple):
return tup
return tuple(tup.cpu().tolist())
for cur_logits, cur_masks, cur_boxes, size, target_size in zip(
out_logits, raw_masks, raw_boxes, processed_sizes, target_sizes
):
# we filter empty queries and detection below threshold
scores, labels = cur_logits.softmax(-1).max(-1)
keep = labels.ne(outputs["pred_logits"].shape[-1] - 1) & (scores > self.threshold)
cur_scores, cur_classes = cur_logits.softmax(-1).max(-1)
cur_scores = cur_scores[keep]
cur_classes = cur_classes[keep]
cur_masks = cur_masks[keep]
cur_masks = interpolate(cur_masks[:, None], to_tuple(size), mode="bilinear").squeeze(1)
cur_boxes = box_ops.box_cxcywh_to_xyxy(cur_boxes[keep])
h, w = cur_masks.shape[-2:]
assert len(cur_boxes) == len(cur_classes)
# It may be that we have several predicted masks for the same stuff class.
# In the following, we track the list of masks ids for each stuff class (they are merged later on)
cur_masks = cur_masks.flatten(1)
stuff_equiv_classes = defaultdict(lambda: [])
for k, label in enumerate(cur_classes):
if not self.is_thing_map[label.item()]:
stuff_equiv_classes[label.item()].append(k)
def get_ids_area(masks, scores, dedup=False):
# This helper function creates the final panoptic segmentation image
# It also returns the area of the masks that appears on the image
m_id = masks.transpose(0, 1).softmax(-1)
if m_id.shape[-1] == 0:
# We didn't detect any mask :(
m_id = torch.zeros((h, w), dtype=torch.long, device=m_id.device)
else:
m_id = m_id.argmax(-1).view(h, w)
if dedup:
# Merge the masks corresponding to the same stuff class
for equiv in stuff_equiv_classes.values():
if len(equiv) > 1:
for eq_id in equiv:
m_id.masked_fill_(m_id.eq(eq_id), equiv[0])
final_h, final_w = to_tuple(target_size)
seg_img = Image.fromarray(id2rgb(m_id.view(h, w).cpu().numpy()))
seg_img = seg_img.resize(size=(final_w, final_h), resample=Image.NEAREST)
np_seg_img = (
torch.ByteTensor(torch.ByteStorage.from_buffer(seg_img.tobytes())).view(final_h, final_w, 3).numpy()
)
m_id = torch.from_numpy(rgb2id(np_seg_img))
area = []
for i in range(len(scores)):
area.append(m_id.eq(i).sum().item())
return area, seg_img
area, seg_img = get_ids_area(cur_masks, cur_scores, dedup=True)
if cur_classes.numel() > 0:
# We know filter empty masks as long as we find some
while True:
filtered_small = torch.as_tensor(
[area[i] <= 4 for i, c in enumerate(cur_classes)], dtype=torch.bool, device=keep.device
)
if filtered_small.any().item():
cur_scores = cur_scores[~filtered_small]
cur_classes = cur_classes[~filtered_small]
cur_masks = cur_masks[~filtered_small]
area, seg_img = get_ids_area(cur_masks, cur_scores)
else:
break
else:
cur_classes = torch.ones(1, dtype=torch.long, device=cur_classes.device)
segments_info = []
for i, a in enumerate(area):
cat = cur_classes[i].item()
segments_info.append({"id": i, "isthing": self.is_thing_map[cat], "category_id": cat, "area": a})
del cur_classes
with io.BytesIO() as out:
seg_img.save(out, format="PNG")
predictions = {"png_string": out.getvalue(), "segments_info": segments_info}
preds.append(predictions)
return preds
================================================
FILE: models/transformer.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
"""
DETR Transformer class.
Copy-paste from torch.nn.Transformer with modifications:
* positional encodings are passed in MHattention
* extra LN at the end of encoder is removed
* decoder returns a stack of activations from all decoding layers
"""
import copy
from typing import Optional, List
import torch
import torch.nn.functional as F
from torch import nn, Tensor
class Transformer(nn.Module):
def __init__(self, d_model=512, nhead=8, num_encoder_layers=6,
num_decoder_layers=6, dim_feedforward=2048, dropout=0.1,
activation="relu", normalize_before=False,
return_intermediate_dec=False):
super().__init__()
encoder_layer = TransformerEncoderLayer(d_model, nhead, dim_feedforward,
dropout, activation, normalize_before)
encoder_norm = nn.LayerNorm(d_model) if normalize_before else None
self.encoder = TransformerEncoder(encoder_layer, num_encoder_layers, encoder_norm)
decoder_layer = TransformerDecoderLayer(d_model, nhead, dim_feedforward,
dropout, activation, normalize_before)
decoder_norm = nn.LayerNorm(d_model)
self.decoder = TransformerDecoder(decoder_layer, num_decoder_layers, decoder_norm,
return_intermediate=return_intermediate_dec)
self._reset_parameters()
self.d_model = d_model
self.nhead = nhead
def _reset_parameters(self):
for p in self.parameters():
if p.dim() > 1:
nn.init.xavier_uniform_(p)
def forward(self, src, mask, query_embed, pos_embed):
# flatten NxCxHxW to HWxNxC
bs, c, h, w = src.shape
src = src.flatten(2).permute(2, 0, 1)
pos_embed = pos_embed.flatten(2).permute(2, 0, 1)
query_embed = query_embed.unsqueeze(1).repeat(1, bs, 1)
mask = mask.flatten(1)
tgt = torch.zeros_like(query_embed)
memory = self.encoder(src, src_key_padding_mask=mask, pos=pos_embed)
hs = self.decoder(tgt, memory, memory_key_padding_mask=mask,
pos=pos_embed, query_pos=query_embed)
return hs.transpose(1, 2), memory.permute(1, 2, 0).view(bs, c, h, w)
class TransformerEncoder(nn.Module):
def __init__(self, encoder_layer, num_layers, norm=None):
super().__init__()
self.layers = _get_clones(encoder_layer, num_layers)
self.num_layers = num_layers
self.norm = norm
def forward(self, src,
mask: Optional[Tensor] = None,
src_key_padding_mask: Optional[Tensor] = None,
pos: Optional[Tensor] = None):
output = src
for layer in self.layers:
output = layer(output, src_mask=mask,
src_key_padding_mask=src_key_padding_mask, pos=pos)
if self.norm is not None:
output = self.norm(output)
return output
class TransformerDecoder(nn.Module):
def __init__(self, decoder_layer, num_layers, norm=None, return_intermediate=False):
super().__init__()
self.layers = _get_clones(decoder_layer, num_layers)
self.num_layers = num_layers
self.norm = norm
self.return_intermediate = return_intermediate
def forward(self, tgt, memory,
tgt_mask: Optional[Tensor] = None,
memory_mask: Optional[Tensor] = None,
tgt_key_padding_mask: Optional[Tensor] = None,
memory_key_padding_mask: Optional[Tensor] = None,
pos: Optional[Tensor] = None,
query_pos: Optional[Tensor] = None):
output = tgt
intermediate = []
for layer in self.layers:
output = layer(output, memory, tgt_mask=tgt_mask,
memory_mask=memory_mask,
tgt_key_padding_mask=tgt_key_padding_mask,
memory_key_padding_mask=memory_key_padding_mask,
pos=pos, query_pos=query_pos)
if self.return_intermediate:
intermediate.append(self.norm(output))
if self.norm is not None:
output = self.norm(output)
if self.return_intermediate:
intermediate.pop()
intermediate.append(output)
if self.return_intermediate:
return torch.stack(intermediate)
return output.unsqueeze(0)
class TransformerEncoderLayer(nn.Module):
def __init__(self, d_model, nhead, dim_feedforward=2048, dropout=0.1,
activation="relu", normalize_before=False):
super().__init__()
self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout)
# Implementation of Feedforward model
self.linear1 = nn.Linear(d_model, dim_feedforward)
self.dropout = nn.Dropout(dropout)
self.linear2 = nn.Linear(dim_feedforward, d_model)
self.norm1 = nn.LayerNorm(d_model)
self.norm2 = nn.LayerNorm(d_model)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
self.activation = _get_activation_fn(activation)
self.normalize_before = normalize_before
def with_pos_embed(self, tensor, pos: Optional[Tensor]):
return tensor if pos is None else tensor + pos
def forward_post(self,
src,
src_mask: Optional[Tensor] = None,
src_key_padding_mask: Optional[Tensor] = None,
pos: Optional[Tensor] = None):
q = k = self.with_pos_embed(src, pos)
src2 = self.self_attn(q, k, value=src, attn_mask=src_mask,
key_padding_mask=src_key_padding_mask)[0]
src = src + self.dropout1(src2)
src = self.norm1(src)
src2 = self.linear2(self.dropout(self.activation(self.linear1(src))))
src = src + self.dropout2(src2)
src = self.norm2(src)
return src
def forward_pre(self, src,
src_mask: Optional[Tensor] = None,
src_key_padding_mask: Optional[Tensor] = None,
pos: Optional[Tensor] = None):
src2 = self.norm1(src)
q = k = self.with_pos_embed(src2, pos)
src2 = self.self_attn(q, k, value=src2, attn_mask=src_mask,
key_padding_mask=src_key_padding_mask)[0]
src = src + self.dropout1(src2)
src2 = self.norm2(src)
src2 = self.linear2(self.dropout(self.activation(self.linear1(src2))))
src = src + self.dropout2(src2)
return src
def forward(self, src,
src_mask: Optional[Tensor] = None,
src_key_padding_mask: Optional[Tensor] = None,
pos: Optional[Tensor] = None):
if self.normalize_before:
return self.forward_pre(src, src_mask, src_key_padding_mask, pos)
return self.forward_post(src, src_mask, src_key_padding_mask, pos)
class TransformerDecoderLayer(nn.Module):
def __init__(self, d_model, nhead, dim_feedforward=2048, dropout=0.1,
activation="relu", normalize_before=False):
super().__init__()
self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout)
self.multihead_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout)
# Implementation of Feedforward model
self.linear1 = nn.Linear(d_model, dim_feedforward)
self.dropout = nn.Dropout(dropout)
self.linear2 = nn.Linear(dim_feedforward, d_model)
self.norm1 = nn.LayerNorm(d_model)
self.norm2 = nn.LayerNorm(d_model)
self.norm3 = nn.LayerNorm(d_model)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
self.dropout3 = nn.Dropout(dropout)
self.activation = _get_activation_fn(activation)
self.normalize_before = normalize_before
def with_pos_embed(self, tensor, pos: Optional[Tensor]):
return tensor if pos is None else tensor + pos
def forward_post(self, tgt, memory,
tgt_mask: Optional[Tensor] = None,
memory_mask: Optional[Tensor] = None,
tgt_key_padding_mask: Optional[Tensor] = None,
memory_key_padding_mask: Optional[Tensor] = None,
pos: Optional[Tensor] = None,
query_pos: Optional[Tensor] = None):
q = k = self.with_pos_embed(tgt, query_pos)
tgt2 = self.self_attn(q, k, value=tgt, attn_mask=tgt_mask,
key_padding_mask=tgt_key_padding_mask)[0]
tgt = tgt + self.dropout1(tgt2)
tgt = self.norm1(tgt)
tgt2 = self.multihead_attn(query=self.with_pos_embed(tgt, query_pos),
key=self.with_pos_embed(memory, pos),
value=memory, attn_mask=memory_mask,
key_padding_mask=memory_key_padding_mask)[0]
tgt = tgt + self.dropout2(tgt2)
tgt = self.norm2(tgt)
tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt))))
tgt = tgt + self.dropout3(tgt2)
tgt = self.norm3(tgt)
return tgt
def forward_pre(self, tgt, memory,
tgt_mask: Optional[Tensor] = None,
memory_mask: Optional[Tensor] = None,
tgt_key_padding_mask: Optional[Tensor] = None,
memory_key_padding_mask: Optional[Tensor] = None,
pos: Optional[Tensor] = None,
query_pos: Optional[Tensor] = None):
tgt2 = self.norm1(tgt)
q = k = self.with_pos_embed(tgt2, query_pos)
tgt2 = self.self_attn(q, k, value=tgt2, attn_mask=tgt_mask,
key_padding_mask=tgt_key_padding_mask)[0]
tgt = tgt + self.dropout1(tgt2)
tgt2 = self.norm2(tgt)
tgt2 = self.multihead_attn(query=self.with_pos_embed(tgt2, query_pos),
key=self.with_pos_embed(memory, pos),
value=memory, attn_mask=memory_mask,
key_padding_mask=memory_key_padding_mask)[0]
tgt = tgt + self.dropout2(tgt2)
tgt2 = self.norm3(tgt)
tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt2))))
tgt = tgt + self.dropout3(tgt2)
return tgt
def forward(self, tgt, memory,
tgt_mask: Optional[Tensor] = None,
memory_mask: Optional[Tensor] = None,
tgt_key_padding_mask: Optional[Tensor] = None,
memory_key_padding_mask: Optional[Tensor] = None,
pos: Optional[Tensor] = None,
query_pos: Optional[Tensor] = None):
if self.normalize_before:
return self.forward_pre(tgt, memory, tgt_mask, memory_mask,
tgt_key_padding_mask, memory_key_padding_mask, pos, query_pos)
return self.forward_post(tgt, memory, tgt_mask, memory_mask,
tgt_key_padding_mask, memory_key_padding_mask, pos, query_pos)
def _get_clones(module, N):
return nn.ModuleList([copy.deepcopy(module) for i in range(N)])
def build_transformer(args):
return Transformer(
d_model=args.hidden_dim,
dropout=args.dropout,
nhead=args.nheads,
dim_feedforward=args.dim_feedforward,
num_encoder_layers=args.enc_layers,
num_decoder_layers=args.dec_layers,
normalize_before=args.pre_norm,
return_intermediate_dec=True,
)
def _get_activation_fn(activation):
"""Return an activation function given a string"""
if activation == "relu":
return F.relu
if activation == "gelu":
return F.gelu
if activation == "glu":
return F.glu
raise RuntimeError(F"activation should be relu/gelu, not {activation}.")
================================================
FILE: requirements.txt
================================================
cython
git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI&egg=pycocotools
submitit
torch>=1.5.0
torchvision>=0.6.0
git+https://github.com/cocodataset/panopticapi.git#egg=panopticapi
scipy
onnx
onnxruntime
================================================
FILE: run_with_submitit.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
"""
A script to run multinode training with submitit.
"""
import argparse
import os
import uuid
from pathlib import Path
import main as detection
import submitit
def parse_args():
detection_parser = detection.get_args_parser()
parser = argparse.ArgumentParser("Submitit for detection", parents=[detection_parser])
parser.add_argument("--ngpus", default=8, type=int, help="Number of gpus to request on each node")
parser.add_argument("--nodes", default=4, type=int, help="Number of nodes to request")
parser.add_argument("--timeout", default=60, type=int, help="Duration of the job")
parser.add_argument("--job_dir", default="", type=str, help="Job dir. Leave empty for automatic.")
return parser.parse_args()
def get_shared_folder() -> Path:
user = os.getenv("USER")
if Path("/checkpoint/").is_dir():
p = Path(f"/checkpoint/{user}/experiments")
p.mkdir(exist_ok=True)
return p
raise RuntimeError("No shared folder available")
def get_init_file():
# Init file must not exist, but it's parent dir must exist.
os.makedirs(str(get_shared_folder()), exist_ok=True)
init_file = get_shared_folder() / f"{uuid.uuid4().hex}_init"
if init_file.exists():
os.remove(str(init_file))
return init_file
class Trainer(object):
def __init__(self, args):
self.args = args
def __call__(self):
import main as detection
self._setup_gpu_args()
detection.main(self.args)
def checkpoint(self):
import os
import submitit
from pathlib import Path
self.args.dist_url = get_init_file().as_uri()
checkpoint_file = os.path.join(self.args.output_dir, "checkpoint.pth")
if os.path.exists(checkpoint_file):
self.args.resume = checkpoint_file
print("Requeuing ", self.args)
empty_trainer = type(self)(self.args)
return submitit.helpers.DelayedSubmission(empty_trainer)
def _setup_gpu_args(self):
import submitit
from pathlib import Path
job_env = submitit.JobEnvironment()
self.args.output_dir = Path(str(self.args.output_dir).replace("%j", str(job_env.job_id)))
self.args.gpu = job_env.local_rank
self.args.rank = job_env.global_rank
self.args.world_size = job_env.num_tasks
print(f"Process group: {job_env.num_tasks} tasks, rank: {job_env.global_rank}")
def main():
args = parse_args()
if args.job_dir == "":
args.job_dir = get_shared_folder() / "%j"
# Note that the folder will depend on the job_id, to easily track experiments
executor = submitit.AutoExecutor(folder=args.job_dir, slurm_max_num_timeout=30)
# cluster setup is defined by environment variables
num_gpus_per_node = args.ngpus
nodes = args.nodes
timeout_min = args.timeout
executor.update_parameters(
mem_gb=40 * num_gpus_per_node,
gpus_per_node=num_gpus_per_node,
tasks_per_node=num_gpus_per_node, # one task per GPU
cpus_per_task=10,
nodes=nodes,
timeout_min=timeout_min, # max is 60 * 72
)
executor.update_parameters(name="detr")
args.dist_url = get_init_file().as_uri()
args.output_dir = args.job_dir
trainer = Trainer(args)
job = executor.submit(trainer)
print("Submitted job_id:", job.job_id)
if __name__ == "__main__":
main()
================================================
FILE: test_all.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
import io
import unittest
import torch
from torch import nn, Tensor
from typing import List
from models.matcher import HungarianMatcher
from models.position_encoding import PositionEmbeddingSine, PositionEmbeddingLearned
from models.backbone import Backbone, Joiner, BackboneBase
from util import box_ops
from util.misc import nested_tensor_from_tensor_list
from hubconf import detr_resnet50, detr_resnet50_panoptic
# onnxruntime requires python 3.5 or above
try:
import onnxruntime
except ImportError:
onnxruntime = None
class Tester(unittest.TestCase):
def test_box_cxcywh_to_xyxy(self):
t = torch.rand(10, 4)
r = box_ops.box_xyxy_to_cxcywh(box_ops.box_cxcywh_to_xyxy(t))
self.assertLess((t - r).abs().max(), 1e-5)
@staticmethod
def indices_torch2python(indices):
return [(i.tolist(), j.tolist()) for i, j in indices]
def test_hungarian(self):
n_queries, n_targets, n_classes = 100, 15, 91
logits = torch.rand(1, n_queries, n_classes + 1)
boxes = torch.rand(1, n_queries, 4)
tgt_labels = torch.randint(high=n_classes, size=(n_targets,))
tgt_boxes = torch.rand(n_targets, 4)
matcher = HungarianMatcher()
targets = [{'labels': tgt_labels, 'boxes': tgt_boxes}]
indices_single = matcher({'pred_logits': logits, 'pred_boxes': boxes}, targets)
indices_batched = matcher({'pred_logits': logits.repeat(2, 1, 1),
'pred_boxes': boxes.repeat(2, 1, 1)}, targets * 2)
self.assertEqual(len(indices_single[0][0]), n_targets)
self.assertEqual(len(indices_single[0][1]), n_targets)
self.assertEqual(self.indices_torch2python(indices_single),
self.indices_torch2python([indices_batched[0]]))
self.assertEqual(self.indices_torch2python(indices_single),
self.indices_torch2python([indices_batched[1]]))
# test with empty targets
tgt_labels_empty = torch.randint(high=n_classes, size=(0,))
tgt_boxes_empty = torch.rand(0, 4)
targets_empty = [{'labels': tgt_labels_empty, 'boxes': tgt_boxes_empty}]
indices = matcher({'pred_logits': logits.repeat(2, 1, 1),
'pred_boxes': boxes.repeat(2, 1, 1)}, targets + targets_empty)
self.assertEqual(len(indices[1][0]), 0)
indices = matcher({'pred_logits': logits.repeat(2, 1, 1),
'pred_boxes': boxes.repeat(2, 1, 1)}, targets_empty * 2)
self.assertEqual(len(indices[0][0]), 0)
def test_position_encoding_script(self):
m1, m2 = PositionEmbeddingSine(), PositionEmbeddingLearned()
mm1, mm2 = torch.jit.script(m1), torch.jit.script(m2) # noqa
def test_backbone_script(self):
backbone = Backbone('resnet50', True, False, False)
torch.jit.script(backbone) # noqa
def test_model_script_detection(self):
model = detr_resnet50(pretrained=False).eval()
scripted_model = torch.jit.script(model)
x = nested_tensor_from_tensor_list([torch.rand(3, 200, 200), torch.rand(3, 200, 250)])
out = model(x)
out_script = scripted_model(x)
self.assertTrue(out["pred_logits"].equal(out_script["pred_logits"]))
self.assertTrue(out["pred_boxes"].equal(out_script["pred_boxes"]))
def test_model_script_panoptic(self):
model = detr_resnet50_panoptic(pretrained=False).eval()
scripted_model = torch.jit.script(model)
x = nested_tensor_from_tensor_list([torch.rand(3, 200, 200), torch.rand(3, 200, 250)])
out = model(x)
out_script = scripted_model(x)
self.assertTrue(out["pred_logits"].equal(out_script["pred_logits"]))
self.assertTrue(out["pred_boxes"].equal(out_script["pred_boxes"]))
self.assertTrue(out["pred_masks"].equal(out_script["pred_masks"]))
def test_model_detection_different_inputs(self):
model = detr_resnet50(pretrained=False).eval()
# support NestedTensor
x = nested_tensor_from_tensor_list([torch.rand(3, 200, 200), torch.rand(3, 200, 250)])
out = model(x)
self.assertIn('pred_logits', out)
# and 4d Tensor
x = torch.rand(1, 3, 200, 200)
out = model(x)
self.assertIn('pred_logits', out)
# and List[Tensor[C, H, W]]
x = torch.rand(3, 200, 200)
out = model([x])
self.assertIn('pred_logits', out)
def test_warpped_model_script_detection(self):
class WrappedDETR(nn.Module):
def __init__(self, model):
super().__init__()
self.model = model
def forward(self, inputs: List[Tensor]):
sample = nested_tensor_from_tensor_list(inputs)
return self.model(sample)
model = detr_resnet50(pretrained=False)
wrapped_model = WrappedDETR(model)
wrapped_model.eval()
scripted_model = torch.jit.script(wrapped_model)
x = [torch.rand(3, 200, 200), torch.rand(3, 200, 250)]
out = wrapped_model(x)
out_script = scripted_model(x)
self.assertTrue(out["pred_logits"].equal(out_script["pred_logits"]))
self.assertTrue(out["pred_boxes"].equal(out_script["pred_boxes"]))
@unittest.skipIf(onnxruntime is None, 'ONNX Runtime unavailable')
class ONNXExporterTester(unittest.TestCase):
@classmethod
def setUpClass(cls):
torch.manual_seed(123)
def run_model(self, model, inputs_list, tolerate_small_mismatch=False, do_constant_folding=True, dynamic_axes=None,
output_names=None, input_names=None):
model.eval()
onnx_io = io.BytesIO()
# export to onnx with the first input
torch.onnx.export(model, inputs_list[0], onnx_io,
do_constant_folding=do_constant_folding, opset_version=12,
dynamic_axes=dynamic_axes, input_names=input_names, output_names=output_names)
# validate the exported model with onnx runtime
for test_inputs in inputs_list:
with torch.no_grad():
if isinstance(test_inputs, torch.Tensor) or isinstance(test_inputs, list):
test_inputs = (nested_tensor_from_tensor_list(test_inputs),)
test_ouputs = model(*test_inputs)
if isinstance(test_ouputs, torch.Tensor):
test_ouputs = (test_ouputs,)
self.ort_validate(onnx_io, test_inputs, test_ouputs, tolerate_small_mismatch)
def ort_validate(self, onnx_io, inputs, outputs, tolerate_small_mismatch=False):
inputs, _ = torch.jit._flatten(inputs)
outputs, _ = torch.jit._flatten(outputs)
def to_numpy(tensor):
if tensor.requires_grad:
return tensor.detach().cpu().numpy()
else:
return tensor.cpu().numpy()
inputs = list(map(to_numpy, inputs))
outputs = list(map(to_numpy, outputs))
ort_session = onnxruntime.InferenceSession(onnx_io.getvalue())
# compute onnxruntime output prediction
ort_inputs = dict((ort_session.get_inputs()[i].name, inpt) for i, inpt in enumerate(inputs))
ort_outs = ort_session.run(None, ort_inputs)
for i, element in enumerate(outputs):
try:
torch.testing.assert_allclose(element, ort_outs[i], rtol=1e-03, atol=1e-05)
except AssertionError as error:
if tolerate_small_mismatch:
self.assertIn("(0.00%)", str(error), str(error))
else:
raise
def test_model_onnx_detection(self):
model = detr_resnet50(pretrained=False).eval()
dummy_image = torch.ones(1, 3, 800, 800) * 0.3
model(dummy_image)
# Test exported model on images of different size, or dummy input
self.run_model(
model,
[(torch.rand(1, 3, 750, 800),)],
input_names=["inputs"],
output_names=["pred_logits", "pred_boxes"],
tolerate_small_mismatch=True,
)
@unittest.skip("CI doesn't have enough memory")
def test_model_onnx_detection_panoptic(self):
model = detr_resnet50_panoptic(pretrained=False).eval()
dummy_image = torch.ones(1, 3, 800, 800) * 0.3
model(dummy_image)
# Test exported model on images of different size, or dummy input
self.run_model(
model,
[(torch.rand(1, 3, 750, 800),)],
input_names=["inputs"],
output_names=["pred_logits", "pred_boxes", "pred_masks"],
tolerate_small_mismatch=True,
)
if __name__ == '__main__':
unittest.main()
================================================
FILE: tox.ini
================================================
[flake8]
max-line-length = 120
ignore = F401,E402,F403,W503,W504
================================================
FILE: util/__init__.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
================================================
FILE: util/box_ops.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
"""
Utilities for bounding box manipulation and GIoU.
"""
import torch
from torchvision.ops.boxes import box_area
def box_cxcywh_to_xyxy(x):
x_c, y_c, w, h = x.unbind(-1)
b = [(x_c - 0.5 * w), (y_c - 0.5 * h),
(x_c + 0.5 * w), (y_c + 0.5 * h)]
return torch.stack(b, dim=-1)
def box_xyxy_to_cxcywh(x):
x0, y0, x1, y1 = x.unbind(-1)
b = [(x0 + x1) / 2, (y0 + y1) / 2,
(x1 - x0), (y1 - y0)]
return torch.stack(b, dim=-1)
# modified from torchvision to also return the union
def box_iou(boxes1, boxes2):
area1 = box_area(boxes1)
area2 = box_area(boxes2)
lt = torch.max(boxes1[:, None, :2], boxes2[:, :2]) # [N,M,2]
rb = torch.min(boxes1[:, None, 2:], boxes2[:, 2:]) # [N,M,2]
wh = (rb - lt).clamp(min=0) # [N,M,2]
inter = wh[:, :, 0] * wh[:, :, 1] # [N,M]
union = area1[:, None] + area2 - inter
iou = inter / union
return iou, union
def generalized_box_iou(boxes1, boxes2):
"""
Generalized IoU from https://giou.stanford.edu/
The boxes should be in [x0, y0, x1, y1] format
Returns a [N, M] pairwise matrix, where N = len(boxes1)
and M = len(boxes2)
"""
# degenerate boxes gives inf / nan results
# so do an early check
assert (boxes1[:, 2:] >= boxes1[:, :2]).all()
assert (boxes2[:, 2:] >= boxes2[:, :2]).all()
iou, union = box_iou(boxes1, boxes2)
lt = torch.min(boxes1[:, None, :2], boxes2[:, :2])
rb = torch.max(boxes1[:, None, 2:], boxes2[:, 2:])
wh = (rb - lt).clamp(min=0) # [N,M,2]
area = wh[:, :, 0] * wh[:, :, 1]
return iou - (area - union) / area
def masks_to_boxes(masks):
"""Compute the bounding boxes around the provided masks
The masks should be in format [N, H, W] where N is the number of masks, (H, W) are the spatial dimensions.
Returns a [N, 4] tensors, with the boxes in xyxy format
"""
if masks.numel() == 0:
return torch.zeros((0, 4), device=masks.device)
h, w = masks.shape[-2:]
y = torch.arange(0, h, dtype=torch.float)
x = torch.arange(0, w, dtype=torch.float)
y, x = torch.meshgrid(y, x)
x_mask = (masks * x.unsqueeze(0))
x_max = x_mask.flatten(1).max(-1)[0]
x_min = x_mask.masked_fill(~(masks.bool()), 1e8).flatten(1).min(-1)[0]
y_mask = (masks * y.unsqueeze(0))
y_max = y_mask.flatten(1).max(-1)[0]
y_min = y_mask.masked_fill(~(masks.bool()), 1e8).flatten(1).min(-1)[0]
return torch.stack([x_min, y_min, x_max, y_max], 1)
================================================
FILE: util/misc.py
================================================
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
"""
Misc functions, including distributed helpers.
Mostly copy-paste from torchvision references.
"""
import os
import subprocess
import time
from collections import defaultdict, deque
import datetime
import pickle
from packaging import version
from typing import Optional, List
import torch
import torch.distributed as dist
from torch import Tensor
# needed due to empty tensor bug in pytorch and torchvision 0.5
import torchvision
if version.parse(torchvision.__version__) < version.parse('0.7'):
from torchvision.ops import _new_empty_tensor
from torchvision.ops.misc import _output_size
class SmoothedValue(object):
"""Track a series of values and provide access to smoothed values over a
window or the global series average.
"""
def __init__(self, window_size=20, fmt=None):
if fmt is None:
fmt = "{median:.4f} ({global_avg:.4f})"
self.deque = deque(maxlen=window_size)
self.total = 0.0
self.count = 0
self.fmt = fmt
def update(self, value, n=1):
self.deque.append(value)
self.count += n
self.total += value * n
def synchronize_between_processes(self):
"""
Warning: does not synchronize the deque!
"""
if not is_dist_avail_and_initialized():
return
t = torch.tensor([self.count, self.total], dtype=torch.float64, device='cuda')
dist.barrier()
dist.all_reduce(t)
t = t.tolist()
self.count = int(t[0])
self.total = t[1]
@property
def median(self):
d = torch.tensor(list(self.deque))
return d.median().item()
@property
def avg(self):
d = torch.tensor(list(self.deque), dtype=torch.float32)
return d.mean().item()
@property
def global_avg(self):
return self.total / self.count
@property
def max(self):
return max(self.deque)
@property
def value(self):
return self.deque[-1]
def __str__(self):
return self.fmt.format(
median=self.median,
avg=self.avg,
global_avg=self.global_avg,
max=self.max,
value=self.value)
def all_gather(data):
"""
Run all_gather on arbitrary picklable data (not necessarily tensors)
Args:
data: any picklable object
Returns:
list[data]: list of data gathered from each rank
"""
world_size = get_world_size()
if world_size == 1:
return [data]
# serialized to a Tensor
buffer = pickle.dumps(data)
storage = torch.ByteStorage.from_buffer(buffer)
tensor = torch.ByteTensor(storage).to("cuda")
# obtain Tensor size of each rank
local_size = torch.tensor([tensor.numel()], device="cuda")
size_list = [torch.tensor([0], device="cuda") for _ in range(world_size)]
dist.all_gather(size_list, local_size)
size_list = [int(size.item()) for size in size_list]
max_size = max(size_list)
# receiving Tensor from all ranks
# we pad the tensor because torch all_gather does not support
# gathering tensors of different shapes
tensor_list = []
for _ in size_list:
tensor_list.append(torch.empty((max_size,), dtype=torch.uint8, device="cuda"))
if local_size != max_size:
padding = torch.empty(size=(max_size - local_size,), dtype=torch.uint8, device="cuda")
tensor = torch.cat((tensor, padding), dim=0)
dist.all_gather(tensor_list, tensor)
data_list = []
for size, tensor in zip(size_list, tensor_list):
buffer = tensor.cpu().numpy().tobytes()[:size]
data_list.append(pickle.loads(buffer))
return data_list
def reduce_dict(input_dict, average=True):
"""
Args:
input_dict (dict): all the values will be reduced
average (bool): whether to do average or sum
Reduce the values in the dictionary from all processes so that all processes
have the averaged results. Returns a dict with the same fields as
input_dict, after reduction.
"""
world_size = get_world_size()
if world_size < 2:
return input_dict
with torch.no_grad():
names = []
values = []
# sort the keys so that they are consistent across processes
for k in sorted(input_dict.keys()):
names.append(k)
values.append(input_dict[k])
values = torch.stack(values, dim=0)
dist.all_reduce(values)
if average:
values /= world_size
reduced_dict = {k: v for k, v in zip(names, values)}
return reduced_dict
class MetricLogger(object):
def __init__(self, delimiter="\t"):
self.meters = defaultdict(SmoothedValue)
self.delimiter = delimiter
def update(self, **kwargs):
for k, v in kwargs.items():
if isinstance(v, torch.Tensor):
v = v.item()
assert isinstance(v, (float, int))
self.meters[k].update(v)
def __getattr__(self, attr):
if attr in self.meters:
return self.meters[attr]
if attr in self.__dict__:
return self.__dict__[attr]
raise AttributeError("'{}' object has no attribute '{}'".format(
type(self).__name__, attr))
def __str__(self):
loss_str = []
for name, meter in self.meters.items():
loss_str.append(
"{}: {}".format(name, str(meter))
)
return self.delimiter.join(loss_str)
def synchronize_between_processes(self):
for meter in self.meters.values():
meter.synchronize_between_processes()
def add_meter(self, name, meter):
self.meters[name] = meter
def log_every(self, iterable, print_freq, header=None):
i = 0
if not header:
header = ''
start_time = time.time()
end = time.time()
iter_time = SmoothedValue(fmt='{avg:.4f}')
data_time = SmoothedValue(fmt='{avg:.4f}')
space_fmt = ':' + str(len(str(len(iterable)))) + 'd'
if torch.cuda.is_available():
log_msg = self.delimiter.join([
header,
'[{0' + space_fmt + '}/{1}]',
'eta: {eta}',
'{meters}',
'time: {time}',
'data: {data}',
'max mem: {memory:.0f}'
])
else:
log_msg = self.delimiter.join([
header,
'[{0' + space_fmt + '}/{1}]',
'eta: {eta}',
'{meters}',
'time: {time}',
'data: {data}'
])
MB = 1024.0 * 1024.0
for obj in iterable:
data_time.update(time.time() - end)
yield obj
iter_time.update(time.time() - end)
if i % print_freq == 0 or i == len(iterable) - 1:
eta_seconds = iter_time.global_avg * (len(iterable) - i)
eta_string = str(datetime.timedelta(seconds=int(eta_seconds)))
if torch.cuda.is_available():
print(log_msg.format(
i, len(iterable), eta=eta_string,
meters=str(self),
time=str(iter_time), data=str(data_time),
memory=torch.cuda.max_memory_allocated() / MB))
else:
print(log_msg.format(
i, len(iterable), eta=eta_string,
meters=str(self),
time=str(iter_time), data=str(data_time)))
i += 1
end = time.time()
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('{} Total time: {} ({:.4f} s / it)'.format(
header, total_time_str, total_time / len(iterable)))
def get_sha():
cwd = os.path.dirname(os.path.abspath(__file__))
def _run(command):
return subprocess.check_output(command, cwd=cwd).decode('ascii').strip()
sha = 'N/A'
diff = "clean"
branch = 'N/A'
try:
sha = _run(['git', 'rev-parse', 'HEAD'])
subprocess.check_output(['git', 'diff'], cwd=cwd)
diff = _run(['git', 'diff-index', 'HEAD'])
diff = "has uncommited changes" if diff else "clean"
branch = _run(['git', 'rev-parse', '--abbrev-ref', 'HEAD'])
except Exception:
pass
message = f"sha: {sha}, status: {diff}, branch: {branch}"
return message
def collate_fn(batch):
batch = list(zip(*batch))
batch[0] = nested_tensor_from_tensor_list(batch[0])
return tuple(batch)
def _max_by_axis(the_list):
# type: (List[List[int]]) -> List[int]
maxes = the_list[0]
for sublist in the_list[1:]:
for index, item in enumerate(sublist):
maxes[index] = max(maxes[index], item)
return maxes
class NestedTensor(object):
def __init__(self, tensors, mask: Optional[Tensor]):
self.tensors = tensors
self.mask = mask
def to(self, device):
# type: (Device) -> NestedTensor # noqa
cast_tensor = self.tensors.to(device)
mask = self.mask
if mask is not None:
assert mask is not None
cast_mask = mask.to(device)
else:
cast_mask = None
return NestedTensor(cast_tensor, cast_mask)
def decompose(self):
return self.tensors, self.mask
def __repr__(self):
return str(self.tensors)
def nested_tensor_from_tensor_list(tensor_list: List[Tensor]):
# TODO make this more general
if tensor_list[0].ndim == 3:
if torchvision._is_tracing():
# nested_tensor_from_tensor_list() does not export well to ONNX
# call _onnx_nested_tensor_from_tensor_list() instead
return _onnx_nested_tensor_from_tensor_list(tensor_list)
# TODO make it support different-sized images
max_size = _max_by_axis([list(img.shape) for img in tensor_list])
# min_size = tuple(min(s) for s in zip(*[img.shape for img in tensor_list]))
batch_shape = [len(tensor_list)] + max_size
b, c, h, w = batch_shape
dtype = tensor_list[0].dtype
device = tensor_list[0].device
tensor = torch.zeros(batch_shape, dtype=dtype, device=device)
mask = torch.ones((b, h, w), dtype=torch.bool, device=device)
for img, pad_img, m in zip(tensor_list, tensor, mask):
pad_img[: img.shape[0], : img.shape[1], : img.shape[2]].copy_(img)
m[: img.shape[1], :img.shape[2]] = False
else:
raise ValueError('not supported')
return NestedTensor(tensor, mask)
# _onnx_nested_tensor_from_tensor_list() is an implementation of
# nested_tensor_from_tensor_list() that is supported by ONNX tracing.
@torch.jit.unused
def _onnx_nested_tensor_from_tensor_list(tensor_list: List[Tensor]) -> NestedTensor:
max_size = []
for i in range(tensor_list[0].dim()):
max_size_i = torch.max(torch.stack([img.shape[i] for img in tensor_list]).to(torch.float32)).to(torch.int64)
max_size.append(max_size_i)
max_size = tuple(max_size)
# work around for
# pad_img[: img.shape[0], : img.shape[1], : img.shape[2]].copy_(img)
# m[: img.shape[1], :img.shape[2]] = False
# which is not yet supported in onnx
padded_imgs = []
padded_masks = []
for img in tensor_list:
padding = [(s1 - s2) for s1, s2 in zip(max_size, tuple(img.shape))]
padded_img = torch.nn.functional.pad(img, (0, padding[2], 0, padding[1], 0, padding[0]))
padded_imgs.append(padded_img)
m = torch.zeros_like(img[0], dtype=torch.int, device=img.device)
padded_mask = torch.nn.functional.pad(m, (0, padding[2], 0, padding[1]), "constant", 1)
padded_masks.append(padded_mask.to(torch.bool))
tensor = torch.stack(padded_imgs)
mask = torch.stack(padded_masks)
return NestedTensor(tensor, mask=mask)
def setup_for_distributed(is_master):
"""
This function disables printing when not in master process
"""
import builtins as __builtin__
builtin_print = __builtin__.print
def print(*args, **kwargs):
force = kwargs.pop('force', False)
if is_master or force:
builtin_print(*args, **kwargs)
__builtin__.print = print
def is_dist_avail_and_initialized():
if not dist.is_available():
return False
if not dist.is_initialized():
return False
return True
def get_world_size():
if not is_dist_avail_and_initialized():
gitextract_f347uvq4/
├── .github/
│ ├── CODE_OF_CONDUCT.md
│ ├── CONTRIBUTING.md
│ └── ISSUE_TEMPLATE/
│ ├── bugs.md
│ ├── questions-help-support.md
│ └── unexpected-problems-bugs.md
├── .gitignore
├── Dockerfile
├── LICENSE
├── README.md
├── d2/
│ ├── README.md
│ ├── configs/
│ │ ├── detr_256_6_6_torchvision.yaml
│ │ └── detr_segm_256_6_6_torchvision.yaml
│ ├── converter.py
│ ├── detr/
│ │ ├── __init__.py
│ │ ├── config.py
│ │ ├── dataset_mapper.py
│ │ └── detr.py
│ └── train_net.py
├── datasets/
│ ├── __init__.py
│ ├── coco.py
│ ├── coco_eval.py
│ ├── coco_panoptic.py
│ ├── panoptic_eval.py
│ └── transforms.py
├── engine.py
├── hubconf.py
├── main.py
├── models/
│ ├── __init__.py
│ ├── backbone.py
│ ├── detr.py
│ ├── matcher.py
│ ├── position_encoding.py
│ ├── segmentation.py
│ └── transformer.py
├── requirements.txt
├── run_with_submitit.py
├── test_all.py
├── tox.ini
└── util/
├── __init__.py
├── box_ops.py
├── misc.py
└── plot_utils.py
SYMBOL INDEX (268 symbols across 26 files)
FILE: d2/converter.py
function parse_args (line 12) | def parse_args():
function main (line 20) | def main():
FILE: d2/detr/config.py
function add_detr_config (line 6) | def add_detr_config(cfg):
FILE: d2/detr/dataset_mapper.py
function build_transform_gen (line 15) | def build_transform_gen(cfg, is_train):
class DetrDatasetMapper (line 42) | class DetrDatasetMapper:
method __init__ (line 55) | def __init__(self, cfg, is_train=True):
method __call__ (line 73) | def __call__(self, dataset_dict):
FILE: d2/detr/detr.py
class MaskedBackbone (line 31) | class MaskedBackbone(nn.Module):
method __init__ (line 34) | def __init__(self, cfg):
method forward (line 41) | def forward(self, images):
method mask_out_padding (line 53) | def mask_out_padding(self, feature_shapes, image_sizes, device):
class Detr (line 70) | class Detr(nn.Module):
method __init__ (line 75) | def __init__(self, cfg):
method forward (line 158) | def forward(self, batched_inputs):
method prepare_targets (line 202) | def prepare_targets(self, targets):
method inference (line 217) | def inference(self, box_cls, box_pred, mask_pred, image_sizes):
method preprocess_image (line 255) | def preprocess_image(self, batched_inputs):
FILE: d2/train_net.py
class Trainer (line 31) | class Trainer(DefaultTrainer):
method build_evaluator (line 37) | def build_evaluator(cls, cfg, dataset_name, output_folder=None):
method build_train_loader (line 49) | def build_train_loader(cls, cfg):
method build_optimizer (line 57) | def build_optimizer(cls, cfg, model):
function setup (line 106) | def setup(args):
function main (line 119) | def main(args):
FILE: datasets/__init__.py
function get_coco_api_from_dataset (line 8) | def get_coco_api_from_dataset(dataset):
function build_dataset (line 18) | def build_dataset(image_set, args):
FILE: datasets/coco.py
class CocoDetection (line 17) | class CocoDetection(torchvision.datasets.CocoDetection):
method __init__ (line 18) | def __init__(self, img_folder, ann_file, transforms, return_masks):
method __getitem__ (line 23) | def __getitem__(self, idx):
function convert_coco_poly_to_mask (line 33) | def convert_coco_poly_to_mask(segmentations, height, width):
class ConvertCocoPolysToMask (line 50) | class ConvertCocoPolysToMask(object):
method __init__ (line 51) | def __init__(self, return_masks=False):
method __call__ (line 54) | def __call__(self, image, target):
function make_coco_transforms (line 115) | def make_coco_transforms(image_set):
function build (line 147) | def build(image_set, args):
FILE: datasets/coco_eval.py
class CocoEvaluator (line 22) | class CocoEvaluator(object):
method __init__ (line 23) | def __init__(self, coco_gt, iou_types):
method update (line 36) | def update(self, predictions):
method synchronize_between_processes (line 55) | def synchronize_between_processes(self):
method accumulate (line 60) | def accumulate(self):
method summarize (line 64) | def summarize(self):
method prepare (line 69) | def prepare(self, predictions, iou_type):
method prepare_for_coco_detection (line 79) | def prepare_for_coco_detection(self, predictions):
method prepare_for_coco_segmentation (line 103) | def prepare_for_coco_segmentation(self, predictions):
method prepare_for_coco_keypoint (line 138) | def prepare_for_coco_keypoint(self, predictions):
function convert_to_xywh (line 165) | def convert_to_xywh(boxes):
function merge (line 170) | def merge(img_ids, eval_imgs):
function create_common_coco_eval (line 192) | def create_common_coco_eval(coco_eval, img_ids, eval_imgs):
function evaluate (line 208) | def evaluate(self):
FILE: datasets/coco_panoptic.py
class CocoPanoptic (line 15) | class CocoPanoptic:
method __init__ (line 16) | def __init__(self, img_folder, ann_folder, ann_file, transforms=None, ...
method __getitem__ (line 34) | def __getitem__(self, idx):
method __len__ (line 70) | def __len__(self):
method get_height_and_width (line 73) | def get_height_and_width(self, idx):
function build (line 80) | def build(image_set, args):
FILE: datasets/panoptic_eval.py
class PanopticEvaluator (line 13) | class PanopticEvaluator(object):
method __init__ (line 14) | def __init__(self, ann_file, ann_folder, output_dir="panoptic_eval"):
method update (line 23) | def update(self, predictions):
method synchronize_between_processes (line 30) | def synchronize_between_processes(self):
method summarize (line 37) | def summarize(self):
FILE: datasets/transforms.py
function crop (line 16) | def crop(image, target, region):
function hflip (line 59) | def hflip(image, target):
function resize (line 76) | def resize(image, target, size, max_size=None):
function pad (line 135) | def pad(image, target, padding):
class RandomCrop (line 148) | class RandomCrop(object):
method __init__ (line 149) | def __init__(self, size):
method __call__ (line 152) | def __call__(self, img, target):
class RandomSizeCrop (line 157) | class RandomSizeCrop(object):
method __init__ (line 158) | def __init__(self, min_size: int, max_size: int):
method __call__ (line 162) | def __call__(self, img: PIL.Image.Image, target: dict):
class CenterCrop (line 169) | class CenterCrop(object):
method __init__ (line 170) | def __init__(self, size):
method __call__ (line 173) | def __call__(self, img, target):
class RandomHorizontalFlip (line 181) | class RandomHorizontalFlip(object):
method __init__ (line 182) | def __init__(self, p=0.5):
method __call__ (line 185) | def __call__(self, img, target):
class RandomResize (line 191) | class RandomResize(object):
method __init__ (line 192) | def __init__(self, sizes, max_size=None):
method __call__ (line 197) | def __call__(self, img, target=None):
class RandomPad (line 202) | class RandomPad(object):
method __init__ (line 203) | def __init__(self, max_pad):
method __call__ (line 206) | def __call__(self, img, target):
class RandomSelect (line 212) | class RandomSelect(object):
method __init__ (line 217) | def __init__(self, transforms1, transforms2, p=0.5):
method __call__ (line 222) | def __call__(self, img, target):
class ToTensor (line 228) | class ToTensor(object):
method __call__ (line 229) | def __call__(self, img, target):
class RandomErasing (line 233) | class RandomErasing(object):
method __init__ (line 235) | def __init__(self, *args, **kwargs):
method __call__ (line 238) | def __call__(self, img, target):
class Normalize (line 242) | class Normalize(object):
method __init__ (line 243) | def __init__(self, mean, std):
method __call__ (line 247) | def __call__(self, image, target=None):
class Compose (line 261) | class Compose(object):
method __init__ (line 262) | def __init__(self, transforms):
method __call__ (line 265) | def __call__(self, image, target):
method __repr__ (line 270) | def __repr__(self):
FILE: engine.py
function train_one_epoch (line 17) | def train_one_epoch(model: torch.nn.Module, criterion: torch.nn.Module,
function evaluate (line 68) | def evaluate(model, criterion, postprocessors, data_loader, base_ds, dev...
FILE: hubconf.py
function _make_detr (line 13) | def _make_detr(backbone_name: str, dilation=False, num_classes=91, mask=...
function detr_resnet50 (line 26) | def detr_resnet50(pretrained=False, num_classes=91, return_postprocessor...
function detr_resnet50_dc5 (line 43) | def detr_resnet50_dc5(pretrained=False, num_classes=91, return_postproce...
function detr_resnet101 (line 62) | def detr_resnet101(pretrained=False, num_classes=91, return_postprocesso...
function detr_resnet101_dc5 (line 79) | def detr_resnet101_dc5(pretrained=False, num_classes=91, return_postproc...
function detr_resnet50_panoptic (line 98) | def detr_resnet50_panoptic(
function detr_resnet50_dc5_panoptic (line 121) | def detr_resnet50_dc5_panoptic(
function detr_resnet101_panoptic (line 147) | def detr_resnet101_panoptic(
FILE: main.py
function get_args_parser (line 20) | def get_args_parser():
function main (line 105) | def main(args):
FILE: models/__init__.py
function build_model (line 5) | def build_model(args):
FILE: models/backbone.py
class FrozenBatchNorm2d (line 19) | class FrozenBatchNorm2d(torch.nn.Module):
method __init__ (line 28) | def __init__(self, n):
method _load_from_state_dict (line 35) | def _load_from_state_dict(self, state_dict, prefix, local_metadata, st...
method forward (line 45) | def forward(self, x):
class BackboneBase (line 58) | class BackboneBase(nn.Module):
method __init__ (line 60) | def __init__(self, backbone: nn.Module, train_backbone: bool, num_chan...
method forward (line 72) | def forward(self, tensor_list: NestedTensor):
class Backbone (line 83) | class Backbone(BackboneBase):
method __init__ (line 85) | def __init__(self, name: str,
class Joiner (line 96) | class Joiner(nn.Sequential):
method __init__ (line 97) | def __init__(self, backbone, position_embedding):
method forward (line 100) | def forward(self, tensor_list: NestedTensor):
function build_backbone (line 112) | def build_backbone(args):
FILE: models/detr.py
class DETR (line 21) | class DETR(nn.Module):
method __init__ (line 23) | def __init__(self, backbone, transformer, num_classes, num_queries, au...
method forward (line 44) | def forward(self, samples: NestedTensor):
method _set_aux_loss (line 75) | def _set_aux_loss(self, outputs_class, outputs_coord):
class SetCriterion (line 83) | class SetCriterion(nn.Module):
method __init__ (line 89) | def __init__(self, num_classes, matcher, weight_dict, eos_coef, losses):
method loss_labels (line 108) | def loss_labels(self, outputs, targets, indices, num_boxes, log=True):
method loss_cardinality (line 130) | def loss_cardinality(self, outputs, targets, indices, num_boxes):
method loss_boxes (line 143) | def loss_boxes(self, outputs, targets, indices, num_boxes):
method loss_masks (line 164) | def loss_masks(self, outputs, targets, indices, num_boxes):
method _get_src_permutation_idx (line 193) | def _get_src_permutation_idx(self, indices):
method _get_tgt_permutation_idx (line 199) | def _get_tgt_permutation_idx(self, indices):
method get_loss (line 205) | def get_loss(self, loss, outputs, targets, indices, num_boxes, **kwargs):
method forward (line 215) | def forward(self, outputs, targets):
class PostProcess (line 258) | class PostProcess(nn.Module):
method forward (line 261) | def forward(self, outputs, target_sizes):
class MLP (line 289) | class MLP(nn.Module):
method __init__ (line 292) | def __init__(self, input_dim, hidden_dim, output_dim, num_layers):
method forward (line 298) | def forward(self, x):
function build (line 304) | def build(args):
FILE: models/matcher.py
class HungarianMatcher (line 12) | class HungarianMatcher(nn.Module):
method __init__ (line 20) | def __init__(self, cost_class: float = 1, cost_bbox: float = 1, cost_g...
method forward (line 35) | def forward(self, outputs, targets):
function build_matcher (line 85) | def build_matcher(args):
FILE: models/position_encoding.py
class PositionEmbeddingSine (line 12) | class PositionEmbeddingSine(nn.Module):
method __init__ (line 17) | def __init__(self, num_pos_feats=64, temperature=10000, normalize=Fals...
method forward (line 28) | def forward(self, tensor_list: NestedTensor):
class PositionEmbeddingLearned (line 51) | class PositionEmbeddingLearned(nn.Module):
method __init__ (line 55) | def __init__(self, num_pos_feats=256):
method reset_parameters (line 61) | def reset_parameters(self):
method forward (line 65) | def forward(self, tensor_list: NestedTensor):
function build_position_encoding (line 79) | def build_position_encoding(args):
FILE: models/segmentation.py
class DETRsegm (line 24) | class DETRsegm(nn.Module):
method __init__ (line 25) | def __init__(self, detr, freeze_detr=False):
method forward (line 37) | def forward(self, samples: NestedTensor):
function _expand (line 65) | def _expand(tensor, length: int):
class MaskHeadSmallConv (line 69) | class MaskHeadSmallConv(nn.Module):
method __init__ (line 75) | def __init__(self, dim, fpn_dims, context_dim):
method forward (line 102) | def forward(self, x: Tensor, bbox_mask: Tensor, fpns: List[Tensor]):
class MHAttentionMap (line 140) | class MHAttentionMap(nn.Module):
method __init__ (line 143) | def __init__(self, query_dim, hidden_dim, num_heads, dropout=0.0, bias...
method forward (line 158) | def forward(self, q, k, mask: Optional[Tensor] = None):
function dice_loss (line 172) | def dice_loss(inputs, targets, num_boxes):
function sigmoid_focal_loss (line 190) | def sigmoid_focal_loss(inputs, targets, num_boxes, alpha: float = 0.25, ...
class PostProcessSegm (line 218) | class PostProcessSegm(nn.Module):
method __init__ (line 219) | def __init__(self, threshold=0.5):
method forward (line 224) | def forward(self, results, outputs, orig_target_sizes, max_target_sizes):
class PostProcessPanoptic (line 241) | class PostProcessPanoptic(nn.Module):
method __init__ (line 245) | def __init__(self, is_thing_map, threshold=0.85):
method forward (line 256) | def forward(self, outputs, processed_sizes, target_sizes=None):
FILE: models/transformer.py
class Transformer (line 18) | class Transformer(nn.Module):
method __init__ (line 20) | def __init__(self, d_model=512, nhead=8, num_encoder_layers=6,
method _reset_parameters (line 42) | def _reset_parameters(self):
method forward (line 47) | def forward(self, src, mask, query_embed, pos_embed):
class TransformerEncoder (line 62) | class TransformerEncoder(nn.Module):
method __init__ (line 64) | def __init__(self, encoder_layer, num_layers, norm=None):
method forward (line 70) | def forward(self, src,
class TransformerDecoder (line 86) | class TransformerDecoder(nn.Module):
method __init__ (line 88) | def __init__(self, decoder_layer, num_layers, norm=None, return_interm...
method forward (line 95) | def forward(self, tgt, memory,
class TransformerEncoderLayer (line 127) | class TransformerEncoderLayer(nn.Module):
method __init__ (line 129) | def __init__(self, d_model, nhead, dim_feedforward=2048, dropout=0.1,
method with_pos_embed (line 146) | def with_pos_embed(self, tensor, pos: Optional[Tensor]):
method forward_post (line 149) | def forward_post(self,
method forward_pre (line 164) | def forward_pre(self, src,
method forward (line 178) | def forward(self, src,
class TransformerDecoderLayer (line 187) | class TransformerDecoderLayer(nn.Module):
method __init__ (line 189) | def __init__(self, d_model, nhead, dim_feedforward=2048, dropout=0.1,
method with_pos_embed (line 209) | def with_pos_embed(self, tensor, pos: Optional[Tensor]):
method forward_post (line 212) | def forward_post(self, tgt, memory,
method forward_pre (line 235) | def forward_pre(self, tgt, memory,
method forward (line 258) | def forward(self, tgt, memory,
function _get_clones (line 272) | def _get_clones(module, N):
function build_transformer (line 276) | def build_transformer(args):
function _get_activation_fn (line 289) | def _get_activation_fn(activation):
FILE: run_with_submitit.py
function parse_args (line 14) | def parse_args():
function get_shared_folder (line 24) | def get_shared_folder() -> Path:
function get_init_file (line 33) | def get_init_file():
class Trainer (line 42) | class Trainer(object):
method __init__ (line 43) | def __init__(self, args):
method __call__ (line 46) | def __call__(self):
method checkpoint (line 52) | def checkpoint(self):
method _setup_gpu_args (line 65) | def _setup_gpu_args(self):
function main (line 77) | def main():
FILE: test_all.py
class Tester (line 23) | class Tester(unittest.TestCase):
method test_box_cxcywh_to_xyxy (line 25) | def test_box_cxcywh_to_xyxy(self):
method indices_torch2python (line 31) | def indices_torch2python(indices):
method test_hungarian (line 34) | def test_hungarian(self):
method test_position_encoding_script (line 63) | def test_position_encoding_script(self):
method test_backbone_script (line 67) | def test_backbone_script(self):
method test_model_script_detection (line 71) | def test_model_script_detection(self):
method test_model_script_panoptic (line 80) | def test_model_script_panoptic(self):
method test_model_detection_different_inputs (line 90) | def test_model_detection_different_inputs(self):
method test_warpped_model_script_detection (line 105) | def test_warpped_model_script_detection(self):
class ONNXExporterTester (line 127) | class ONNXExporterTester(unittest.TestCase):
method setUpClass (line 129) | def setUpClass(cls):
method run_model (line 132) | def run_model(self, model, inputs_list, tolerate_small_mismatch=False,...
method ort_validate (line 151) | def ort_validate(self, onnx_io, inputs, outputs, tolerate_small_mismat...
method test_model_onnx_detection (line 178) | def test_model_onnx_detection(self):
method test_model_onnx_detection_panoptic (line 193) | def test_model_onnx_detection_panoptic(self):
FILE: util/box_ops.py
function box_cxcywh_to_xyxy (line 9) | def box_cxcywh_to_xyxy(x):
function box_xyxy_to_cxcywh (line 16) | def box_xyxy_to_cxcywh(x):
function box_iou (line 24) | def box_iou(boxes1, boxes2):
function generalized_box_iou (line 40) | def generalized_box_iou(boxes1, boxes2):
function masks_to_boxes (line 64) | def masks_to_boxes(masks):
FILE: util/misc.py
class SmoothedValue (line 27) | class SmoothedValue(object):
method __init__ (line 32) | def __init__(self, window_size=20, fmt=None):
method update (line 40) | def update(self, value, n=1):
method synchronize_between_processes (line 45) | def synchronize_between_processes(self):
method median (line 59) | def median(self):
method avg (line 64) | def avg(self):
method global_avg (line 69) | def global_avg(self):
method max (line 73) | def max(self):
method value (line 77) | def value(self):
method __str__ (line 80) | def __str__(self):
function all_gather (line 89) | def all_gather(data):
function reduce_dict (line 132) | def reduce_dict(input_dict, average=True):
class MetricLogger (line 159) | class MetricLogger(object):
method __init__ (line 160) | def __init__(self, delimiter="\t"):
method update (line 164) | def update(self, **kwargs):
method __getattr__ (line 171) | def __getattr__(self, attr):
method __str__ (line 179) | def __str__(self):
method synchronize_between_processes (line 187) | def synchronize_between_processes(self):
method add_meter (line 191) | def add_meter(self, name, meter):
method log_every (line 194) | def log_every(self, iterable, print_freq, header=None):
function get_sha (line 249) | def get_sha():
function collate_fn (line 269) | def collate_fn(batch):
function _max_by_axis (line 275) | def _max_by_axis(the_list):
class NestedTensor (line 284) | class NestedTensor(object):
method __init__ (line 285) | def __init__(self, tensors, mask: Optional[Tensor]):
method to (line 289) | def to(self, device):
method decompose (line 300) | def decompose(self):
method __repr__ (line 303) | def __repr__(self):
function nested_tensor_from_tensor_list (line 307) | def nested_tensor_from_tensor_list(tensor_list: List[Tensor]):
function _onnx_nested_tensor_from_tensor_list (line 335) | def _onnx_nested_tensor_from_tensor_list(tensor_list: List[Tensor]) -> N...
function setup_for_distributed (line 363) | def setup_for_distributed(is_master):
function is_dist_avail_and_initialized (line 378) | def is_dist_avail_and_initialized():
function get_world_size (line 386) | def get_world_size():
function get_rank (line 392) | def get_rank():
function is_main_process (line 398) | def is_main_process():
function save_on_master (line 402) | def save_on_master(*args, **kwargs):
function init_distributed_mode (line 407) | def init_distributed_mode(args):
function accuracy (line 433) | def accuracy(output, target, topk=(1,)):
function interpolate (line 451) | def interpolate(input, size=None, scale_factor=None, mode="nearest", ali...
FILE: util/plot_utils.py
function plot_logs (line 13) | def plot_logs(logs, fields=('class_error', 'loss_bbox_unscaled', 'mAP'),...
function plot_precision_recall (line 76) | def plot_precision_recall(files, naming_scheme='iter'):
Condensed preview — 42 files, each showing path, character count, and a content snippet. Download the .json file or copy for the full structured content (211K chars).
[
{
"path": ".github/CODE_OF_CONDUCT.md",
"chars": 244,
"preview": "# Code of Conduct\n\nFacebook has adopted a Code of Conduct that we expect project participants to adhere to.\nPlease read "
},
{
"path": ".github/CONTRIBUTING.md",
"chars": 1611,
"preview": "# Contributing to DETR\nWe want to make contributing to this project as easy and transparent as\npossible.\n\n## Our Develop"
},
{
"path": ".github/ISSUE_TEMPLATE/bugs.md",
"chars": 719,
"preview": "---\nname: \"🐛 Bugs\"\nabout: Report bugs in DETR\ntitle: Please read & provide the following\n\n---\n\n## Instructions To Reprod"
},
{
"path": ".github/ISSUE_TEMPLATE/questions-help-support.md",
"chars": 787,
"preview": "---\nname: \"How to do something❓\"\nabout: How to do something using DETR?\n\n---\n\n## ❓ How to do something using DETR\n\nDescr"
},
{
"path": ".github/ISSUE_TEMPLATE/unexpected-problems-bugs.md",
"chars": 1145,
"preview": "---\nname: \"Unexpected behaviors\"\nabout: Run into unexpected behaviors when using DETR\ntitle: Please read & provide the f"
},
{
"path": ".gitignore",
"chars": 189,
"preview": ".nfs*\n*.ipynb\n*.pyc\n.dumbo.json\n.DS_Store\n.*.swp\n*.pth\n**/__pycache__/**\n.ipynb_checkpoints/\ndatasets/data/\nexperiment-*"
},
{
"path": "Dockerfile",
"chars": 328,
"preview": "FROM pytorch/pytorch:1.5-cuda10.1-cudnn7-runtime\n\nENV DEBIAN_FRONTEND=noninteractive\n\nRUN apt-get update -qq && \\\n ap"
},
{
"path": "LICENSE",
"chars": 11354,
"preview": " Apache License\n Version 2.0, January 2004\n "
},
{
"path": "README.md",
"chars": 11890,
"preview": "**DE⫶TR**: End-to-End Object Detection with Transformers\n========\n\n[![Support Ukraine](https://img.shields.io/badge/Supp"
},
{
"path": "d2/README.md",
"chars": 2228,
"preview": "Detectron2 wrapper for DETR\n=======\n\nWe provide a Detectron2 wrapper for DETR, thus providing a way to better integrate "
},
{
"path": "d2/configs/detr_256_6_6_torchvision.yaml",
"chars": 1012,
"preview": "MODEL:\n META_ARCHITECTURE: \"Detr\"\n WEIGHTS: \"detectron2://ImageNetPretrained/torchvision/R-50.pkl\"\n PIXEL_MEAN: [123."
},
{
"path": "d2/configs/detr_segm_256_6_6_torchvision.yaml",
"chars": 1033,
"preview": "MODEL:\n META_ARCHITECTURE: \"Detr\"\n# WEIGHTS: \"detectron2://ImageNetPretrained/torchvision/R-50.pkl\"\n PIXEL_MEAN: [123"
},
{
"path": "d2/converter.py",
"chars": 2590,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\n\"\"\"\nHelper script to convert models trained with "
},
{
"path": "d2/detr/__init__.py",
"chars": 176,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\nfrom .config import add_detr_config\nfrom .detr im"
},
{
"path": "d2/detr/config.py",
"chars": 888,
"preview": "# -*- coding: utf-8 -*-\n# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\nfrom detectron2.config im"
},
{
"path": "d2/detr/dataset_mapper.py",
"chars": 4570,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\nimport copy\nimport logging\n\nimport numpy as np\nim"
},
{
"path": "d2/detr/detr.py",
"chars": 11143,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\nimport logging\nimport math\nfrom typing import Lis"
},
{
"path": "d2/train_net.py",
"chars": 4999,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\n\"\"\"\nDETR Training Script.\n\nThis script is a simpl"
},
{
"path": "datasets/__init__.py",
"chars": 897,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\nimport torch.utils.data\nimport torchvision\n\nfrom "
},
{
"path": "datasets/coco.py",
"chars": 5253,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\n\"\"\"\nCOCO dataset which returns image_id for evalu"
},
{
"path": "datasets/coco_eval.py",
"chars": 8735,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\n\"\"\"\nCOCO evaluator that works in distributed mode"
},
{
"path": "datasets/coco_panoptic.py",
"chars": 3723,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\nimport json\nfrom pathlib import Path\n\nimport nump"
},
{
"path": "datasets/panoptic_eval.py",
"chars": 1493,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\nimport json\nimport os\n\nimport util.misc as utils\n"
},
{
"path": "datasets/transforms.py",
"chars": 8524,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\n\"\"\"\nTransforms and data augmentation for both ima"
},
{
"path": "engine.py",
"chars": 6626,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\n\"\"\"\nTrain and eval functions used in main.py\n\"\"\"\n"
},
{
"path": "hubconf.py",
"chars": 6265,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\nimport torch\n\nfrom models.backbone import Backbon"
},
{
"path": "main.py",
"chars": 11532,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\nimport argparse\nimport datetime\nimport json\nimpor"
},
{
"path": "models/__init__.py",
"chars": 143,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\nfrom .detr import build\n\n\ndef build_model(args):\n"
},
{
"path": "models/backbone.py",
"chars": 4437,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\n\"\"\"\nBackbone modules.\n\"\"\"\nfrom collections import"
},
{
"path": "models/detr.py",
"chars": 17088,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\n\"\"\"\nDETR model and criterion classes.\n\"\"\"\nimport "
},
{
"path": "models/matcher.py",
"chars": 4250,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\n\"\"\"\nModules to compute the matching cost and solv"
},
{
"path": "models/position_encoding.py",
"chars": 3336,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\n\"\"\"\nVarious positional encodings for the transfor"
},
{
"path": "models/segmentation.py",
"chars": 15573,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\n\"\"\"\nThis file provides the definition of the conv"
},
{
"path": "models/transformer.py",
"chars": 12162,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\n\"\"\"\nDETR Transformer class.\n\nCopy-paste from torc"
},
{
"path": "requirements.txt",
"chars": 224,
"preview": "cython\ngit+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI&egg=pycocotools\nsubmitit\ntorch>=1.5.0\ntorch"
},
{
"path": "run_with_submitit.py",
"chars": 3476,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\n\"\"\"\nA script to run multinode training with submi"
},
{
"path": "test_all.py",
"chars": 8806,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\nimport io\nimport unittest\n\nimport torch\nfrom torc"
},
{
"path": "tox.ini",
"chars": 65,
"preview": "[flake8]\nmax-line-length = 120\nignore = F401,E402,F403,W503,W504\n"
},
{
"path": "util/__init__.py",
"chars": 71,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\n"
},
{
"path": "util/box_ops.py",
"chars": 2561,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\n\"\"\"\nUtilities for bounding box manipulation and G"
},
{
"path": "util/misc.py",
"chars": 15356,
"preview": "# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\n\"\"\"\nMisc functions, including distributed helpers"
},
{
"path": "util/plot_utils.py",
"chars": 4514,
"preview": "\"\"\"\nPlotting utilities to visualize training logs.\n\"\"\"\nimport torch\nimport pandas as pd\nimport numpy as np\nimport seabor"
}
]
About this extraction
This page contains the full source code of the facebookresearch/detr GitHub repository, extracted and formatted as plain text for AI agents and large language models (LLMs). The extraction includes 42 files (197.3 KB), approximately 49.5k tokens, and a symbol index with 268 extracted functions, classes, methods, constants, and types. Use this with OpenClaw, Claude, ChatGPT, Cursor, Windsurf, or any other AI tool that accepts text input. You can copy the full output to your clipboard or download it as a .txt file.
Extracted by GitExtract — free GitHub repo to text converter for AI. Built by Nikandr Surkov.