Repository: DeNA/Chainer_Mask_R-CNN
Branch: master
Commit: 315a5b098978
Files: 25
Total size: 113.7 KB
Directory structure:
gitextract_iecqmj1q/
├── .gitignore
├── LICENSE
├── README.md
├── README_JP.md
├── coco_dataset.py
├── demo.py
├── getcoco.sh
├── mask_rcnn.py
├── mask_rcnn_resnet.py
├── mask_rcnn_train_chain.py
├── mask_rcnn_train_chain_batch.py
├── train.py
└── utils/
├── __init__.py
├── bn_utils.py
├── box_utils.py
├── cocoapi_evaluator.py
├── detection_coco_evaluator.py
├── detectron_parser.py
├── eval_detection_coco.py
├── makecocolist.py
├── proposal_target_creator.py
├── region_proposal_network.py
├── roi_align_2d.py
├── updater.py
└── vis_bbox.py
================================================
FILE CONTENTS
================================================
================================================
FILE: .gitignore
================================================
__pycache*
result
*.png
================================================
FILE: LICENSE
================================================
Copyright (c) 2018 DeNA Co., Ltd.
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, and/or sublicense
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software; and
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY
CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
#####################################################################################
# Chainer_Mask_R-CNN is designed based on chainercv's API.
# Chainer_Mask_R-CNN's source code and documents contain the original chainercv ones.
#####################################################################################
Copyright (c) 2017 Yusuke Niitani.
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:
* Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
* Redistributions in binary form must reproduce the above
copyright notice, this list of conditions and the following
disclaimer in the documentation and/or other materials provided
with the distribution.
* Neither the name of the chainercv Developers nor the names of any
contributors may be used to endorse or promote products derived
from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#####################################################################################
# Chainer_Mask_R-CNN is designed based on chainer's API.
# Chainer_Mask_R-CNN's source code and documents contain the original chainer ones.
#####################################################################################
Copyright (c) 2015 Preferred Infrastructure, Inc.
Copyright (c) 2015 Preferred Networks, Inc.
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:
* Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
* Redistributions in binary form must reproduce the above
copyright notice, this list of conditions and the following
disclaimer in the documentation and/or other materials provided
with the distribution.
* Neither the name of the chainer Developers nor the names of any
contributors may be used to endorse or promote products derived
from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
~
================================================
FILE: README.md
================================================
# Chainer\_Mask\_R-CNN
Chainer implementation of Mask R-CNN - the multi-task network for object detection, object classification, and instance segmentation.
(https://arxiv.org/abs/1703.06870)
<a href="README_JP.md">日本語版 README</a>
## What's New
- Training result for R-50-C4 model has been evaluated!
- COCO box AP = 0.346 using our trainer (0.355 with official boxes)
- COCO mask AP = 0.287 using our trainer (0.314 with official boxes)
## Examples
- to be updated
## Requirements
- [Chainer](https://github.com/pfnet/chainer)
- [Chainercv](https://github.com/chainer/chainercv)
- [Cupy](https://github.com/cupy/cupy)
(operable if your environment can run chainer > v3 with cuda and cudnn.)
(verified as operable: chainer==3.1.0, chainercv==0.7.0, cupy==1.0.3)
```
$ pip install chainer
$ pip install chainercv
$ pip install cupy
```
- Python 3.0+
- NumPy
- Matplotlib
- OpenCV
## TODOs
- [x] Precision Evaluator (bbox, COCO metric)
- [x] Detectron Model Parser
- [x] Modify ROIAlign
- [x] Mask inference using refined ROIs
- [x] Precision Evaluator (mask, COCO metric)
- [ ] Improve segmentation AP for R-50-C4 model
- [ ] Feature Pyramid Network (R-50-FPN)
- [ ] Keypoint Detection (R-50-FPN, Keypoints)
## Benchmark Results
<table><tbody>
<tr><th align="left" bgcolor=#f8f8f8> </th> <td bgcolor=white> Box AP 50:95</td><td bgcolor=white> Segm AP 50:95</td></tr>
<tr><th align="left" bgcolor=#f8f8f8>Ours (1 GPU)</th> <td bgcolor=white> 0.346 </td><td bgcolor=white> 0.287 </td></tr>
<tr><th align="left" bgcolor=#f8f8f8>Detectron model</th> <td bgcolor=white> 0.350 </td><td bgcolor=white> 0.295 </td></tr>
<tr><th align="left" bgcolor=#f8f8f8>Detectron caffe2</th> <td bgcolor=white> 0.355 </td><td bgcolor=white> 0.314 </td></tr>
</table></tbody>
## Inference with Pretrained Models
- Download the pretrained model from the [Model Zoo] (https://github.com/facebookresearch/Detectron/blob/master/MODEL_ZOO.md)
(`model` link of `R-50-C4 Mask` at `End-to-End Faster & Mask R-CNN Baselines`)
- Make `modelfiles` directory and put the downloaded file `model_final.pkl` in it
- Execute:
```
python utils/detectron_parser.py
```
- And the converted model file is saved in `modelfiles`
- Run the demo:
```
python demo.py --bn2affine --modelfile modelfiles/e2e_mask_rcnn_R-50-C4_1x_d2c.npz --image <input image>
```
## Prerequisites for training
- Download 'ResNet-50-model.caffemodel' from the "OneDrive download" of [ResNet pretrained models](https://github.com/KaimingHe/deep-residual-networks#models)
for model initialization and place it in ~/.chainer/dataset/pfnet/chainer/models/
- COCO 2017 dataset :
the COCO dataset can be downloaded and unzipped by:
```
bash getcoco.sh
```
Setup the COCO API:
```
git clone https://github.com/waleedka/coco
cd coco/PythonAPI/
make
python setup.py install
cd ../../
```
note: the official coco repository is not python3 compatible.
Use the repository above in order to run our evaluation.
## Train
```
python train.py
```
arguments and the default conditions are defined as follows:
```
'--dataset', choices=('coco2017'), default='coco2017'
'--extractor', choices=('resnet50','resnet101'), default='resnet50', help='extractor network'
'--gpu', '-g', type=int, default=0
'--lr', '-l', type=float, default=1e-4
'--batchsize', '-b', type=int, default=8
'--freeze_bn', action='store_true', default=False, help='freeze batchnorm gamma/beta'
'--bn2affine', action='store_true', default=False, help='batchnorm to affine'
'--out', '-o', default='result', help='output directory'
'--seed', '-s', type=int, default=0
'--roialign', action='store_true', default=True, help='True: ROIAlign, False: ROIpooling'
'--step_size', '-ss', type=int, default=400000
'--lr_step', '-ls', type=int, default=480000
'--lr_initialchange', '-li', type=int, default=800
'--pretrained', '-p', type=str, default='imagenet'
'--snapshot', type=int, default=4000
'--validation', type=int, default=30000
'--resume', type=str
'--iteration', '-i', type=int, default=800000
'--roi_size', '-r', type=int, default=14, help='ROI size for mask head input'
'--gamma', type=float, default=1, help='mask loss balancing factor'
```
note that we use a subdivision-based updater to enable training with large batch size.
## Demo
Segment the objects in the input image by executing:
```
python demo.py --image <input image> --modelfile result/snapshot_model.npz --contour
```
## Evaluation
Evaluate the trained model with COCO metric (bounding box, segmentation) :
```
python train.py --lr 0 --iteration 1 --validation 1 --resume <trained_model>
```
## Citation
Please cite the original paper in your publications if it helps your research:
@article{DBLP:journals/corr/HeGDG17,
author = {Kaiming He and
Georgia Gkioxari and
Piotr Doll{\'{a}}r and
Ross B. Girshick},
title = {Mask {R-CNN}},
journal = {CoRR},
volume = {abs/1703.06870},
year = {2017},
url = {http://arxiv.org/abs/1703.06870},
archivePrefix = {arXiv},
eprint = {1703.06870},
timestamp = {Wed, 07 Jun 2017 14:42:32 +0200},
biburl = {http://dblp.org/rec/bib/journals/corr/HeGDG17},
bibsource = {dblp computer science bibliography, http://dblp.org}
}
================================================
FILE: README_JP.md
================================================
# Chainer\_Mask\_R-CNN
マルチタスク検出器Mask R-CNNのchainer実装
(https://arxiv.org/abs/1703.06870)
## 実行例
- 準備中
## 必要環境
- [Chainer](https://github.com/pfnet/chainer)
- [Chainercv](https://github.com/chainer/chainercv)
- [Cupy](https://github.com/cupy/cupy)
(動作確認済み: chainer==3.1.0, chainercv==0.7.0, verified: cupy==1.0.3)
```
$ pip install chainer
$ pip install chainercv
$ pip install cupy==1.0.3
```
- Python 3.0+
- NumPy
- Matplotlib
- OpenCV
## TODOs
- [x] Precision Evaluator (bbox, COCO metric)
- [x] Detectron Model Parser
- [x] Modify ROIAlign
- [x] Mask inference using refined ROIs
- [x] Precision Evaluator (mask, COCO metric)
- [ ] Feature Pyramid Network (R-50-FPN)
- [ ] Keypoint Detection (R-50-FPN, Keypoints)
## 学習済みモデルの使用
- [Model Zoo] (https://github.com/facebookresearch/Detectron/blob/master/MODEL_ZOO.md) からモデルファイルをダウンロード
( `End-to-End Faster & Mask R-CNN Baselines` の `R-50-C4 Mask` 行の `model` リンク)
- `modelfiles` ディレクトリを作り、ダウンロードした `model_final.pkl` を置く
- 以下を実行
```
python utils/detectron_parser.py
```
- `modelfiles` の中に変換されたモデルファイルが保存されます。
- 以下によりデモを実行
```
python demo.py --bn2affine --modelfile modelfiles/e2e_mask_rcnn_R-50-C4_1x_d2c.npz --image <input image>
```
## 学習のための準備
- 学習済みモデルのダウンロード
・以下リンク先の'OneDrive download'から、ResNet-50-model.caffemodelをダウンロード
[ResNet pretrained models](https://github.com/KaimingHe/deep-residual-networks#models)
・~/.chainer/dataset/pfnet/chainer/models/ に置く
- COCO 2017 データセット
COCOデータセットのダウンロードと解凍:
```
bash getcoco.sh
```
- COCO APIのセットアップ:
```
git clone https://github.com/waleedka/coco
cd coco/PythonAPI/
make
python setup.py install
cd ../../
```
## 学習
```
python train.py
```
引数は以下です:
```
'--dataset', choices=('coco2017'), default='coco2017'
'--extractor', choices=('resnet50','resnet101'), default='resnet50', help='extractor network'
'--gpu', '-g', type=int, default=0
'--lr', '-l', type=float, default=1e-4
'--batchsize', '-b', type=int, default=8
'--freeze_bn', action='store_true', default=False, help='freeze batchnorm gamma/beta'
'--bn2affine', action='store_true', default=False, help='batchnorm to affine'
'--out', '-o', default='result', help='output directory'
'--seed', '-s', type=int, default=0
'--roialign', action='store_true', default=True, help='True: ROIAlign, False: ROIpooling'
'--step_size', '-ss', type=int, default=400000
'--lr_step', '-ls', type=int, default=480000
'--lr_initialchange', '-li', type=int, default=800
'--pretrained', '-p', type=str, default='imagenet'
'--snapshot', type=int, default=4000
'--validation', type=int, default=30000
'--resume', type=str
'--iteration', '-i', type=int, default=800000
'--roi_size', '-r', type=int, default=14, help='ROI size for mask head input'
'--gamma', type=float, default=1, help='mask loss balancing factor'
```
本実装ではsubdivisionを用いたupdateを行なっているため、batch size = 1 相当のGPUメモリでbatch size=8等を指定可能です
## デモ
入力画像のインスタンス・セグメンテーションを実行します:
```
python demo.py --image <input image> --modelfile result/snapshot_model.npz --contour
```
### 評価
COCO metric (Bounding Box, Segmentation) によるモデルの評価を実行します。
```
python train.py --lr 0 --iteration 1 --validation 1 --resume <trained_model>
```
## 引用
Please cite the original paper in your publications if it helps your research:
@article{DBLP:journals/corr/HeGDG17,
author = {Kaiming He and
Georgia Gkioxari and
Piotr Doll{\'{a}}r and
Ross B. Girshick},
title = {Mask {R-CNN}},
journal = {CoRR},
volume = {abs/1703.06870},
year = {2017},
url = {http://arxiv.org/abs/1703.06870},
archivePrefix = {arXiv},
eprint = {1703.06870},
timestamp = {Wed, 07 Jun 2017 14:42:32 +0200},
biburl = {http://dblp.org/rec/bib/journals/corr/HeGDG17},
bibsource = {dblp computer science bibliography, http://dblp.org}
}
================================================
FILE: coco_dataset.py
================================================
import numpy as np
from skimage.draw import polygon
import json
import os
import cv2
import pycocotools
from pycocotools.coco import COCO
import chainer
from chainercv.utils import read_image
class COCODataset(chainer.dataset.DatasetMixin):
def __init__(self, data_dir='COCO/', json_file='instances_train2017.json',
name='train2017', id_list_file='train2017.txt', sizemin=10):
self.data_dir = data_dir
self.json_file = json_file
self.coco = COCO(self.data_dir + 'annotations/'+self.json_file)
self.ids = self.coco.getImgIds()
self.name = name
self.sizemin = sizemin
self.class_ids = sorted(self.coco.getCatIds())
def __len__(self):
return len(self.ids)
def ann2rle(self, ann, height, width):
if isinstance(ann, list):
rles = pycocotools.mask.frPyObjects(ann, height, width)
rle = pycocotools.mask.merge(rles)
elif isinstance(ann['counts'], list):
rle = pycocotools.mask.frPyObjects(ann, height, width)
else:
rle = ann
return rle
def get_example(self, i):
#i = i % 500 # for limiting data size
numofboxes=0
while True:
id_ = self.ids[i]
annot_labels, annot_bboxes, annot_segs= list(), list(), list()
anno_ids = self.coco.getAnnIds(imgIds=[int(id_)], iscrowd=None)
annotations = self.coco.loadAnns(anno_ids)
for a in annotations:
if a['bbox'][2] > self.sizemin and a['bbox'][3] > self.sizemin \
and a['iscrowd']==0:
annot_labels.append(a['category_id'])
annot_bboxes.append(a['bbox'])
annot_segs.append(a['segmentation'])
numofboxes=len(annot_labels)
if numofboxes > 0 or chainer.config.train == False:
break
else:
i = i - 1
img_file = os.path.join(self.data_dir, self.name, '{:012}'.format(id_) + '.jpg')
img = read_image(img_file, color=True)
_, h, w = img.shape
annot_masks = []
for annot_seg_polygons in annot_segs:
rle = self.ann2rle(annot_seg_polygons, h, w)
annot_masks.append(pycocotools.mask.decode(rle))
if numofboxes > 0:
annot_masks = np.stack(annot_masks).astype(np.uint8) #y,x
annot_bboxes = np.stack(annot_bboxes).astype(np.float32)
annot_labels = np.stack(annot_labels).astype(np.int32)
else:
annot_labels, annot_bboxes, annot_masks = [], [], []
return img, annot_labels, annot_bboxes, annot_masks, i
================================================
FILE: demo.py
================================================
import argparse
import chainer
import numpy as np
from mask_rcnn_train_chain import MaskRCNNTrainChain
from utils.bn_utils import freeze_bn, bn_to_affine
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--gpu', type=int, default=0)
parser.add_argument('--modelfile')
parser.add_argument('--image', type=str)
parser.add_argument('--roi_size', '-r', type=int, default=14, help='ROI size for mask head input')
parser.add_argument('--roialign', action='store_false', default=True, help='default: True')
parser.add_argument('--contour', action='store_true', default=False, help='visualize contour')
parser.add_argument('--background', action='store_true', default=False, help='background(no-display mode)')
parser.add_argument('--bn2affine', action='store_true', default=False, help='batchnorm to affine')
parser.add_argument('--extractor', choices=('resnet50','resnet101'),
default='resnet50', help='extractor network')
args = parser.parse_args()
#network class id --> coco label id
test_class_ids = [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]
if args.background:
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plot
from utils.vis_bbox import vis_bbox
from chainercv.datasets import voc_bbox_label_names
from mask_rcnn_resnet import MaskRCNNResNet
from chainercv import utils
if args.extractor=='resnet50':
model = MaskRCNNResNet(n_fg_class=80, roi_size=args.roi_size, pretrained_model=args.modelfile, n_layers=50, roi_align=args.roialign, class_ids=test_class_ids)
elif args.extractor=='resnet101':
model = MaskRCNNResNet(n_fg_class=80, roi_size=args.roi_size, pretrained_model=args.modelfile, n_layers=101, roi_align=args.roialign, class_ids=test_class_ids)
chainer.serializers.load_npz(args.modelfile, model)
if args.gpu >= 0:
chainer.cuda.get_device_from_id(args.gpu).use()
model.to_gpu()
if args.bn2affine:
bn_to_affine(model)
img = utils.read_image(args.image, color=True)
bboxes, labels, scores, masks = model.predict([img])
bbox, label, score, mask = bboxes[0], np.asarray(labels[0],dtype=np.int32), scores[0], masks[0]
#print(bbox, np.asarray(label,dtype=np.int32), score, mask)
coco_label_names=('background', # class zero
'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light',
'fire hydrant', 'street sign', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow',
'elephant', 'bear', 'zebra', 'giraffe', 'hat', 'backpack', 'umbrella', 'shoe', 'eye glasses', 'handbag', 'tie', 'suitcase', 'frisbee',
'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard', 'tennis racket', 'bottle',
'plate', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', 'sandwich', 'orange',
'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch', 'potted plant', 'bed',
'mirror', 'dining table', 'window', 'desk','toilet', 'door', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', 'microwave', 'oven',
'toaster', 'sink', 'refrigerator', 'blender', 'book', 'clock', 'vase', 'scissors', 'teddy bear', 'hair drier', 'toothbrush'
)
vis_bbox(
img, bbox, label=label, score=score, mask=mask, label_names=coco_label_names, contour=args.contour, labeldisplay=True)
plot.show()
filename = "output.png"
plot.savefig(filename)
if __name__ == '__main__':
main()
================================================
FILE: getcoco.sh
================================================
# get COCO dataset
mkdir COCO
cd COCO
wget http://images.cocodataset.org/zips/train2017.zip
wget http://images.cocodataset.org/zips/val2017.zip
wget http://images.cocodataset.org/annotations/annotations_trainval2017.zip
unzip train2017.zip
unzip val2017.zip
unzip annotations_trainval2017.zip
rm -f train2017.zip
rm -f val2017.zip
rm -f annotations_trainval2017.zip
================================================
FILE: mask_rcnn.py
================================================
from __future__ import division
import numpy as np
import chainer
from chainer import cuda
import chainer.functions as F
from chainercv.links.model.faster_rcnn.utils.loc2bbox import loc2bbox
from chainercv.utils import non_maximum_suppression
from chainercv.transforms.image.resize import resize
import cv2
import pycocotools
from utils.box_utils import bbox_yxyx2xywh, im_mask
class MaskRCNN(chainer.Chain):
def __init__(self, extractor, rpn, head, mean,
min_size=600, max_size=1000,
loc_normalize_mean=(0., 0., 0., 0.),
loc_normalize_std=(0.1, 0.1, 0.2, 0.2),
class_ids=[]
):
print("MaskRCNN initialization")
super(MaskRCNN, self).__init__()
with self.init_scope():
self.extractor = extractor
self.rpn = rpn
self.head = head
self.mean = mean
self.min_size = min_size
self.max_size = max_size
self.loc_normalize_mean = loc_normalize_mean
self.loc_normalize_std = loc_normalize_std
self.use_preset('visualize')
if class_ids==[]:
raise ValueError('set class ids')
self.class_ids = class_ids
self.preset = 'visualize'
@property
def n_class(self):
return self.head.n_class
def __call__(self, x, scale=1.):
img_size = x.shape[2:]
h = self.extractor(x) #VGG
rpn_locs, rpn_scores, rois, roi_indices, anchor = \
self.rpn(h, img_size, scale) #Region Proposal Network
hres5 = self.head.res5head(h, rois, roi_indices)
roi_cls_locs, roi_scores = self.head.boxhead(hres5)
return roi_cls_locs, roi_scores, rois, roi_indices, h
def use_preset(self, preset):
if preset == 'visualize':
self.nms_thresh = 0.3
self.score_thresh = 0.7
self.preset = 'visualize'
elif preset == 'evaluate':
self.nms_thresh = 0.5
self.score_thresh = 0.05
self.preset = 'evaluate'
else:
raise ValueError('preset must be visualize or evaluate')
def prepare(self, img):
_, H, W = img.shape
scale = self.min_size / min(H, W)
if scale * max(H, W) > self.max_size:
scale = self.max_size / max(H, W)
#img = resize(img, (int(H * scale), int(W * scale)))
img = img.transpose((1,2,0))
img = cv2.resize(img, None, None, fx=scale, fy=scale,
interpolation=cv2.INTER_LINEAR)
img = img.transpose((2,0,1))
img = (img - self.mean).astype(np.float32, copy=False)
img = img[::-1, :, :] # RGB to BGR order for resnet pretrained model
return img
def _suppress(self, raw_cls_bbox, raw_cls_roi, raw_prob):
bbox = list()
roi = list()
label = list()
score = list()
mask = list()
for l in range(1, self.n_class):
cls_bbox_l = raw_cls_bbox.reshape((-1, self.n_class, 4))[:, l, :]
cls_roi_l = raw_cls_roi.reshape((-1, self.n_class, 4))[:, l, :]
prob_l = raw_prob[:, l]
lmask = prob_l > self.score_thresh
cls_bbox_l = cls_bbox_l[lmask]
cls_roi_l = cls_roi_l[lmask]
prob_l = prob_l[lmask]
keep = non_maximum_suppression(cls_bbox_l, self.nms_thresh, prob_l)
bbox.append(cls_bbox_l[keep])
roi.append(cls_roi_l[keep])
label.append((l - 1) * np.ones((len(keep),)))
score.append(prob_l[keep])
bbox = np.concatenate(bbox, axis=0).astype(np.float32)
roi = np.concatenate(roi, axis=0).astype(np.float32)
label = np.concatenate(label, axis=0).astype(np.float32)
score = np.concatenate(score, axis=0).astype(np.float32)
return bbox, roi, label, score
def predict(self, imgs):
prepared_imgs = list()
sizes = list()
#print("predicting!")
for img in imgs:
size = img.shape[1:]
img = self.prepare(img.astype(np.float32))
prepared_imgs.append(img)
sizes.append(size)
bboxes = list()
out_rois = list()
labels = list()
scores = list()
masks = list()
for img, size in zip(prepared_imgs, sizes):
with chainer.using_config('train', False), \
chainer.function.no_backprop_mode():
img_var = chainer.Variable(self.xp.asarray(img[None]))
scale = img_var.shape[3] / size[1]
roi_cls_locs, roi_scores, rois, _, h = self.__call__(img_var, scale=scale)
#assuming batch size = 1
roi_cls_loc = roi_cls_locs.data
roi_score = roi_scores.data
roi = rois / scale
mean = self.xp.tile(self.xp.asarray(self.loc_normalize_mean), self.n_class)
std = self.xp.tile(self.xp.asarray(self.loc_normalize_std), self.n_class)
roi_cls_loc = (roi_cls_loc * std + mean).astype(np.float32)
roi_cls_loc = roi_cls_loc.reshape((-1, self.n_class, 4))
roi = self.xp.broadcast_to(roi[:, None], roi_cls_loc.shape).reshape((-1, 4))
cls_bbox = loc2bbox(roi, roi_cls_loc.reshape((-1, 4)))
cls_bbox = cls_bbox.reshape((-1, self.n_class * 4))
cls_roi = roi.reshape((-1, self.n_class * 4))
#clip the bbox
cls_bbox[:, 0::2] = self.xp.clip(cls_bbox[:, 0::2], 0, size[0])
cls_bbox[:, 1::2] = self.xp.clip(cls_bbox[:, 1::2], 0, size[1])
cls_roi[:, 0::2] = self.xp.clip(cls_roi[:, 0::2], 0, size[0])
cls_roi[:, 1::2] = self.xp.clip(cls_roi[:, 1::2], 0, size[1])
prob = F.softmax(roi_score).data
raw_cls_bbox = cuda.to_cpu(cls_bbox)
raw_cls_roi = cuda.to_cpu(cls_roi)
raw_prob = cuda.to_cpu(prob)
bbox, out_roi, label, score = self._suppress(raw_cls_bbox, raw_cls_roi, raw_prob)
mask=[]
if len(bbox) > 0:
# mask head
roi_indices = self.xp.zeros((len(bbox),), dtype=np.int32)
with chainer.using_config('train', False), \
chainer.function.no_backprop_mode():
hres5 = self.head.res5head(h, cuda.to_gpu(bbox * scale), roi_indices)
roi_masks = self.head.maskhead(hres5)
roi_mask = F.sigmoid(roi_masks).data
raw_mask = cuda.to_cpu(roi_mask)
# postprocess
if self.preset == 'evaluate':
bboxes.append(bbox_yxyx2xywh(bbox))
wmasks = []
for m, b, l in zip(raw_mask, bbox, label):
wm = im_mask(m[int(l+1)], size, b)
# encode the mask
wm = pycocotools.mask.encode(np.asfortranarray(wm))
wm['counts'] = wm['counts'].decode('ascii')
mask.append(wm)
elif self.preset == 'visualize':
bboxes.append(bbox)
wmasks = []
for m, b, l in zip(raw_mask, bbox, label):
wm = im_mask(m[int(l+1)], size, b)
mask.append(wm)
elif self.preset == 'evaluate':
# len(bbox) = 0
wm = np.zeros((size[0], size[1]), dtype=np.uint8)
wm = pycocotools.mask.encode(np.asfortranarray(wm))
wm['counts'] = wm['counts'].decode('ascii')
mask.append(wm)
bboxes.append(bbox_yxyx2xywh(bbox))
labels.append([self.class_ids[int(l)] for l in label.tolist()])
scores.append(score)
masks.append(mask)
return bboxes, labels, scores, masks
================================================
FILE: mask_rcnn_resnet.py
================================================
import numpy as np
import chainer
import chainer.functions as F
import chainer.links as L
from mask_rcnn import MaskRCNN
#from chainercv.links.model.faster_rcnn.region_proposal_network import \
# RegionProposalNetwork
from utils.region_proposal_network import RegionProposalNetwork
from utils import roi_align_2d
from chainer.links.model.vision.resnet import BuildingBlock, _retrieve
from chainer.links.connection.convolution_2d import Convolution2D
from chainer.links.connection.linear import Linear
from chainer.links.normalization.batch_normalization import BatchNormalization
from chainer.initializers import constant
class ExtractorResNet(chainer.link.Chain):
def __init__(self, pretrained_model='auto', n_layers=50, roi_size=14):
super(ExtractorResNet, self).__init__()
print('Extractor ResNet',n_layers,' initialization')
kwargs = {'initialW': constant.Zero()}
if pretrained_model=='auto':
if n_layers == 50:
pretrained_model = 'ResNet-50-model.caffemodel'
block = [3, 4, 6, 3]
elif n_layers == 101:
pretrained_model = 'ResNet-101-model.caffemodel'
block = [3, 4, 23, 3]
with self.init_scope():
self.conv1 = Convolution2D(3, 64, 7, 2, 3, **kwargs, nobias=True)
self.bn1 = BatchNormalization(64)
self.res2 = BuildingBlock(block[0], 64, 64, 256, 1, **kwargs)
self.res3 = BuildingBlock(block[1], 256, 128, 512, 2, **kwargs)
self.res4 = BuildingBlock(block[2], 512, 256, 1024, 2, **kwargs)
self.res5 = BuildingBlock(block[3], 1024, 512, 2048, roi_size//7, **kwargs)
self.fc6 = Linear(2048, 1000)
if pretrained_model and pretrained_model.endswith('.caffemodel'):
_retrieve(n_layers, 'ResNet-{}-model.npz'.format(n_layers),
pretrained_model, self)
elif pretrained_model:
npz.load_npz(pretrained_model, self)
del self.fc6
def __call__(self, x):
h = F.relu(self.bn1(self.conv1(x)))
_, _, H, W = h.shape
Hpool = (H + 1)//2
Wpool = (W + 1)//2
h = F.max_pooling_2d(h, ksize=3, stride=2, pad=1)
h = h[:, :, :Hpool, :Wpool]
h = self.res2(h)
h = self.res3(h)
h = self.res4(h)
return h
class MaskRCNNResNet(MaskRCNN):
feat_stride = 16
def __init__(self,
n_fg_class=None,
pretrained_model=None,
min_size=800, max_size=1333,
ratios=[0.5 ,1, 2], anchor_scales=[2, 4, 8, 16, 32],
initialW=None, rpn_initialW=None,
loc_initialW=None, score_initialW=None,
proposal_creator_params={"n_test_pre_nms":6000,"n_test_post_nms": 1000,"min_size":4},
roi_size=14,
class_ids=[],
n_layers=50,
roi_align=True
):
print("MaskRNNResNet initialization")
if n_fg_class is None:
raise ValueError('supply n_fg_class!')
if loc_initialW is None:
loc_initialW = chainer.initializers.Normal(0.001)
if score_initialW is None:
score_initialW = chainer.initializers.Normal(0.01)
if rpn_initialW is None:
rpn_initialW = chainer.initializers.Normal(0.01)
if initialW is None:# and pretrained_model:
print("setting initialW")
initialW = chainer.initializers.Normal(0.01)
self.roi_size=roi_size
if pretrained_model is not None:
pretrained_model = 'auto'
extractor = ExtractorResNet(pretrained_model, n_layers=n_layers, roi_size=roi_size)
rpn = RegionProposalNetwork(
1024, 1024,
ratios=ratios, anchor_scales=anchor_scales,
feat_stride=self.feat_stride,
initialW=rpn_initialW,
proposal_creator_params=proposal_creator_params,
)
head = MaskRCNNHead(
n_fg_class + 1,
roi_size=self.roi_size, spatial_scale=1. / self.feat_stride,
initialW=initialW, loc_initialW=loc_initialW, score_initialW=score_initialW,
roi_align=roi_align, reslayer=extractor.res5
)
del extractor.res5
super(MaskRCNNResNet, self).__init__(
extractor, rpn, head,
mean=np.array([122.7717, 115.9465, 102.9801], dtype=np.float32)[:, None, None],
min_size=min_size, max_size=max_size, class_ids=class_ids
)
class MaskRCNNHead(chainer.Chain):
def __init__(self, n_class, roi_size, spatial_scale,
initialW=None, loc_initialW=None, score_initialW=None, roi_align=True, reslayer=None):
super(MaskRCNNHead, self).__init__()
with self.init_scope():
self.res5 = reslayer#BuildingBlock(3, 1024, 512, 2048, 1, initialW=initialW)
#class / loc branch
self.cls_loc = L.Linear(2048, n_class * 4, initialW=initialW)
self.score = L.Linear(2048, n_class, initialW=score_initialW)
#Mask-RCNN branch
self.deconvm1 = L.Deconvolution2D(2048, 256, 2, 2, initialW=initialW)
self.convm2 = L.Convolution2D(256, n_class, 1, 1, pad=0,initialW=initialW)
self.n_class = n_class
self.roi_size = roi_size
self.spatial_scale = spatial_scale
self.roi_align = roi_align
print("ROI Align=",roi_align)
def res5head(self, x, rois, roi_indices):
# extracted feature map -> pooling -> res5 block
roi_indices = roi_indices.astype(np.float32)
indices_and_rois = self.xp.concatenate(
(roi_indices[:, None], rois), axis=1)
#x: (batch, channel, w, h)
#rois: (128, 4) (ROI indices)
if self.roi_align:
pool = _roi_align_2d_yx(
x, indices_and_rois, self.roi_size,self.roi_size,
self.spatial_scale)
else:
pool = _roi_pooling_2d_yx(
x, indices_and_rois, self.roi_size,self.roi_size,
self.spatial_scale)
hres5 = self.res5(pool)
return hres5
def maskhead(self, hres5):
# mask branch
h = F.relu(self.deconvm1(hres5))
masks=self.convm2(h)
return masks
def boxhead(self, hres5):
# box branch
h = F.average_pooling_2d(hres5, self.roi_size//2, stride=7)
roi_cls_locs = self.cls_loc(h)
roi_scores = self.score(h)
return roi_cls_locs, roi_scores
def _roi_pooling_2d_yx(x, indices_and_rois, outh, outw, spatial_scale):
xy_indices_and_rois = indices_and_rois[:, [0, 2, 1, 4, 3]]
pool = F.roi_pooling_2d(
x, xy_indices_and_rois, outh, outw, spatial_scale)
return pool
def _roi_align_2d_yx(x, indices_and_rois, outh, outw, spatial_scale):
xy_indices_and_rois = indices_and_rois[:, [0, 2, 1, 4, 3]]
pool = roi_align_2d.roi_align_2d(
x, xy_indices_and_rois, outh, outw, spatial_scale)
return pool
================================================
FILE: mask_rcnn_train_chain.py
================================================
import numpy as np
import chainer
from chainer import cuda
import chainer.functions as F
from chainercv.links.model.faster_rcnn.utils.anchor_target_creator import AnchorTargetCreator
from utils.proposal_target_creator import ProposalTargetCreator
from chainer import computational_graph as c
from chainercv.links import PixelwiseSoftmaxClassifier
class MaskRCNNTrainChain(chainer.Chain):
def __init__(self, mask_rcnn, rpn_sigma=3., roi_sigma=1., gamma=1,
anchor_target_creator=AnchorTargetCreator(),
roi_size=14):
super(MaskRCNNTrainChain, self).__init__()
with self.init_scope():
self.mask_rcnn = mask_rcnn
self.rpn_sigma = rpn_sigma
self.roi_sigma = roi_sigma
self.anchor_target_creator = anchor_target_creator
self.proposal_target_creator = ProposalTargetCreator(roi_size=roi_size//2)
self.loc_normalize_mean = mask_rcnn.loc_normalize_mean
self.loc_normalize_std = mask_rcnn.loc_normalize_std
self.decayrate=0.99
self.avg_loss = None
self.gamma=gamma
def __call__(self, imgs, bboxes, labels, scale, masks, i):
if isinstance(bboxes, chainer.Variable):
bboxes = bboxes.data
if isinstance(labels, chainer.Variable):
labels = labels.data
if isinstance(scale, chainer.Variable):
scale = scale.data
if isinstance(masks, chainer.Variable):
masks = masks.data
scale = np.asscalar(cuda.to_cpu(scale))
n = bboxes.shape[0]
if n != 1:
raise ValueError('only batch size 1 is supported')
_, _, H, W = imgs.shape
img_size = (H, W)
#Extractor (VGG) : img -> features
with chainer.using_config('train', False):
features = self.mask_rcnn.extractor(imgs)
#Region Proposal Network : features -> rpn_locs, rpn_scores, rois
rpn_locs, rpn_scores, rois, roi_indices, anchor = self.mask_rcnn.rpn(
features, img_size, scale)
bbox, label, mask, rpn_score, rpn_loc, roi = \
bboxes[0], labels[0], masks[0], rpn_scores[0], rpn_locs[0], rois # batch size=1
#proposal target : roi(proposed) , bbox(GT), label(GT) -> sample_roi, gt_roi_loc, gt_roi_label
#the targets are compared with the head output.
sample_roi, gt_roi_loc, gt_roi_label, gt_roi_mask = self.proposal_target_creator(
roi, bbox, label, mask, self.loc_normalize_mean, self.loc_normalize_std)
sample_roi_index = self.xp.zeros((len(sample_roi),), dtype=np.int32)
#Head Network : features, sample_roi -> roi_cls_loc, roi_score
with chainer.using_config('train', False):
hres5 = self.mask_rcnn.head.res5head(features, sample_roi, sample_roi_index)
roi_cls_loc, roi_score = self.mask_rcnn.head.boxhead(hres5)
roi_cls_mask = self.mask_rcnn.head.maskhead(hres5)
del(hres5)
#RPN losses
gt_rpn_loc, gt_rpn_label = self.anchor_target_creator(bbox, anchor, img_size)
rpn_loc_loss = _fast_rcnn_loc_loss(rpn_loc, gt_rpn_loc, gt_rpn_label, self.rpn_sigma)
rpn_cls_loss = F.sigmoid_cross_entropy(rpn_score, gt_rpn_label)
#Head output losses
n_sample = roi_cls_loc.shape[0]
roi_cls_loc = roi_cls_loc.reshape((n_sample, -1, 4))
roi_loc = roi_cls_loc[self.xp.arange(n_sample), gt_roi_label]
roi_mask = roi_cls_mask[self.xp.arange(n_sample), gt_roi_label]
roi_loc_loss = _fast_rcnn_loc_loss(roi_loc, gt_roi_loc, gt_roi_label, self.roi_sigma)
roi_cls_loss = F.softmax_cross_entropy(roi_score, gt_roi_label)
#mask loss: average binary cross-entropy loss
mask_loss = F.sigmoid_cross_entropy(roi_mask[0:gt_roi_mask.shape[0]], gt_roi_mask)
#total loss
loss = rpn_loc_loss + rpn_cls_loss + roi_loc_loss + roi_cls_loss + self.gamma * mask_loss
#avg loss calculation
if self.avg_loss is None:
self.avg_loss = loss.data
else:
self.avg_loss = self.avg_loss * self.decayrate + loss.data*(1-self.decayrate)
chainer.reporter.report({'rpn_loc_loss':rpn_loc_loss,
'rpn_cls_loss':rpn_cls_loss,
'roi_loc_loss':roi_loc_loss,
'roi_cls_loss':roi_cls_loss,
'roi_mask_loss':self.gamma * mask_loss,
'avg_loss':self.avg_loss,
'loss':loss}, self)
return loss
def _smooth_l1_loss(x, t, in_weight, sigma):
sigma2 = sigma ** 2
diff = in_weight * (x - t)
abs_diff = F.absolute(diff)
flag = (abs_diff.data < (1. / sigma2)).astype(np.float32)
y = (flag * (sigma2 / 2.) * F.square(diff) +
(1 - flag) * (abs_diff - 0.5 / sigma2))
return F.sum(y)
def _fast_rcnn_loc_loss(pred_loc, gt_loc, gt_label, sigma):
xp = chainer.cuda.get_array_module(pred_loc)
in_weight = xp.zeros_like(gt_loc)
in_weight[gt_label > 0] = 1
loc_loss = _smooth_l1_loss(pred_loc, gt_loc, in_weight, sigma)
loc_loss /= xp.sum(gt_label >= 0)
return loc_loss
================================================
FILE: mask_rcnn_train_chain_batch.py
================================================
import numpy as np
import chainer
from chainer import cuda
import chainer.functions as F
from chainercv.links.model.faster_rcnn.utils.anchor_target_creator import AnchorTargetCreator
from utils.proposal_target_creator import ProposalTargetCreator
from chainer import computational_graph as c
from chainercv.links import PixelwiseSoftmaxClassifier
class MaskRCNNTrainChain(chainer.Chain):
def __init__(self, mask_rcnn, rpn_sigma=3., roi_sigma=1., gamma=1,
anchor_target_creator=AnchorTargetCreator(),
roi_size=7):
super(MaskRCNNTrainChain, self).__init__()
with self.init_scope():
self.mask_rcnn = mask_rcnn
self.rpn_sigma = rpn_sigma
self.roi_sigma = roi_sigma
self.anchor_target_creator = anchor_target_creator
self.proposal_target_creator = ProposalTargetCreator(roi_size=roi_size)
self.loc_normalize_mean = mask_rcnn.loc_normalize_mean
self.loc_normalize_std = mask_rcnn.loc_normalize_std
self.decayrate=0.99
self.avg_loss = None
self.gamma=gamma
def __call__(self, imgs, bboxes, labels, scale, masks):
if isinstance(bboxes, chainer.Variable):
bboxes = bboxes.data
if isinstance(labels, chainer.Variable):
labels = labels.data
if isinstance(scale, chainer.Variable):
scale = scale.data
if isinstance(masks, chainer.Variable):
masks = masks.data
scale = np.asscalar(cuda.to_cpu(scale[0]))
n = bboxes.shape[0]
#if n != 1:
# raise ValueError('only batch size 1 is supported')
_, _, H, W = imgs.shape
img_size = (H, W)
#Extractor (VGG) : img -> features
features = self.mask_rcnn.extractor(imgs)
#Region Proposal Network : features -> rpn_locs, rpn_scores, rois
rpn_loc_loss,rpn_cls_loss, roi_loc_loss, roi_cls_loss, mask_loss= 0,0,0,0,0
for i in range(n):
rpn_locs, rpn_scores, rois, roi_indices, anchor = self.mask_rcnn.rpn(
features[i:i+1], img_size, scale)
bbox, label, mask, rpn_score, rpn_loc, roi = \
bboxes[i], labels[i], masks[i], rpn_scores[0], rpn_locs[0], rois
mask[mask>1]=0
numdata = sum(label>=0)
label = label[0:numdata]
bbox = bbox[0:numdata]
mask = mask[0:numdata]
#proposal target : roi(proposed) , bbox(GT), label(GT) -> sample_roi, gt_roi_loc, gt_roi_label
#the targets are compared with the head output.
sample_roi, gt_roi_loc, gt_roi_label, gt_roi_mask = self.proposal_target_creator(
roi, bbox, label, mask, self.loc_normalize_mean, self.loc_normalize_std)
sample_roi_index = self.xp.zeros((len(sample_roi),), dtype=np.int32)
#Head Network : features, sample_roi -> roi_cls_loc, roi_score
roi_cls_loc, roi_score, roi_cls_mask = self.mask_rcnn.head(
features[i:i+1], sample_roi, sample_roi_index)
#RPN losses
gt_rpn_loc, gt_rpn_label = self.anchor_target_creator(bbox, anchor, img_size)
rpn_loc_loss += _fast_rcnn_loc_loss(rpn_loc, gt_rpn_loc, gt_rpn_label, self.rpn_sigma)
rpn_cls_loss += F.softmax_cross_entropy(rpn_score, gt_rpn_label)
#Head output losses
n_sample = roi_cls_loc.shape[0]
roi_cls_loc = roi_cls_loc.reshape((n_sample, -1, 4))
roi_loc = roi_cls_loc[self.xp.arange(n_sample), gt_roi_label]
roi_mask = roi_cls_mask[self.xp.arange(n_sample), gt_roi_label]
roi_loc_loss += _fast_rcnn_loc_loss(roi_loc, gt_roi_loc, gt_roi_label, self.roi_sigma)
roi_cls_loss += F.softmax_cross_entropy(roi_score, gt_roi_label)
#mask loss: average binary cross-entropy loss
mask_loss += F.sigmoid_cross_entropy(roi_mask[0:gt_roi_mask.shape[0]], gt_roi_mask)
#total loss
loss = rpn_loc_loss + rpn_cls_loss + roi_loc_loss + roi_cls_loss + self.gamma * mask_loss
loss /= n
#avg loss calculation
if self.avg_loss is None:
self.avg_loss = loss.data
else:
self.avg_loss = self.avg_loss * self.decayrate + loss.data*(1-self.decayrate)
chainer.reporter.report({'rpn_loc_loss':rpn_loc_loss/n,
'rpn_cls_loss':rpn_cls_loss/n,
'roi_loc_loss':roi_loc_loss/n,
'roi_cls_loss':roi_cls_loss/n,
'roi_mask_loss':self.gamma * mask_loss/n,
'avg_loss':self.avg_loss,
'loss':loss}, self)
return loss
def _smooth_l1_loss(x, t, in_weight, sigma):
sigma2 = sigma ** 2
diff = in_weight * (x - t)
abs_diff = F.absolute(diff)
flag = (abs_diff.data < (1. / sigma2)).astype(np.float32)
y = (flag * (sigma2 / 2.) * F.square(diff) +
(1 - flag) * (abs_diff - 0.5 / sigma2))
return F.sum(y)
def _fast_rcnn_loc_loss(pred_loc, gt_loc, gt_label, sigma):
xp = chainer.cuda.get_array_module(pred_loc)
in_weight = xp.zeros_like(gt_loc)
in_weight[gt_label > 0] = 1
loc_loss = _smooth_l1_loss(pred_loc, gt_loc, in_weight, sigma)
loc_loss /= xp.sum(gt_label >= 0)
return loc_loss
================================================
FILE: train.py
================================================
import chainer
from chainer import training
from chainer.training import extensions, ParallelUpdater
from chainer.training.triggers import ManualScheduleTrigger
from chainer.datasets import TransformDataset
from chainercv.datasets import VOCBboxDataset, voc_bbox_label_names
from chainercv import transforms
from chainercv.transforms.image.resize import resize
import argparse
import numpy as np
import time
#from mask_rcnn_vgg import MaskRCNNVGG16
from mask_rcnn_resnet import MaskRCNNResNet
from coco_dataset import COCODataset
from mask_rcnn_train_chain import MaskRCNNTrainChain
from utils.bn_utils import freeze_bn, bn_to_affine
from utils.cocoapi_evaluator import COCOAPIEvaluator
from utils.detection_coco_evaluator import DetectionCOCOEvaluator
import logging
import traceback
from utils.updater import SubDivisionUpdater
import cv2
def resize_bbox(bbox, in_size, out_size):
bbox_o = bbox.copy()
y_scale = float(out_size[0]) / in_size[0]
x_scale = float(out_size[1]) / in_size[1]
bbox_o[:, 0] = y_scale * bbox[:, 1]
bbox_o[:, 2] = y_scale * (bbox[:, 1]+bbox[:, 3])
bbox_o[:, 1] = x_scale * bbox[:, 0]
bbox_o[:, 3] = x_scale * (bbox[:, 0]+bbox[:, 2])
return bbox_o
def parse():
parser = argparse.ArgumentParser(
description='Mask RCNN trainer')
parser.add_argument('--dataset', choices=('coco2017'),
default='coco2017')
parser.add_argument('--extractor', choices=('resnet50','resnet101'),
default='resnet50', help='extractor network')
parser.add_argument('--gpu', '-g', type=int, default=0)
parser.add_argument('--lr', '-l', type=float, default=1e-4)
parser.add_argument('--batchsize', '-b', type=int, default=8)
parser.add_argument('--freeze_bn', action='store_true', default=False, help='freeze batchnorm gamma/beta')
parser.add_argument('--bn2affine', action='store_true', default=False, help='batchnorm to affine')
parser.add_argument('--out', '-o', default='result',
help='Output directory')
parser.add_argument('--seed', '-s', type=int, default=0)
parser.add_argument('--roialign', action='store_false', default=True, help='default: True')
parser.add_argument('--lr_step', '-ls', type=int, default=120000)
parser.add_argument('--lr_initialchange', '-li', type=int, default=400)
parser.add_argument('--pretrained', '-p', type=str, default='imagenet')
parser.add_argument('--snapshot', type=int, default=4000)
parser.add_argument('--validation', type=int, default=30000)
parser.add_argument('--resume', type=str)
parser.add_argument('--iteration', '-i', type=int, default=180000)
parser.add_argument('--roi_size', '-r', type=int, default=14, help='ROI size for mask head input')
parser.add_argument('--gamma', type=float, default=1, help='mask loss weight')
return parser.parse_args()
class Transform(object):
def __init__(self, net, labelids):
self.net = net
self.labelids = labelids
def __call__(self, in_data):
if len(in_data)==5:
img, label, bbox, mask, i = in_data
elif len(in_data)==4:
img, bbox, label, i= in_data
label = [self.labelids.index(l) for l in label]
_, H, W = img.shape
if chainer.config.train:
img = self.net.prepare(img)
_, o_H, o_W = img.shape
scale = o_H / H
if len(bbox)==0:
return img, [],[],1
bbox = resize_bbox(bbox, (H, W), (o_H, o_W))
mask = resize(mask,(o_H, o_W))
if chainer.config.train:
#horizontal flip
img, params = transforms.random_flip(
img, x_random=True, return_param=True)
bbox = transforms.flip_bbox(
bbox, (o_H, o_W), x_flip=params['x_flip'])
mask = transforms.flip(mask, x_flip=params['x_flip'])
return img, bbox, label, scale, mask, i
def convert(batch, device):
return chainer.dataset.convert.concat_examples(batch, device, padding=-1)
def main():
args = parse()
np.random.seed(args.seed)
print('arguments: ', args)
# Model setup
if args.dataset == 'coco2017':
train_data = COCODataset()
test_data = COCODataset(json_file='instances_val2017.json', name='val2017', id_list_file='val2017.txt')
train_class_ids =train_data.class_ids
test_ids = test_data.ids
cocoanns = test_data.coco
if args.extractor=='vgg16':
mask_rcnn = MaskRCNNVGG16(n_fg_class=80, pretrained_model=args.pretrained, roi_size=args.roi_size, roi_align = args.roialign)
elif args.extractor=='resnet50':
mask_rcnn = MaskRCNNResNet(n_fg_class=80, pretrained_model=args.pretrained,roi_size=args.roi_size, n_layers=50, roi_align = args.roialign, class_ids=train_class_ids)
elif args.extractor=='resnet101':
mask_rcnn = MaskRCNNResNet(n_fg_class=80, pretrained_model=args.pretrained,roi_size=args.roi_size, n_layers=101, roi_align = args.roialign, class_ids=train_class_ids)
mask_rcnn.use_preset('evaluate')
model = MaskRCNNTrainChain(mask_rcnn, gamma=args.gamma, roi_size=args.roi_size)
# Trainer setup
if args.gpu >= 0:
chainer.cuda.get_device_from_id(args.gpu).use()
model.to_gpu()
optimizer = chainer.optimizers.MomentumSGD(lr=args.lr, momentum=0.9)
#optimizer = chainer.optimizers.Adam()#alpha=0.001, beta1=0.9, beta2=0.999 , eps=0.00000001)
optimizer.setup(model)
optimizer.add_hook(chainer.optimizer.WeightDecay(rate=0.0001))
train_data=TransformDataset(train_data, Transform(mask_rcnn, train_class_ids))
test_data=TransformDataset(test_data, Transform(mask_rcnn, train_class_ids))
train_iter = chainer.iterators.SerialIterator(
train_data, batch_size=args.batchsize)
test_iter = chainer.iterators.SerialIterator(
test_data, batch_size=1, repeat=False, shuffle=False)
updater = SubDivisionUpdater(train_iter, optimizer, device=args.gpu, subdivisions=args.batchsize)
#updater = ParallelUpdater(train_iter, optimizer, devices={"main": 0, "second": 1}, converter=convert ) #for training with multiple GPUs
trainer = training.Trainer(
updater, (args.iteration, 'iteration'), out=args.out)
# Extensions
trainer.extend(
extensions.snapshot_object(model.mask_rcnn, 'snapshot_model.npz'),
trigger=(args.snapshot, 'iteration'))
trainer.extend(extensions.ExponentialShift('lr', 10),
trigger=ManualScheduleTrigger(
[args.lr_initialchange], 'iteration'))
trainer.extend(extensions.ExponentialShift('lr', 0.1),
trigger=(args.lr_step, 'iteration'))
if args.resume is not None:
chainer.serializers.load_npz(args.resume, model.mask_rcnn)
if args.freeze_bn:
freeze_bn(model.mask_rcnn)
if args.bn2affine:
bn_to_affine(model.mask_rcnn)
log_interval = 40, 'iteration'
plot_interval = 160, 'iteration'
print_interval = 40, 'iteration'
#trainer.extend(extensions.Evaluator(test_iter, model, device=args.gpu), trigger=(args.validation, 'iteration'))
#trainer.extend(DetectionCOCOEvaluator(test_iter, model.mask_rcnn), trigger=(args.validation, 'iteration')) #COCO AP Evaluator with VOC metric
trainer.extend(COCOAPIEvaluator(test_iter, model.mask_rcnn, test_ids, cocoanns), trigger=(args.validation, 'iteration')) #COCO AP Evaluator
trainer.extend(chainer.training.extensions.observe_lr(),
trigger=log_interval)
trainer.extend(extensions.LogReport(trigger=log_interval))
trainer.extend(extensions.PrintReport(
['iteration', 'epoch', 'elapsed_time', 'lr',
'main/loss',
'main/avg_loss',
'main/roi_loc_loss',
'main/roi_cls_loss',
'main/roi_mask_loss',
'main/rpn_loc_loss',
'main/rpn_cls_loss',
'validation/main/loss',
'validation/main/map',
]), trigger=print_interval)
trainer.extend(extensions.ProgressBar(update_interval=1000))
#trainer.extend(extensions.dump_graph('main/loss'))
try:
trainer.run()
except:
traceback.print_exc()
if __name__ == '__main__':
main()
================================================
FILE: utils/__init__.py
================================================
================================================
FILE: utils/bn_utils.py
================================================
import numpy as np
import cupy
def freeze_bn(model):
# freeze batchnorm update
def disableupdate(block):
for name in block._forward:
l = getattr(block, name)
l.bn1.disable_update()
l.bn2.disable_update()
l.bn3.disable_update()
if name=='a':
l.bn4.disable_update()
model.extractor.bn1.disable_update()
disableupdate(model.extractor.res2)
disableupdate(model.extractor.res3)
disableupdate(model.extractor.res4)
disableupdate(model.head.res5)
print("batchnorm update disabled!")
def bn_to_affine(model):
# change batchnorm layers to affine layers (mean -> 0, var -> 1)
def bn_to_affine_block(block):
for name in block._forward:
l = getattr(block, name)
l.bn1.avg_mean = cupy.zeros(l.bn1.avg_mean.shape, dtype=np.float32)
l.bn1.avg_var = cupy.ones(l.bn1.avg_var.shape, dtype=np.float32) - l.bn1.eps
l.bn2.avg_mean = cupy.zeros(l.bn2.avg_mean.shape, dtype=np.float32)
l.bn2.avg_var = cupy.ones(l.bn2.avg_var.shape, dtype=np.float32) - l.bn1.eps
l.bn3.avg_mean = cupy.zeros(l.bn3.avg_mean.shape, dtype=np.float32)
l.bn3.avg_var = cupy.ones(l.bn3.avg_var.shape, dtype=np.float32) - l.bn1.eps
if name=='a':
l.bn4.avg_mean = cupy.zeros(l.bn4.avg_mean.shape, dtype=np.float32)
l.bn4.avg_var = cupy.ones(l.bn4.avg_var.shape, dtype=np.float32) - l.bn1.eps
model.extractor.bn1.avg_mean = cupy.zeros(model.extractor.bn1.avg_mean.shape, dtype=np.float32)
model.extractor.bn1.avg_var = cupy.ones(model.extractor.bn1.avg_var.shape, dtype=np.float32) - model.extractor.bn1.eps
bn_to_affine_block(model.extractor.res2)
bn_to_affine_block(model.extractor.res3)
bn_to_affine_block(model.extractor.res4)
bn_to_affine_block(model.head.res5)
print("converted batchnorm to affine")
================================================
FILE: utils/box_utils.py
================================================
import numpy as np
import cupy
import cv2
def resize_bbox(bbox, in_size, out_size):
bbox_o = bbox.copy()
y_scale = float(out_size[0]) / in_size[0]
x_scale = float(out_size[1]) / in_size[1]
bbox_o[:, 0] = y_scale * bbox[:, 1]
bbox_o[:, 2] = y_scale * (bbox[:, 1]+bbox[:, 3])
bbox_o[:, 1] = x_scale * bbox[:, 0]
bbox_o[:, 3] = x_scale * (bbox[:, 0]+bbox[:, 2])
return bbox_o
def bbox_yxyx2xywh(bbox):
bbox_o = bbox.copy()
bbox_o[:, 0] = bbox[:, 1]
bbox_o[:, 2] = bbox[:, 3] - bbox[:, 1]
bbox_o[:, 1] = bbox[:, 0]
bbox_o[:, 3] = bbox[:, 2] - bbox[:, 0]
return bbox_o
def im_mask(mask, size, bbox):
# bboxes are already clipped to [0, w], [0, h]
masksize = mask.shape[0]
# pad the mask to avoid cv2.resize artifacts
pmask = np.zeros((masksize + 2, masksize + 2), dtype=np.float32)
pmask[1:-1, 1:-1] = mask
# extend the boxhead
scale = (masksize + 2) / masksize
ex_w = (bbox[3] - bbox[1]) * scale
ex_h = (bbox[2] - bbox[0]) * scale
ex_x0 = (bbox[3] + bbox[1] - ex_w) / 2
ex_y0 = (bbox[2] + bbox[0] - ex_h) / 2
ex_x1 = (bbox[3] + bbox[1] + ex_w) / 2
ex_y1 = (bbox[2] + bbox[0] + ex_h) / 2
ex_bbox = np.asarray([ex_y0, ex_x0, ex_y1, ex_x1], dtype=np.int32)
# whole-image-sized mask
immask = np.zeros((size[0],size[1]), dtype=np.uint8)
x0, x1 = max(ex_bbox[1], 0), min(ex_bbox[3] + 1, size[1])
y0, y1= max(ex_bbox[0], 0), min(ex_bbox[2] + 1, size[0])
immask_roi = cv2.resize(pmask, (x1 - x0, y1 - y0))
immask[y0:y1, x0:x1] = np.round(immask_roi).astype(np.uint8)
return immask
================================================
FILE: utils/cocoapi_evaluator.py
================================================
import copy
import numpy as np
from chainer import reporter
import chainer.training.extensions
from utils import eval_detection_coco
from chainercv.utils import apply_prediction_to_iterator
import pycocotools
from pycocotools.coco import COCO
from pycocotools.cocoeval import COCOeval
class COCOAPIEvaluator(chainer.training.extensions.Evaluator):
trigger = 1, 'epoch'
default_name = 'validation'
priority = chainer.training.PRIORITY_WRITER
def __init__(
self, iterator, target, ids, cocoanns, label_names=None):
super(COCOAPIEvaluator, self).__init__(
iterator, target)
self.ids = ids
self.cocoanns = cocoanns
def evaluate(self):
iterator = self._iterators['main']
target = self._targets['main']
annType = ['segm','bbox','keypoints']
if hasattr(iterator, 'reset'):
iterator.reset()
it = iterator
else:
it = copy.copy(iterator)
in_values, out_values, rest_values = apply_prediction_to_iterator(
target.predict, it)
# delete unused iterators explicitly
del in_values
pred_bboxes, pred_labels, pred_scores, pred_masks = out_values
if len(rest_values) == 3:
gt_bboxes, gt_labels, gt_difficults = rest_values
elif len(rest_values) == 2:
gt_bboxes, gt_labels = rest_values
gt_difficults = None
elif len(rest_values) == 5:
gt_bboxes, gt_labels, _, _, i = rest_values
gt_difficults = None
pred_bboxes = iter(list(pred_bboxes))
pred_labels = iter(list(pred_labels))
pred_scores = iter(list(pred_scores))
gt_bboxes = iter(list(gt_bboxes))
gt_labels = iter(list(gt_labels))
data_dict = []
for i, (pred_bbox, pred_label, pred_score, pred_mask) in \
enumerate(zip(pred_bboxes, pred_labels, pred_scores, pred_masks)):
for bbox, label, score, mask in zip(pred_bbox, pred_label, pred_score, pred_mask):
A={"image_id":int(self.ids[i]), "category_id":int(label), "bbox":bbox.tolist(),
"score":float(score), "segmentation": mask}
data_dict.append(A)
if len(data_dict)>0:
for i in range(2): # 'segm','bbox'
cocoGt=self.cocoanns
cocoDt=cocoGt.loadRes(data_dict)
cocoEval = COCOeval(self.cocoanns, cocoDt, annType[i])
cocoEval.params.imgIds = [int(id_) for id_ in self.ids]
cocoEval.evaluate()
cocoEval.accumulate()
cocoEval.summarize()
report = {'map': cocoEval.stats[0]} # report COCO AP (IoU=0.5:0:95)
else:
report = {'map': 0}
observation = {}
with reporter.report_scope(observation):
reporter.report(report, target)
return observation
================================================
FILE: utils/detection_coco_evaluator.py
================================================
import copy
import numpy as np
from chainer import reporter
import chainer.training.extensions
from utils import eval_detection_coco
from chainercv.utils import apply_prediction_to_iterator
class DetectionCOCOEvaluator(chainer.training.extensions.Evaluator):
"""An extension that evaluates a detection model by PASCAL VOC metric.
This extension iterates over an iterator and evaluates the prediction
results by average precisions (APs) and mean of them
(mean Average Precision, mAP).
This extension reports the following values with keys.
Please note that :obj:`'ap/<label_names[l]>'` is reported only if
:obj:`label_names` is specified.
* :obj:`'map'`: Mean of average precisions (mAP).
* :obj:`'ap/<label_names[l]>'`: Average precision for class \
:obj:`label_names[l]`, where :math:`l` is the index of the class. \
For example, this evaluator reports :obj:`'ap/aeroplane'`, \
:obj:`'ap/bicycle'`, etc. if :obj:`label_names` is \
:obj:`~chainercv.datasets.voc_bbox_label_names`. \
If there is no bounding box assigned to class :obj:`label_names[l]` \
in either ground truth or prediction, it reports :obj:`numpy.nan` as \
its average precision. \
In this case, mAP is computed without this class.
Args:
iterator (chainer.Iterator): An iterator. Each sample should be
following tuple :obj:`img, bbox, label` or
:obj:`img, bbox, label, difficult`.
:obj:`img` is an image, :obj:`bbox` is coordinates of bounding
boxes, :obj:`label` is labels of the bounding boxes and
:obj:`difficult` is whether the bounding boxes are difficult or
not. If :obj:`difficult` is returned, difficult ground truth
will be ignored from evaluation.
target (chainer.Link): A detection link. This link must have
:meth:`predict` method that takes a list of images and returns
:obj:`bboxes`, :obj:`labels` and :obj:`scores`.
use_07_metric (bool): Whether to use PASCAL VOC 2007 evaluation metric
for calculating average precision. The default value is
:obj:`False`.
label_names (iterable of strings): An iterable of names of classes.
If this value is specified, average precision for each class is
also reported with the key :obj:`'ap/<label_names[l]>'`.
"""
trigger = 1, 'epoch'
default_name = 'validation'
priority = chainer.training.PRIORITY_WRITER
def __init__(
self, iterator, target, use_07_metric=False, label_names=None):
super(DetectionCOCOEvaluator, self).__init__(
iterator, target)
self.use_07_metric = use_07_metric
self.label_names = ['background', # class zero
'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light',
'fire hydrant', 'street sign', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow',
'elephant', 'bear', 'zebra', 'giraffe', 'hat', 'backpack', 'umbrella', 'shoe', 'eye glasses', 'handbag', 'tie', 'suitcase', 'frisbee',
'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard', 'tennis racket', 'bottle',
'plate', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', 'sandwich', 'orange',
'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch', 'potted plant', 'bed',
'mirror', 'dining table', 'window', 'desk','toilet', 'door', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', 'microwave', 'oven',
'toaster', 'sink', 'refrigerator', 'blender', 'book', 'clock', 'vase', 'scissors', 'teddy bear', 'hair drier', 'toothbrush']
def evaluate(self):
iterator = self._iterators['main']
target = self._targets['main']
if hasattr(iterator, 'reset'):
iterator.reset()
it = iterator
else:
it = copy.copy(iterator)
in_values, out_values, rest_values = apply_prediction_to_iterator(
target.predict, it)
# delete unused iterators explicitly
del in_values
pred_bboxes, _, pred_labels, pred_scores, _ = out_values
if len(rest_values) == 3:
gt_bboxes, gt_labels, gt_difficults = rest_values
elif len(rest_values) == 2:
gt_bboxes, gt_labels = rest_values
gt_difficults = None
elif len(rest_values) == 5:
gt_bboxes, gt_labels, _, _, i = rest_values
gt_difficults = None
result = eval_detection_coco.eval_detection_coco(
pred_bboxes, pred_labels, pred_scores,
gt_bboxes, gt_labels, gt_difficults,
use_07_metric=self.use_07_metric)
report = {'map': result['map']}
if self.label_names is not None:
for l, label_name in enumerate(self.label_names):
try:
report['ap/{:s}'.format(label_name)] = result['ap'][l]
except IndexError:
report['ap/{:s}'.format(label_name)] = np.nan
if True:
print(report)
observation = {}
with reporter.report_scope(observation):
reporter.report(report, target)
return observation
================================================
FILE: utils/detectron_parser.py
================================================
import numpy as np
import os
path = os.path.join(os.path.dirname(__file__), '../')
import sys
sys.path.append(path)
from mask_rcnn_resnet import MaskRCNNResNet
from chainer import serializers
import pickle
model = MaskRCNNResNet(n_fg_class=80, roi_size=14, pretrained_model='auto', anchor_scales=[2, 4, 8, 16, 32], n_layers=50, class_ids=[[1]])
modeldir = "modelfiles"
if os.path.exists(modeldir)==False:
os.mkdir(modeldir)
# resnet50, end-to-end, C4
d_model_file = "modelfiles/model_final.pkl"
c_model_file = "modelfiles/e2e_mask_rcnn_R-50-C4_1x_d2c.npz"
with open(d_model_file, 'rb') as f:
d = pickle.load(f, encoding='latin-1')['blobs']
d_key = sorted(d)
parsecount = 0
for bl in d_key:
if 'res' in bl:
stage = bl[3] # resnet stage, 2, 3, 4, 5
block = bl[5] # resnet block, a or b
if stage=='_': # non-resnet layers
continue
else:
stage = int(stage) - 1
if stage == 4:
netname='head'
else:
netname='extractor'
if 'branch2a' in bl:
c_nlayer = 1
elif 'branch2b' in bl:
c_nlayer = 2
elif 'branch2c' in bl:
c_nlayer = 3
elif 'branch1' in bl:
c_nlayer = 4
else:
c_nlayer = 0
# do not copy
if bl.endswith('_b') and 'bn_b' not in bl:
continue
if 'momentum' in bl:
continue
# conv / bn gamma / bn beta
if '_w' in bl:
c_kind = 'conv%d.W' % c_nlayer
elif 'bn_s' in bl:
c_kind = 'bn%d.gamma' % c_nlayer
elif 'bn_b' in bl:
c_kind = 'bn%d.beta' % c_nlayer
# chainer block kind
if block == '0':
c_block = 'a'
else:
c_block = 'b'+block
# shape checker
exec("c_shape = model.%s.res%d.%s.%s.data.shape" % (netname, stage + 1, c_block, c_kind))
exec("d_shape = d['%s'].shape" % bl)
if c_shape == d_shape:
# execute copy
txt = "model.%s.res%d.%s.%s.data = d['%s']" % (netname, stage + 1, c_block, c_kind, bl )
print(txt)
exec(txt)
parsecount += 1
else:
print("shape mismatch error!")
# copy the other layers
layer_pairs = \
[('extractor.conv1.W', 'conv1_w'), ('extractor.bn1.gamma', 'res_conv1_bn_s'), ('extractor.bn1.beta', 'res_conv1_bn_b'),
('rpn.conv1.W', 'conv_rpn_w'), ('rpn.conv1.b', 'conv_rpn_b'),
('rpn.loc.W', 'rpn_bbox_pred_w'), ('rpn.loc.b', 'rpn_bbox_pred_b'),
('rpn.score.W', 'rpn_cls_logits_w'), ('rpn.score.b', 'rpn_cls_logits_b'),
('head.score.W', 'cls_score_w'), ('head.score.b', 'cls_score_b'),
('head.cls_loc.W', 'bbox_pred_w'), ('head.cls_loc.b', 'bbox_pred_b'),
('head.deconvm1.W', 'conv5_mask_w'), ('head.deconvm1.b', 'conv5_mask_b'),
('head.convm2.W', 'mask_fcn_logits_w'), ('head.convm2.b', 'mask_fcn_logits_b'),
]
def xytrans(src):
sh = src.shape
dst = src.reshape(sh[0]//4, 4, -1)[:,[1, 0, 3, 2]].reshape(sh)
return dst
for layer_pair in layer_pairs:
exec("c_shape = model.%s.data.shape" % layer_pair[0])
exec("d_shape = d['%s'].shape" % layer_pair[1])
if 'bbox_pred' in layer_pair[1]:
d[layer_pair[1]] = xytrans(d[layer_pair[1]])
if c_shape == d_shape:
txt = "model.%s.data = d['%s']" % layer_pair
print(txt)
exec(txt)
parsecount += 1
else:
print("shape mismatch error!")
print(parsecount, " layers copied")
serializers.save_npz(c_model_file, model)
print("save weights file to a chainer model", c_model_file)
================================================
FILE: utils/eval_detection_coco.py
================================================
from __future__ import division
from collections import defaultdict
import itertools
import numpy as np
import six
from chainercv.utils.bbox.bbox_iou import bbox_iou
def eval_detection_coco(
pred_bboxes, pred_labels, pred_scores, gt_bboxes, gt_labels,
gt_difficults=None,
iou_thresh=0.5, use_07_metric=False):
"""Calculate average precisions based on evaluation code of PASCAL VOC.
This function evaluates predicted bounding boxes obtained from a dataset
which has :math:`N` images by using average precision for each class.
The code is based on the evaluation code used in PASCAL VOC Challenge.
Args:
pred_bboxes (iterable of numpy.ndarray): An iterable of :math:`N`
sets of bounding boxes.
Its index corresponds to an index for the base dataset.
Each element of :obj:`pred_bboxes` is a set of coordinates
of bounding boxes. This is an array whose shape is :math:`(R, 4)`,
where :math:`R` corresponds
to the number of bounding boxes, which may vary among boxes.
The second axis corresponds to
:math:`y_{min}, x_{min}, y_{max}, x_{max}` of a bounding box.
pred_labels (iterable of numpy.ndarray): An iterable of labels.
Similar to :obj:`pred_bboxes`, its index corresponds to an
index for the base dataset. Its length is :math:`N`.
pred_scores (iterable of numpy.ndarray): An iterable of confidence
scores for predicted bounding boxes. Similar to :obj:`pred_bboxes`,
its index corresponds to an index for the base dataset.
Its length is :math:`N`.
gt_bboxes (iterable of numpy.ndarray): An iterable of ground truth
bounding boxes
whose length is :math:`N`. An element of :obj:`gt_bboxes` is a
bounding box whose shape is :math:`(R, 4)`. Note that the number of
bounding boxes in each image does not need to be same as the number
of corresponding predicted boxes.
gt_labels (iterable of numpy.ndarray): An iterable of ground truth
labels which are organized similarly to :obj:`gt_bboxes`.
gt_difficults (iterable of numpy.ndarray): An iterable of boolean
arrays which is organized similarly to :obj:`gt_bboxes`.
This tells whether the
corresponding ground truth bounding box is difficult or not.
By default, this is :obj:`None`. In that case, this function
considers all bounding boxes to be not difficult.
iou_thresh (float): A prediction is correct if its Intersection over
Union with the ground truth is above this value.
use_07_metric (bool): Whether to use PASCAL VOC 2007 evaluation metric
for calculating average precision. The default value is
:obj:`False`.
Returns:
dict:
The keys, value-types and the description of the values are listed
below.
* **ap** (*numpy.ndarray*): An array of average precisions. \
The :math:`l`-th value corresponds to the average precision \
for class :math:`l`. If class :math:`l` does not exist in \
either :obj:`pred_labels` or :obj:`gt_labels`, the corresponding \
value is set to :obj:`numpy.nan`.
* **map** (*float*): The average of Average Precisions over classes.
"""
prec, rec = calc_detection_coco_prec_rec(
pred_bboxes, pred_labels, pred_scores,
gt_bboxes, gt_labels, gt_difficults,
iou_thresh=iou_thresh)
ap = calc_detection_coco_ap(prec, rec, use_07_metric=use_07_metric)
#for name, ap0 in zip(coconames, ap):
# if ~(ap0==ap0):
# ap0 = -1
# apresults.append([name, ap0])
#print("average precision evaluation results: ", apresults)
return {'ap': ap, 'map': np.nanmean(ap)}
def calc_detection_coco_prec_rec(
pred_bboxes, pred_labels, pred_scores, gt_bboxes, gt_labels,
gt_difficults=None,
iou_thresh=0.5):
"""Calculate precision and recall based on evaluation code of PASCAL VOC.
This function calculates precision and recall of
predicted bounding boxes obtained from a dataset which has :math:`N`
images.
The code is based on the evaluation code used in PASCAL VOC Challenge.
Args:
pred_bboxes (iterable of numpy.ndarray): An iterable of :math:`N`
sets of bounding boxes.
Its index corresponds to an index for the base dataset.
Each element of :obj:`pred_bboxes` is a set of coordinates
of bounding boxes. This is an array whose shape is :math:`(R, 4)`,
where :math:`R` corresponds
to the number of bounding boxes, which may vary among boxes.
The second axis corresponds to
:math:`y_{min}, x_{min}, y_{max}, x_{max}` of a bounding box.
pred_labels (iterable of numpy.ndarray): An iterable of labels.
Similar to :obj:`pred_bboxes`, its index corresponds to an
index for the base dataset. Its length is :math:`N`.
pred_scores (iterable of numpy.ndarray): An iterable of confidence
scores for predicted bounding boxes. Similar to :obj:`pred_bboxes`,
its index corresponds to an index for the base dataset.
Its length is :math:`N`.
gt_bboxes (iterable of numpy.ndarray): An iterable of ground truth
bounding boxes
whose length is :math:`N`. An element of :obj:`gt_bboxes` is a
bounding box whose shape is :math:`(R, 4)`. Note that the number of
bounding boxes in each image does not need to be same as the number
of corresponding predicted boxes.
gt_labels (iterable of numpy.ndarray): An iterable of ground truth
labels which are organized similarly to :obj:`gt_bboxes`.
gt_difficults (iterable of numpy.ndarray): An iterable of boolean
arrays which is organized similarly to :obj:`gt_bboxes`.
This tells whether the
corresponding ground truth bounding box is difficult or not.
By default, this is :obj:`None`. In that case, this function
considers all bounding boxes to be not difficult.
iou_thresh (float): A prediction is correct if its Intersection over
Union with the ground truth is above this value..
Returns:
tuple of two lists:
This function returns two lists: :obj:`prec` and :obj:`rec`.
* :obj:`prec`: A list of arrays. :obj:`prec[l]` is precision \
for class :math:`l`. If class :math:`l` does not exist in \
either :obj:`pred_labels` or :obj:`gt_labels`, :obj:`prec[l]` is \
set to :obj:`None`.
* :obj:`rec`: A list of arrays. :obj:`rec[l]` is recall \
for class :math:`l`. If class :math:`l` that is not marked as \
difficult does not exist in \
:obj:`gt_labels`, :obj:`rec[l]` is \
set to :obj:`None`.
"""
pred_bboxes = iter(list(pred_bboxes))
pred_labels = iter(list(pred_labels))
pred_scores = iter(list(pred_scores))
gt_bboxes = iter(list(gt_bboxes))
gt_labels = iter(list(gt_labels))
if gt_difficults is None:
gt_difficults = itertools.repeat(None)
else:
gt_difficults = iter(gt_difficults)
n_pos = defaultdict(int)
score = defaultdict(list)
match = defaultdict(list)
for pred_bbox, pred_label, pred_score, gt_bbox, gt_label, gt_difficult in \
six.moves.zip(
pred_bboxes, pred_labels, pred_scores,
gt_bboxes, gt_labels, gt_difficults):
if gt_difficult is None:
gt_difficult = np.zeros(gt_bbox.shape[0], dtype=bool)
for l in np.unique(np.concatenate((pred_label, gt_label)).astype(int)):
pred_mask_l = pred_label == l
pred_bbox_l = pred_bbox[pred_mask_l]
pred_score_l = pred_score[pred_mask_l]
# sort by score
order = pred_score_l.argsort()[::-1]
pred_bbox_l = pred_bbox_l[order]
pred_score_l = pred_score_l[order]
gt_mask_l = gt_label == l
gt_bbox_l = gt_bbox[gt_mask_l]
gt_difficult_l = gt_difficult[gt_mask_l]
n_pos[l] += np.logical_not(gt_difficult_l).sum()
score[l].extend(pred_score_l)
if len(pred_bbox_l) == 0:
continue
if len(gt_bbox_l) == 0:
match[l].extend((0,) * pred_bbox_l.shape[0])
continue
# VOC evaluation follows integer typed bounding boxes.
pred_bbox_l = pred_bbox_l.copy()
pred_bbox_l[:, 2:] += 1
gt_bbox_l = gt_bbox_l.copy()
gt_bbox_l[:, 2:] += 1
iou = bbox_iou(pred_bbox_l, gt_bbox_l)
gt_index = iou.argmax(axis=1)
# set -1 if there is no matching ground truth
gt_index[iou.max(axis=1) < iou_thresh] = -1
del iou
selec = np.zeros(gt_bbox_l.shape[0], dtype=bool)
for gt_idx in gt_index:
if gt_idx >= 0:
if gt_difficult_l[gt_idx]:
match[l].append(-1)
else:
if not selec[gt_idx]:
match[l].append(1)
else:
match[l].append(0)
selec[gt_idx] = True
else:
match[l].append(0)
for iter_ in (
pred_bboxes, pred_labels, pred_scores,
gt_bboxes, gt_labels, gt_difficults):
if next(iter_, None) is not None:
raise ValueError('Length of input iterables need to be same.')
n_fg_class = max(n_pos.keys()) + 1
prec = [None] * n_fg_class
rec = [None] * n_fg_class
for l in n_pos.keys():
score_l = np.array(score[l])
match_l = np.array(match[l], dtype=np.int8)
order = score_l.argsort()[::-1]
match_l = match_l[order]
tp = np.cumsum(match_l == 1)
fp = np.cumsum(match_l == 0)
# If an element of fp + tp is 0,
# the corresponding element of prec[l] is nan.
prec[l] = tp / (fp + tp)
# If n_pos[l] is 0, rec[l] is None.
if n_pos[l] > 0:
rec[l] = tp / n_pos[l]
return prec, rec
def calc_detection_coco_ap(prec, rec, use_07_metric=False):
"""Calculate average precisions based on evaluation code of PASCAL VOC.
This function calculates average precisions
from given precisions and recalls.
The code is based on the evaluation code used in PASCAL VOC Challenge.
Args:
prec (list of numpy.array): A list of arrays.
:obj:`prec[l]` indicates precision for class :math:`l`.
If :obj:`prec[l]` is :obj:`None`, this function returns
:obj:`numpy.nan` for class :math:`l`.
rec (list of numpy.array): A list of arrays.
:obj:`rec[l]` indicates recall for class :math:`l`.
If :obj:`rec[l]` is :obj:`None`, this function returns
:obj:`numpy.nan` for class :math:`l`.
use_07_metric (bool): Whether to use PASCAL VOC 2007 evaluation metric
for calculating average precision. The default value is
:obj:`False`.
Returns:
~numpy.ndarray:
This function returns an array of average precisions.
The :math:`l`-th value corresponds to the average precision
for class :math:`l`. If :obj:`prec[l]` or :obj:`rec[l]` is
:obj:`None`, the corresponding value is set to :obj:`numpy.nan`.
"""
n_fg_class = len(prec)
ap = np.empty(n_fg_class)
for l in six.moves.range(n_fg_class):
if prec[l] is None or rec[l] is None:
ap[l] = np.nan
continue
if use_07_metric:
# 11 point metric
ap[l] = 0
for t in np.arange(0., 1.1, 0.1):
if np.sum(rec[l] >= t) == 0:
p = 0
else:
p = np.max(np.nan_to_num(prec[l])[rec[l] >= t])
ap[l] += p / 11
else:
# correct AP calculation
# first append sentinel values at the end
mpre = np.concatenate(([0], np.nan_to_num(prec[l]), [0]))
mrec = np.concatenate(([0], rec[l], [1]))
mpre = np.maximum.accumulate(mpre[::-1])[::-1]
# to calculate area under PR curve, look for points
# where X axis (recall) changes value
i = np.where(mrec[1:] != mrec[:-1])[0]
# and sum (\Delta recall) * prec
ap[l] = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1])
return ap
================================================
FILE: utils/makecocolist.py
================================================
import glob
fnames = glob.glob('COCO/train2017/*.jpg')
with open("COCO/train2017.txt", "w") as f:
for fname in fnames:
f.write(fname.split('/')[-1].split('.')[0]+'\n')
f.close()
fnames = glob.glob('COCO/val2017/*.jpg')
with open("COCO/val2017.txt", "w") as f:
for i, fname in enumerate(fnames):
f.write(fname.split('/')[-1].split('.')[0]+'\n')
if i > 1000:
break
f.close()
================================================
FILE: utils/proposal_target_creator.py
================================================
import numpy as np
from chainer import cuda
from chainercv.links.model.faster_rcnn.utils.bbox2loc import bbox2loc
from chainercv.utils.bbox.bbox_iou import bbox_iou
import cv2
class ProposalTargetCreator(object):
"""Assign ground truth bounding boxes to given RoIs.
The :meth:`__call__` of this class generates training targets
for each object proposal.
This is used to train Faster RCNN [#]_.
.. [#] Shaoqing Ren, Kaiming He, Ross Girshick, Jian Sun. \
Faster R-CNN: Towards Real-Time Object Detection with \
Region Proposal Networks. NIPS 2015.
Args:
n_sample (int): The number of sampled regions.
pos_ratio (float): Fraction of regions that is labeled as a
foreground.
pos_iou_thresh (float): IoU threshold for a RoI to be considered as a
foreground.
neg_iou_thresh_hi (float): RoI is considered to be the background
if IoU is in
[:obj:`neg_iou_thresh_hi`, :obj:`neg_iou_thresh_hi`).
neg_iou_thresh_lo (float): See above.
"""
def __init__(self,
n_sample=128,
pos_ratio=0.25, pos_iou_thresh=0.5,
neg_iou_thresh_hi=0.5, neg_iou_thresh_lo=0.0,
roi_size=7
):
self.roi_size=roi_size
self.n_sample = n_sample
self.pos_ratio = pos_ratio
self.pos_iou_thresh = pos_iou_thresh
self.neg_iou_thresh_hi = neg_iou_thresh_hi
self.neg_iou_thresh_lo = neg_iou_thresh_lo
def __call__(self, roi, bbox, label, mask,
loc_normalize_mean=(0., 0., 0., 0.),
loc_normalize_std=(0.1, 0.1, 0.2, 0.2)):
"""Assigns ground truth to sampled proposals.
This function samples total of :obj:`self.n_sample` RoIs
from the combination of :obj:`roi` and :obj:`bbox`.
The RoIs are assigned with the ground truth class labels as well as
bounding box offsets and scales to match the ground truth bounding
boxes. As many as :obj:`pos_ratio * self.n_sample` RoIs are
sampled as foregrounds.
Offsets and scales of bounding boxes are calculated using
:func:`chainercv.links.model.faster_rcnn.bbox2loc`.
Also, types of input arrays and output arrays are same.
Here are notations.
* :math:`S` is the total number of sampled RoIs, which equals \
:obj:`self.n_sample`.
* :math:`L` is number of object classes possibly including the \
background.
Args:
roi (array): Region of Interests (RoIs) from which we sample.
Its shape is :math:`(R, 4)`
bbox (array): The coordinates of ground truth bounding boxes.
Its shape is :math:`(R', 4)`.
label (array): Ground truth bounding box labels. Its shape
is :math:`(R',)`. Its range is :math:`[0, L - 1]`, where
:math:`L` is the number of foreground classes.
loc_normalize_mean (tuple of four floats): Mean values to normalize
coordinates of bouding boxes.
loc_normalize_std (tupler of four floats): Standard deviation of
the coordinates of bounding boxes.
Returns:
(array, array, array):
* **sample_roi**: Regions of interests that are sampled. \
Its shape is :math:`(S, 4)`.
* **gt_roi_loc**: Offsets and scales to match \
the sampled RoIs to the ground truth bounding boxes. \
Its shape is :math:`(S, 4)`.
* **gt_roi_label**: Labels assigned to sampled RoIs. Its shape is \
:math:`(S,)`. Its range is :math:`[0, L]`. The label with \
value 0 is the background.
"""
xp = cuda.get_array_module(roi)
roi = cuda.to_cpu(roi)
bbox = cuda.to_cpu(bbox)
label = cuda.to_cpu(label)
mask = cuda.to_cpu(mask)
n_bbox, _ = bbox.shape
roi = np.concatenate((roi, bbox), axis=0)
pos_roi_per_image = np.round(self.n_sample * self.pos_ratio)
iou = bbox_iou(roi, bbox)
gt_assignment = iou.argmax(axis=1)
max_iou = iou.max(axis=1)
# Offset range of classes from [0, n_fg_class - 1] to [1, n_fg_class].
# The label with value 0 is the background.
gt_roi_label = label[gt_assignment] + 1
# Select foreground RoIs as those with >= pos_iou_thresh IoU.
pos_index = np.where(max_iou >= self.pos_iou_thresh)[0]
pos_roi_per_this_image = int(min(pos_roi_per_image, pos_index.size))
if pos_index.size > 0:
pos_index = np.random.choice(
pos_index, size=pos_roi_per_this_image, replace=False)
# Select background RoIs as those within
# [neg_iou_thresh_lo, neg_iou_thresh_hi).
neg_index = np.where((max_iou < self.neg_iou_thresh_hi) &
(max_iou >= self.neg_iou_thresh_lo))[0]
neg_roi_per_this_image = self.n_sample - pos_roi_per_this_image
neg_roi_per_this_image = int(min(neg_roi_per_this_image,
neg_index.size))
if neg_index.size > 0:
neg_index = np.random.choice(
neg_index, size=neg_roi_per_this_image, replace=False)
# The indices that we're selecting (both positive and negative).
keep_index = np.append(pos_index, neg_index)
gt_roi_label = gt_roi_label[keep_index]
gt_roi_label[pos_roi_per_this_image:] = 0 # negative labels --> 0
sample_roi = roi[keep_index]# sampled <- proposed
# Compute offsets and scales to match sampled RoIs to the GTs.
gt_roi_loc = bbox2loc(sample_roi, bbox[gt_assignment[keep_index]])
gt_roi_loc = ((gt_roi_loc - np.array(loc_normalize_mean, np.float32)
) / np.array(loc_normalize_std, np.float32))
# Prepare groundtruth masks
gt_roi_mask=[]
_, h, w = mask.shape
for i , idx in enumerate(gt_assignment[pos_index]):
A=mask[idx, np.max((int(sample_roi[i,0]),0)):np.min((int(sample_roi[i,2]),h)), np.max((int(sample_roi[i,1]),0)):np.min((int(sample_roi[i,3]),w))]
gt_roi_mask.append(cv2.resize(A, (self.roi_size*2,self.roi_size*2)))
#debug: visualize masks
#cv2.imwrite("gt_assignment_mask.png",mask[0,np.max((int(sample_roi[0,0]),0)):np.min((int(sample_roi[0,2]),h)), np.max((int(sample_roi[0,1]),0)):np.min((int(sample_roi[0,3]),w))]*255)
#cv2.imwrite("gt_roi_mask.png",gt_roi_mask[0]*244)#
if xp != np:
sample_roi = cuda.to_gpu(sample_roi)
gt_roi_loc = cuda.to_gpu(gt_roi_loc)
gt_roi_label = cuda.to_gpu(gt_roi_label)
gt_roi_mask = cuda.to_gpu(np.stack(gt_roi_mask).astype(np.int32))
else:
gt_roi_mask = np.stack(gt_roi_mask).astype(np.int32)
return sample_roi, gt_roi_loc, gt_roi_label, gt_roi_mask
================================================
FILE: utils/region_proposal_network.py
================================================
import numpy as np
import chainer
from chainer import cuda
import chainer.functions as F
import chainer.links as L
from chainercv.links.model.faster_rcnn.utils.generate_anchor_base import \
generate_anchor_base
from chainercv.links.model.faster_rcnn.utils.proposal_creator import \
ProposalCreator
class RegionProposalNetwork(chainer.Chain):
"""Region Proposal Network introduced in Faster R-CNN.
This is Region Proposal Network introduced in Faster R-CNN [#]_.
This takes features extracted from images and propose
class agnostic bounding boxes around "objects".
.. [#] Shaoqing Ren, Kaiming He, Ross Girshick, Jian Sun. \
Faster R-CNN: Towards Real-Time Object Detection with \
Region Proposal Networks. NIPS 2015.
Args:
in_channels (int): The channel size of input.
mid_channels (int): The channel size of the intermediate tensor.
ratios (list of floats): This is ratios of width to height of
the anchors.
anchor_scales (list of numbers): This is areas of anchors.
Those areas will be the product of the square of an element in
:obj:`anchor_scales` and the original area of the reference
window.
feat_stride (int): Stride size after extracting features from an
image.
initialW (callable): Initial weight value. If :obj:`None` then this
function uses Gaussian distribution scaled by 0.1 to
initialize weight.
May also be a callable that takes an array and edits its values.
proposal_creator_params (dict): Key valued paramters for
:class:`~chainercv.links.model.faster_rcnn.ProposalCreator`.
.. seealso::
:class:`~chainercv.links.model.faster_rcnn.ProposalCreator`
"""
def __init__(
self, in_channels=512, mid_channels=512, ratios=[0.5, 1, 2],
anchor_scales=[8, 16, 32], feat_stride=16,
initialW=None,
proposal_creator_params={},
):
self.anchor_base = generate_anchor_base(
anchor_scales=anchor_scales, ratios=ratios)
self.feat_stride = feat_stride
self.proposal_layer = ProposalCreator(**proposal_creator_params)
n_anchor = self.anchor_base.shape[0]
super(RegionProposalNetwork, self).__init__()
with self.init_scope():
self.conv1 = L.Convolution2D(
in_channels, mid_channels, 3, 1, 1, initialW=initialW)
self.score = L.Convolution2D(
mid_channels, n_anchor * 1, 1, 1, 0, initialW=initialW)
self.loc = L.Convolution2D(
mid_channels, n_anchor * 4, 1, 1, 0, initialW=initialW)
def __call__(self, x, img_size, scale=1.):
"""Forward Region Proposal Network.
Here are notations.
* :math:`N` is batch size.
* :math:`C` channel size of the input.
* :math:`H` and :math:`W` are height and witdh of the input feature.
* :math:`A` is number of anchors assigned to each pixel.
Args:
x (~chainer.Variable): The Features extracted from images.
Its shape is :math:`(N, C, H, W)`.
img_size (tuple of ints): A tuple :obj:`height, width`,
which contains image size after scaling.
scale (float): The amount of scaling done to the input images after
reading them from files.
Returns:
(~chainer.Variable, ~chainer.Variable, array, array, array):
This is a tuple of five following values.
* **rpn_locs**: Predicted bounding box offsets and scales for \
anchors. Its shape is :math:`(N, H W A, 4)`.
* **rpn_scores**: Predicted foreground scores for \
anchors. Its shape is :math:`(N, H W A, 2)`.
* **rois**: A bounding box array containing coordinates of \
proposal boxes. This is a concatenation of bounding box \
arrays from multiple images in the batch. \
Its shape is :math:`(R', 4)`. Given :math:`R_i` predicted \
bounding boxes from the :math:`i` th image, \
:math:`R' = \\sum _{i=1} ^ N R_i`.
* **roi_indices**: An array containing indices of images to \
which RoIs correspond to. Its shape is :math:`(R',)`.
* **anchor**: Coordinates of enumerated shifted anchors. \
Its shape is :math:`(H W A, 4)`.
"""
n, _, hh, ww = x.shape
anchor = _enumerate_shifted_anchor(
self.xp.array(self.anchor_base), self.feat_stride, hh, ww)
n_anchor = anchor.shape[0] // (hh * ww)
h = F.relu(self.conv1(x))
rpn_locs = self.loc(h)
rpn_scores = self.score(h)
rpn_locs = rpn_locs.transpose((0, 2, 3, 1)).reshape((n, -1, 4))
rpn_scores = rpn_scores.transpose((0, 2, 3, 1))
rpn_fg_scores =\
rpn_scores.reshape((n, hh, ww, n_anchor))[:, :, :, :] # modified from chainercv
rpn_fg_scores = rpn_fg_scores.reshape((n, -1))
rpn_scores = rpn_scores.reshape((n, -1)) # modified from chainercv
rois = []
roi_indices = []
for i in range(n):
roi = self.proposal_layer(
rpn_locs[i].array, rpn_fg_scores[i].array, anchor, img_size,
scale=scale)
batch_index = i * self.xp.ones((len(roi),), dtype=np.int32)
rois.append(roi)
roi_indices.append(batch_index)
rois = self.xp.concatenate(rois, axis=0)
roi_indices = self.xp.concatenate(roi_indices, axis=0)
return rpn_locs, rpn_scores, rois, roi_indices, anchor
def _enumerate_shifted_anchor(anchor_base, feat_stride, height, width):
# Enumerate all shifted anchors:
#
# add A anchors (1, A, 4) to
# cell K shifts (K, 1, 4) to get
# shift anchors (K, A, 4)
# reshape to (K*A, 4) shifted anchors
xp = cuda.get_array_module(anchor_base)
shift_y = xp.arange(0, height * feat_stride, feat_stride)
shift_x = xp.arange(0, width * feat_stride, feat_stride)
shift_x, shift_y = xp.meshgrid(shift_x, shift_y)
shift = xp.stack((shift_y.ravel(), shift_x.ravel(),
shift_y.ravel(), shift_x.ravel()), axis=1)
A = anchor_base.shape[0]
K = shift.shape[0]
anchor = anchor_base.reshape((1, A, 4)) + \
shift.reshape((1, K, 4)).transpose((1, 0, 2))
anchor = anchor.reshape((K * A, 4)).astype(np.float32)
return anchor
================================================
FILE: utils/roi_align_2d.py
================================================
# Modified work as ROIAlign:
# -----------------------------------------------------------------------------
# Copyright (c) 2018 DeNA
# -----------------------------------------------------------------------------
# Modified work:
# -----------------------------------------------------------------------------
# Copyright (c) 2015 Preferred Infrastructure, Inc.
# Copyright (c) 2015 Preferred Networks, Inc.
# -----------------------------------------------------------------------------
# Original work of forward_gpu and backward_gpu:
# -----------------------------------------------------------------------------
# Fast R-CNN
# Copyright (c) 2015 Microsoft
# Licensed under The MIT License [see fast-rcnn/LICENSE for details]
# Written by Ross Girshick
# -----------------------------------------------------------------------------
import numpy
import six
from chainer import cuda
from chainer import function
from chainer.utils import type_check
class ROIAlign2D(function.Function):
"""RoI align over a set of 2d planes."""
def __init__(self, outh, outw, spatial_scale):
self.outh, self.outw = outh, outw
self.spatial_scale = spatial_scale
def check_type_forward(self, in_types):
type_check.expect(in_types.size() == 2)
x_type, roi_type = in_types
type_check.expect(
x_type.dtype == numpy.float32,
x_type.ndim == 4,
roi_type.dtype == numpy.float32,
roi_type.ndim == 2,
roi_type.shape[1] == 5,
)
def forward_gpu(self, inputs):
self.retain_inputs((1,))
self._bottom_data_shape = inputs[0].shape
bottom_data, bottom_rois = inputs
#e.g. (batch, channel, h, w)=(1, 512, 38, 53) (n_rois, )=(128, 5)
channels, height, width = bottom_data.shape[1:]
n_rois = bottom_rois.shape[0]
top_data = cuda.cupy.empty((n_rois, channels, self.outh,
self.outw), dtype=numpy.float32)
cuda.cupy.ElementwiseKernel(
'''
raw float32 bottom_data, float32 spatial_scale, int32 channels,
int32 height, int32 width, int32 pooled_height, int32 pooled_width,
raw float32 bottom_rois
''',
'float32 top_data',
'''
// pos in output filter
int pw = i % pooled_width;
int ph = (i / pooled_width) % pooled_height;
int c = (i / pooled_width / pooled_height) % channels;
int num = i / pooled_width / pooled_height / channels;
// scale the ROI coordinates (1/16)
float roi_batch_ind = bottom_rois[num * 5 + 0];
float roi_start_w = bottom_rois[num * 5 + 1] * spatial_scale;
float roi_start_h = bottom_rois[num * 5 + 2] * spatial_scale;
float roi_end_w = bottom_rois[num * 5 + 3] * spatial_scale;
float roi_end_h = bottom_rois[num * 5 + 4] * spatial_scale;
// Force malformed ROIs to be 1x1
float roi_width = max(roi_end_w - roi_start_w, 1.0);
float roi_height = max(roi_end_h - roi_start_h, 1.0);
// float bin size
float bin_size_h = roi_height / static_cast<float>(pooled_height);
float bin_size_w = roi_width / static_cast<float>(pooled_width);
float maxval = 0;
int maxidx = -1;
for (int j = 0; j < 4; j++) {
int ih = j / 2;
int iw = j % 2;
float val = 0;
// ROIAlign using the center of the bin
float fh = roi_start_h + (static_cast<float>(ph) + 0.25 + static_cast<float>(ih) * 0.5f) * bin_size_h;
float fw = roi_start_w + (static_cast<float>(pw) + 0.25 + static_cast<float>(iw) * 0.5f) * bin_size_w;
if (fh < -1.0 || fh > height || fw < -1.0 || fw > width) {
continue;
}
int hstart = static_cast<int>(floor(fh));
int wstart = static_cast<int>(floor(fw));
int hend = hstart + 1;
int wend = wstart + 1;
if (hstart >= height - 1) {
hend = hstart = height - 1;
fh = static_cast<float>(hstart);
} else {
hend = hstart + 1;
}
if (wstart >= width - 1) {
wend = wstart = width - 1;
fw = static_cast<float>(wstart);
} else {
wend = wstart + 1;
}
float dh = fh - static_cast<float>(hstart);
float dw = fw - static_cast<float>(wstart);
//compute the max value in the bin
int data_offset = (roi_batch_ind * channels + c) * height * width;
val += (1.0 - dh) * (1.0 - dw) * bottom_data[data_offset + hstart * width + wstart];
val += (1.0 - dh) * dw * bottom_data[data_offset + hstart * width + wend];
val += dh * (1.0 - dw) * bottom_data[data_offset + hend * width + wstart];
val += dh * dw * bottom_data[data_offset + hend * width + wend];
maxval += val;
}
top_data = maxval / 4;
''', 'roi_pooling_2d_fwd'
)(bottom_data, self.spatial_scale, channels, height, width,
self.outh, self.outw, bottom_rois, top_data)
return top_data,
def backward_gpu(self, inputs, gy):
bottom_rois = inputs[1]
channels, height, width = self._bottom_data_shape[1:]
bottom_diff = cuda.cupy.zeros(self._bottom_data_shape, numpy.float32)
cuda.cupy.ElementwiseKernel(
'''
raw float32 top_diff, int32 num_rois,
float32 spatial_scale, int32 channels, int32 height, int32 width,
int32 pooled_height, int32 pooled_width, raw float32 bottom_rois
''',
'raw float32 bottom_diff',
'''
// pos in output filter
int pw = i % pooled_width;
int ph = (i / pooled_width) % pooled_height;
int c = (i / pooled_width / pooled_height) % channels;
int num = i / pooled_width / pooled_height / channels;
// scale the ROI coordinates (1/16)
float roi_batch_ind = bottom_rois[num * 5 + 0];
float roi_start_w = bottom_rois[num * 5 + 1] * spatial_scale;
float roi_start_h = bottom_rois[num * 5 + 2] * spatial_scale;
float roi_end_w = bottom_rois[num * 5 + 3] * spatial_scale;
float roi_end_h = bottom_rois[num * 5 + 4] * spatial_scale;
// Force malformed ROIs to be 1x1
float roi_width = max(roi_end_w - roi_start_w, 1.0);
float roi_height = max(roi_end_h - roi_start_h, 1.0);
// float bin size
float bin_size_h = roi_height / static_cast<float>(pooled_height);
float bin_size_w = roi_width / static_cast<float>(pooled_width);
int data_offset = (roi_batch_ind * channels + c) * height * width;
for (int j = 0; j < 4; j++) {
int ih = j / 2;
int iw = j % 2;
// ROIAlign using the center of the bin
float fh = roi_start_h + (static_cast<float>(ph) + 0.25 + static_cast<float>(ih) * 0.5f) * bin_size_h;
float fw = roi_start_w + (static_cast<float>(pw) + 0.25 + static_cast<float>(iw) * 0.5f) * bin_size_w;
if (fh < -1.0 || fh > height || fw < -1.0 || fw > width) {
continue;
}
int hstart = static_cast<int>(floor(fh));
int wstart = static_cast<int>(floor(fw));
int hend = hstart + 1;
int wend = wstart + 1;
if (hstart >= height - 1) {
hend = hstart = height - 1;
fh = static_cast<float>(hstart);
} else {
hend = hstart + 1;
}
if (wstart >= width - 1) {
wend = wstart = width - 1;
fw = static_cast<float>(wstart);
} else {
wend = wstart + 1;
}
float dh = fh - static_cast<float>(hstart);
float dw = fw - static_cast<float>(wstart);
//atomic add: pointer, value
atomicAdd(&bottom_diff[data_offset + hstart * width + wstart], top_diff[i] * (1.0 - dh) * (1.0 - dw) / 4);
atomicAdd(&bottom_diff[data_offset + hstart * width + wend], top_diff[i] * (1.0 - dh) * dw / 4);
atomicAdd(&bottom_diff[data_offset + hend * width + wstart], top_diff[i] * dh * (1.0 - dw) / 4);
atomicAdd(&bottom_diff[data_offset + hend * width + wend], top_diff[i] * dh * dw / 4);
}
''', 'roi_pooling_2d_bwd'
)(gy[0], bottom_rois.shape[0], self.spatial_scale,
channels, height, width, self.outh, self.outw,
bottom_rois, bottom_diff, size=gy[0].size)
return bottom_diff, None
def roi_align_2d(x, rois, outh, outw, spatial_scale):
"""Spatial Region of Interest (ROI) align function.
This function acts similarly to :class:`~functions.MaxPooling2D`, but
it computes the maximum of input spatial patch for each channel
with the region of interest.
Args:
x (~chainer.Variable): Input variable. The shape is expected to be
4 dimentional: (n: batch, c: channel, h, height, w: width).
rois (~chainer.Variable): Input roi variable. The shape is expected to
be (n: data size, 5), and each datum is set as below:
(batch_index, x_min, y_min, x_max, y_max).
outh (int): Height of output image after pooled.
outw (int): Width of output image after pooled.
spatial_scale (float): Scale of the roi is resized.
Returns:
~chainer.Variable: Output variable.
See the original paper proposing ROIPooling:
`Fast R-CNN <https://arxiv.org/abs/1504.08083>`_.
"""
return ROIAlign2D(outh, outw, spatial_scale)(x, rois)
================================================
FILE: utils/updater.py
================================================
import copy
import six
from chainer.dataset import convert
from chainer.dataset import iterator as iterator_module
from chainer import function, variable
from chainer.training.updater import StandardUpdater
from chainer import reporter
from chainer import cuda
class SubDivisionUpdater(StandardUpdater):
def __init__(self, iterator, optimizer, converter=convert.concat_examples,
subdivisions=1, device=None, loss_func=None):
super(SubDivisionUpdater, self).__init__(
iterator=iterator,
optimizer=optimizer,
converter=converter,
device=device,
loss_func=loss_func,
)
self._batchsize = self._iterators['main'].batch_size
self._subdivisions = subdivisions
self._n = int(self._batchsize / self._subdivisions)
assert self._batchsize % self._subdivisions == 0, (self._batchsize, self._subdivisions)
def update_core(self):
batch = self._iterators['main'].next()
#print(self._n)
in_arrays_list = []
for i in range(self._subdivisions):
in_arrays_list.append(self.converter(batch[i::self._subdivisions], self.device))
#in_arrays_list.append(self.converter(batch, self.device))
optimizer = self._optimizers['main']
loss_func = self.loss_func or optimizer.target
loss_func.cleargrads()
losses=[]
for i, in_arrays in enumerate(in_arrays_list):
if isinstance(in_arrays, tuple):
in_vars = list(variable.Variable(x) for x in in_arrays)
loss = loss_func(*in_vars)
elif isinstance(in_arrays, dict):
in_vars = {key: variable.Variable(x) for key, x in six.iteritems(in_arrays)}
loss = loss_func(in_vars)
else:
print(type(in_arrays))
loss.backward()
#loss = {k: cuda.to_cpu(v.data) for k, v in loss.items()} # for logging
loss = cuda.to_cpu(loss.data)
losses.append(loss)
optimizer.update()
# minibatch average
if isinstance(loss, dict):
avg_loss = {k: 0. for k in losses[0].keys()}
for loss in losses:
for k, v in loss.items():
avg_loss[k] += v
#avg_loss = {k: v / float(self._batchsize) for k, v in avg_loss.items()}
avg_loss = {k: v / float(len(losses)) for k, v in avg_loss.items()}
#avg_loss = {k: v for k, v in avg_loss.items()}
# report all the loss values
for k, v in avg_loss.items():
reporter.report({k: v}, loss_func)
reporter.report({'loss': sum(list(avg_loss.values()))}, loss_func)
else:
avg_loss = 0.
for loss in losses:
avg_loss += loss
avg_loss /= float(self._subdivisions)
reporter.report({'loss': avg_loss}, loss_func)
================================================
FILE: utils/vis_bbox.py
================================================
from chainercv.visualizations.vis_image import vis_image
import numpy as np
from skimage.measure import find_contours
from matplotlib.patches import Polygon
import cv2
def vis_bbox(img, bbox, label=None, score=None, mask=None, label_names=None, ax=None, contour=False, labeldisplay=True):
"""Visualize bounding boxes inside image.
Example:
>>> from chainercv.datasets import VOCDetectionDataset
>>> from chainercv.datasets import voc_bbox_label_names
>>> from chainercv.visualizations import vis_bbox
>>> import matplotlib.pyplot as plot
>>> dataset = VOCDetectionDataset()
>>> img, bbox, label = dataset[60]
>>> vis_bbox(img, bbox, label,
... label_names=voc_bbox_label_names)
>>> plot.show()
Args:
img (~numpy.ndarray): An array of shape :math:`(3, height, width)`.
This is in RGB format and the range of its value is
:math:`[0, 255]`.
bbox (~numpy.ndarray): An array of shape :math:`(R, 4)`, where
:math:`R` is the number of bounding boxes in the image.
Each element is organized
by :obj:`(y_min, x_min, y_max, x_max)` in the second axis.
label (~numpy.ndarray): An integer array of shape :math:`(R,)`.
The values correspond to id for label names stored in
:obj:`label_names`. This is optional.
score (~numpy.ndarray): A float array of shape :math:`(R,)`.
Each value indicates how confident the prediction is.
This is optional.
label_names (iterable of strings): Name of labels ordered according
to label ids. If this is :obj:`None`, labels will be skipped.
ax (matplotlib.axes.Axis): The visualization is displayed on this
axis. If this is :obj:`None` (default), a new axis is created.
Returns:
~matploblib.axes.Axes:
Returns the Axes object with the plot for further tweaking.
"""
from matplotlib import pyplot as plot
if label is not None and not len(bbox) == len(label):
raise ValueError('The length of label must be same as that of bbox')
if score is not None and not len(bbox) == len(score):
raise ValueError('The length of score must be same as that of bbox')
# alpha-blend the masks
COLOR=[(1,1,0), (1,0,1),(0,1,1),(0,0,1),(0,1,0), (1,0,0),(0.1,1,0.2)]
dst = img.astype(float)
for i, m in enumerate(mask):
alpha = np.tile(np.round(m), (3, 1, 1)).astype(float) * 0.4
src1 = np.ones(dst.shape).astype(float)
for j, col in enumerate(COLOR[i%len(COLOR)]):
src1[j] *= col * 255
dst = cv2.multiply(src1, alpha) + cv2.multiply(dst, 1 - alpha)
# Returns newly instantiated matplotlib.axes.Axes object if ax is None
ax = vis_image(dst, ax=ax)
# If there is no bounding box to display, visualize the image and exit.
if len(bbox) == 0:
return ax
# add boxes, contours and labels
for i, bb in enumerate(bbox):
# boxes
xy = (bb[1], bb[0])
height = int(bb[2]) - int(bb[0])
width = int(bb[3]) - int(bb[1])
ax.add_patch(plot.Rectangle(
xy, width, height, fill=False, edgecolor='red', linewidth=1))
# contours
if contour:
Mcontours = find_contours(mask[i].T, 0.5)
for verts in Mcontours:
p = Polygon(verts, facecolor="none", edgecolor=[0.5,0.5,0.5])
ax.add_patch(p)
#labels
caption = list()
if label is not None and label_names is not None:
lb = label[i]
print(lb)
if not (0 <= lb < len(label_names)):
raise ValueError('No corresponding name is given')
caption.append(label_names[lb])
if score is not None:
sc = score[i]
caption.append('{:.2f}'.format(sc))
if len(caption) > 0 and labeldisplay:
ax.text(bb[1], bb[0],
': '.join(caption),
style='italic',
fontsize=8,
color='white'
)#'facecolor': 'white', 'alpha': 0.7, 'pad': 10})
return ax
gitextract_iecqmj1q/
├── .gitignore
├── LICENSE
├── README.md
├── README_JP.md
├── coco_dataset.py
├── demo.py
├── getcoco.sh
├── mask_rcnn.py
├── mask_rcnn_resnet.py
├── mask_rcnn_train_chain.py
├── mask_rcnn_train_chain_batch.py
├── train.py
└── utils/
├── __init__.py
├── bn_utils.py
├── box_utils.py
├── cocoapi_evaluator.py
├── detection_coco_evaluator.py
├── detectron_parser.py
├── eval_detection_coco.py
├── makecocolist.py
├── proposal_target_creator.py
├── region_proposal_network.py
├── roi_align_2d.py
├── updater.py
└── vis_bbox.py
SYMBOL INDEX (75 symbols across 18 files)
FILE: coco_dataset.py
class COCODataset (line 11) | class COCODataset(chainer.dataset.DatasetMixin):
method __init__ (line 12) | def __init__(self, data_dir='COCO/', json_file='instances_train2017.js...
method __len__ (line 22) | def __len__(self):
method ann2rle (line 25) | def ann2rle(self, ann, height, width):
method get_example (line 35) | def get_example(self, i):
FILE: demo.py
function main (line 7) | def main():
FILE: mask_rcnn.py
class MaskRCNN (line 15) | class MaskRCNN(chainer.Chain):
method __init__ (line 16) | def __init__(self, extractor, rpn, head, mean,
method n_class (line 40) | def n_class(self):
method __call__ (line 43) | def __call__(self, x, scale=1.):
method use_preset (line 52) | def use_preset(self, preset):
method prepare (line 64) | def prepare(self, img):
method _suppress (line 78) | def _suppress(self, raw_cls_bbox, raw_cls_roi, raw_prob):
method predict (line 103) | def predict(self, imgs):
FILE: mask_rcnn_resnet.py
class ExtractorResNet (line 17) | class ExtractorResNet(chainer.link.Chain):
method __init__ (line 18) | def __init__(self, pretrained_model='auto', n_layers=50, roi_size=14):
method __call__ (line 43) | def __call__(self, x):
class MaskRCNNResNet (line 55) | class MaskRCNNResNet(MaskRCNN):
method __init__ (line 57) | def __init__(self,
class MaskRCNNHead (line 106) | class MaskRCNNHead(chainer.Chain):
method __init__ (line 107) | def __init__(self, n_class, roi_size, spatial_scale,
method res5head (line 125) | def res5head(self, x, rois, roi_indices):
method maskhead (line 143) | def maskhead(self, hres5):
method boxhead (line 149) | def boxhead(self, hres5):
function _roi_pooling_2d_yx (line 156) | def _roi_pooling_2d_yx(x, indices_and_rois, outh, outw, spatial_scale):
function _roi_align_2d_yx (line 162) | def _roi_align_2d_yx(x, indices_and_rois, outh, outw, spatial_scale):
FILE: mask_rcnn_train_chain.py
class MaskRCNNTrainChain (line 12) | class MaskRCNNTrainChain(chainer.Chain):
method __init__ (line 13) | def __init__(self, mask_rcnn, rpn_sigma=3., roi_sigma=1., gamma=1,
method __call__ (line 28) | def __call__(self, imgs, bboxes, labels, scale, masks, i):
function _smooth_l1_loss (line 101) | def _smooth_l1_loss(x, t, in_weight, sigma):
function _fast_rcnn_loc_loss (line 110) | def _fast_rcnn_loc_loss(pred_loc, gt_loc, gt_label, sigma):
FILE: mask_rcnn_train_chain_batch.py
class MaskRCNNTrainChain (line 12) | class MaskRCNNTrainChain(chainer.Chain):
method __init__ (line 13) | def __init__(self, mask_rcnn, rpn_sigma=3., roi_sigma=1., gamma=1,
method __call__ (line 28) | def __call__(self, imgs, bboxes, labels, scale, masks):
function _smooth_l1_loss (line 104) | def _smooth_l1_loss(x, t, in_weight, sigma):
function _fast_rcnn_loc_loss (line 113) | def _fast_rcnn_loc_loss(pred_loc, gt_loc, gt_label, sigma):
FILE: train.py
function resize_bbox (line 25) | def resize_bbox(bbox, in_size, out_size):
function parse (line 35) | def parse():
class Transform (line 62) | class Transform(object):
method __init__ (line 63) | def __init__(self, net, labelids):
method __call__ (line 66) | def __call__(self, in_data):
function convert (line 90) | def convert(batch, device):
function main (line 93) | def main():
FILE: utils/bn_utils.py
function freeze_bn (line 4) | def freeze_bn(model):
function bn_to_affine (line 21) | def bn_to_affine(model):
FILE: utils/box_utils.py
function resize_bbox (line 5) | def resize_bbox(bbox, in_size, out_size):
function bbox_yxyx2xywh (line 15) | def bbox_yxyx2xywh(bbox):
function im_mask (line 23) | def im_mask(mask, size, bbox):
FILE: utils/cocoapi_evaluator.py
class COCOAPIEvaluator (line 13) | class COCOAPIEvaluator(chainer.training.extensions.Evaluator):
method __init__ (line 18) | def __init__(
method evaluate (line 25) | def evaluate(self):
FILE: utils/detection_coco_evaluator.py
class DetectionCOCOEvaluator (line 11) | class DetectionCOCOEvaluator(chainer.training.extensions.Evaluator):
method __init__ (line 58) | def __init__(
method evaluate (line 73) | def evaluate(self):
FILE: utils/detectron_parser.py
function xytrans (line 92) | def xytrans(src):
FILE: utils/eval_detection_coco.py
function eval_detection_coco (line 11) | def eval_detection_coco(
function calc_detection_coco_prec_rec (line 89) | def calc_detection_coco_prec_rec(
function calc_detection_coco_ap (line 250) | def calc_detection_coco_ap(prec, rec, use_07_metric=False):
FILE: utils/proposal_target_creator.py
class ProposalTargetCreator (line 10) | class ProposalTargetCreator(object):
method __init__ (line 34) | def __init__(self,
method __call__ (line 47) | def __call__(self, roi, bbox, label, mask,
FILE: utils/region_proposal_network.py
class RegionProposalNetwork (line 13) | class RegionProposalNetwork(chainer.Chain):
method __init__ (line 48) | def __init__(
method __call__ (line 69) | def __call__(self, x, img_size, scale=1.):
function _enumerate_shifted_anchor (line 138) | def _enumerate_shifted_anchor(anchor_base, feat_stride, height, width):
FILE: utils/roi_align_2d.py
class ROIAlign2D (line 27) | class ROIAlign2D(function.Function):
method __init__ (line 31) | def __init__(self, outh, outw, spatial_scale):
method check_type_forward (line 35) | def check_type_forward(self, in_types):
method forward_gpu (line 47) | def forward_gpu(self, inputs):
method backward_gpu (line 138) | def backward_gpu(self, inputs, gy):
function roi_align_2d (line 219) | def roi_align_2d(x, rois, outh, outw, spatial_scale):
FILE: utils/updater.py
class SubDivisionUpdater (line 11) | class SubDivisionUpdater(StandardUpdater):
method __init__ (line 14) | def __init__(self, iterator, optimizer, converter=convert.concat_examp...
method update_core (line 28) | def update_core(self):
FILE: utils/vis_bbox.py
function vis_bbox (line 7) | def vis_bbox(img, bbox, label=None, score=None, mask=None, label_names=N...
Condensed preview — 25 files, each showing path, character count, and a content snippet. Download the .json file or copy for the full structured content (122K chars).
[
{
"path": ".gitignore",
"chars": 23,
"preview": "__pycache*\nresult\n*.png"
},
{
"path": "LICENSE",
"chars": 4901,
"preview": "Copyright (c) 2018 DeNA Co., Ltd.\n\nPermission is hereby granted, free of charge, to any person obtaining a copy\n of this"
},
{
"path": "README.md",
"chars": 5410,
"preview": "# Chainer\\_Mask\\_R-CNN \nChainer implementation of Mask R-CNN - the multi-task network for object detection, object cla"
},
{
"path": "README_JP.md",
"chars": 3935,
"preview": "# Chainer\\_Mask\\_R-CNN \nマルチタスク検出器Mask R-CNNのchainer実装\n(https://arxiv.org/abs/1703.06870) \n\n## 実行例\n- 準備中\n\n## 必要環境\n- ["
},
{
"path": "coco_dataset.py",
"chars": 2678,
"preview": "import numpy as np\nfrom skimage.draw import polygon\nimport json\nimport os\nimport cv2\nimport pycocotools\nfrom pycocotools"
},
{
"path": "demo.py",
"chars": 4033,
"preview": "import argparse\r\nimport chainer\r\nimport numpy as np\r\nfrom mask_rcnn_train_chain import MaskRCNNTrainChain\r\nfrom utils.bn"
},
{
"path": "getcoco.sh",
"chars": 368,
"preview": "# get COCO dataset\nmkdir COCO\ncd COCO\n\nwget http://images.cocodataset.org/zips/train2017.zip\nwget http://images.cocodata"
},
{
"path": "mask_rcnn.py",
"chars": 7857,
"preview": "from __future__ import division\n\nimport numpy as np\n\nimport chainer\nfrom chainer import cuda\nimport chainer.functions as"
},
{
"path": "mask_rcnn_resnet.py",
"chars": 7233,
"preview": "import numpy as np\r\n\r\nimport chainer\r\nimport chainer.functions as F\r\nimport chainer.links as L\r\nfrom mask_rcnn import Ma"
},
{
"path": "mask_rcnn_train_chain.py",
"chars": 5339,
"preview": "import numpy as np\r\n\r\nimport chainer\r\nfrom chainer import cuda\r\nimport chainer.functions as F\r\n\r\nfrom chainercv.links.mo"
},
{
"path": "mask_rcnn_train_chain_batch.py",
"chars": 5519,
"preview": "import numpy as np\r\n\r\nimport chainer\r\nfrom chainer import cuda\r\nimport chainer.functions as F\r\n\r\nfrom chainercv.links.mo"
},
{
"path": "train.py",
"chars": 8408,
"preview": "import chainer\r\nfrom chainer import training\r\nfrom chainer.training import extensions, ParallelUpdater\r\nfrom chainer.tra"
},
{
"path": "utils/__init__.py",
"chars": 0,
"preview": ""
},
{
"path": "utils/bn_utils.py",
"chars": 1967,
"preview": "import numpy as np\nimport cupy\n\ndef freeze_bn(model):\n # freeze batchnorm update \n def disableupdate(block):\n "
},
{
"path": "utils/box_utils.py",
"chars": 1614,
"preview": "import numpy as np\nimport cupy\nimport cv2\n\ndef resize_bbox(bbox, in_size, out_size):\n bbox_o = bbox.copy()\n y_scal"
},
{
"path": "utils/cocoapi_evaluator.py",
"chars": 2930,
"preview": "import copy\nimport numpy as np\n\nfrom chainer import reporter\nimport chainer.training.extensions\n\nfrom utils import eval_"
},
{
"path": "utils/detection_coco_evaluator.py",
"chars": 5444,
"preview": "import copy\nimport numpy as np\n\nfrom chainer import reporter\nimport chainer.training.extensions\n\nfrom utils import eval_"
},
{
"path": "utils/detectron_parser.py",
"chars": 3835,
"preview": "import numpy as np\nimport os\npath = os.path.join(os.path.dirname(__file__), '../')\nimport sys\nsys.path.append(path)\nfrom"
},
{
"path": "utils/eval_detection_coco.py",
"chars": 12859,
"preview": "from __future__ import division\n\nfrom collections import defaultdict\nimport itertools\nimport numpy as np\nimport six\n\nfro"
},
{
"path": "utils/makecocolist.py",
"chars": 421,
"preview": "import glob\nfnames = glob.glob('COCO/train2017/*.jpg')\n\nwith open(\"COCO/train2017.txt\", \"w\") as f:\n for fname in fnam"
},
{
"path": "utils/proposal_target_creator.py",
"chars": 7034,
"preview": "import numpy as np\n\nfrom chainer import cuda\n\nfrom chainercv.links.model.faster_rcnn.utils.bbox2loc import bbox2loc\nfrom"
},
{
"path": "utils/region_proposal_network.py",
"chars": 6568,
"preview": "import numpy as np\n\nimport chainer\nfrom chainer import cuda\nimport chainer.functions as F\nimport chainer.links as L\n\nfro"
},
{
"path": "utils/roi_align_2d.py",
"chars": 10625,
"preview": "# Modified work as ROIAlign:\r\n# -----------------------------------------------------------------------------\r\n# Copyrig"
},
{
"path": "utils/updater.py",
"chars": 3035,
"preview": "import copy\r\nimport six\r\n\r\nfrom chainer.dataset import convert\r\nfrom chainer.dataset import iterator as iterator_module\r"
},
{
"path": "utils/vis_bbox.py",
"chars": 4354,
"preview": "from chainercv.visualizations.vis_image import vis_image\r\nimport numpy as np\r\nfrom skimage.measure import find_contours\r"
}
]
About this extraction
This page contains the full source code of the DeNA/Chainer_Mask_R-CNN GitHub repository, extracted and formatted as plain text for AI agents and large language models (LLMs). The extraction includes 25 files (113.7 KB), approximately 30.5k tokens, and a symbol index with 75 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.