Repository: arimousa/DDAD
Branch: main
Commit: 4323a150dd63
Files: 16
Total size: 77.1 KB
Directory structure:
gitextract_ujrg7voc/
├── LICENSE
├── README.md
├── anomaly_map.py
├── config.yaml
├── dataset.py
├── ddad.py
├── feature_extractor.py
├── loss.py
├── main.py
├── metrics.py
├── reconstruction.py
├── requirements.txt
├── resnet.py
├── train.py
├── unet.py
└── visualize.py
================================================
FILE CONTENTS
================================================
================================================
FILE: LICENSE
================================================
MIT License
Copyright (c) 2023 Arian Mousakhan
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, sublicense, and/or sell
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.
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.
================================================
FILE: README.md
================================================
# Anomaly Detection with Conditioned Denoising Diffusion Models.
Official implementation of [DDAD](https://arxiv.org/abs/2305.15956)
[](https://paperswithcode.com/sota/anomaly-detection-on-mvtec-ad?p=anomaly-detection-with-conditioned-denoising) [](https://paperswithcode.com/sota/anomaly-detection-on-visa?p=anomaly-detection-with-conditioned-denoising)
## Requirements
This repository is implemented and tested on Python 3.8 and PyTorch 2.1.
To install requirements:
```setup
pip install -r requirements.txt
```
## Train and Evaluation of the Model
You can download the model checkpoints directly from [Checkpoints](https://drive.google.com/drive/u/0/folders/1FF83llo3a-mN5pJN8-_mw0hL5eZqe9fC)
To train the denoising UNet, run:
```train
python main.py --train True
```
Modify the settings in the config.yaml file to train the model on different categories.
For fine-tuning the feature extractor, use the following command:
```domain_adaptation
python main.py --domain_adaptation True
```
To evaluate and test the model, run:
```detection
python main.py --detection True
```
## Dataset
You can download [MVTec AD: MVTec Software](https://www.mvtec.com/company/research/datasets/mvtec-ad/) and [VisA](https://amazon-visual-anomaly.s3.us-west-2.amazonaws.com/VisA_20220922.tar) Benchmarks.
For preprocessing of VisA dataset check out the [Data preparation](https://github.com/amazon-science/spot-diff/tree/main) section of this repository.
The dataset should be placed in the 'datasets' folder. The training dataset should only contain one subcategory consisting of nominal samples, which should be named 'good'. The test dataset should include one category named 'good' for nominal samples, and any other subcategories of anomalous samples. It should be made as follows:
```shell
Name_of_Dataset
|-- Category
|-----|----- ground_truth
|-----|----- test
|-----|--------|------ good
|-----|--------|------ ...
|-----|--------|------ ...
|-----|----- train
|-----|--------|------ good
```
## Results
Running the code as explained in this file should achieve the following results for MVTec AD:
Anomaly Detection (Image AUROC) and Anomaly Localization (Pixel AUROC, PRO)
Expected results for MVTec AD:
| Category | Carpet | Grid | Leather | Tile | Wood | Bottle | Cable | Capsule | Hazel nut | Metalnut | Pill | Screw | Toothbrush | Transistor | Zipper |Average
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Detection | 99.3% | 100% | 100% | 100% | 100% | 100% | 99.4% | 99.4% | 100% | 100% | 100% | 99.0% | 100% | 100% | 100% | 99.8%
| Localization | (98.7%,93.9%) | (99.4%,97.3%) | (99.4%,97.7%) | (98.2%,93.1%) | (95.0%,82.9%) | (98.7%,91.8%) | (98.1%,88.9%) | (95.7%,93.4%) | (98.4%,86.7%) | (99.0%,91.1%) | (99.1%,95.5%) | (99.3%,96.3%) | (98.7%,92.6%) | (95.3%,90.1%) | (98.2%,93.2%) | (98,1%,92.3%)
The settings used for these results are detailed in the table.
| **Categories** | Carpet | Grid | Leather | Tile | Wood | Bottle | Cable | Capsule | Hazelnut | Metal nut | Pill | Screw | Toothbrush | Transistor | Zipper |
| -------------- | ------ | ---- | ------- | ---- | ---- | ------ | ----- | ------- | -------- | --------- | ---- | ----- | ----------- | ---------- | ------ |
| **\(w\)** | 0 | 4 | 11 | 4 | 11 | 3 | 3 | 8 | 5 | 7 | 9 | 2 | 0 | 0 | 10 |
| **Training epochs** | 2500 | 2000 | 2000 | 1000 | 2000 | 1000 | 3000 | 1500 | 2000 | 3000 | 1000 | 2000 | 2000 | 2000 | 1000 |
| **FE epochs** | 0 | 6 | 8 | 0 | 16 | 5 | 0 | 8 | 3 | 1 | 4 | 4 | 2 | 0 | 6 |
Following is the expected results on VisA Dataset.
| Category | Candle | Capsules | Cashew | Chewing gum | Fryum | Macaroni1 | Macaroni2 | PCB1 | PCB2 | PCB3 | PCB4 | Pipe fryum | Average
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Detection | 99.9% | 100% | 94.5% | 98.1% | 99.0% | 99.2% | 99.2% | 100% | 99.7% | 97.2% | 100% | 100% | 98.9%
| Localization | (98.7%,96.6%) | (99.5%,95.0%) | (97.4%,80.3%) | (96.5%,85.2%) | (96.9%,94.2%) | (98.7%,98.5%) | (98.2%,99.3%) | (93.4%,93.3%) | (97.4%,93.3%) | (96.3%,86.6%) | (98.5%,95.5%) | (99.5%,94.7%) |(97.6%,92.7%)
The settings used for these results are detailed in the table.
| **Categories** | Candle | Capsules | Cashew | Chewing gum | Fryum | Macaroni1 | Macaroni2 | PCB1 | PCB2 | PCB3 | PCB4 | Pipe fryum |
| ---------------- | ------ | -------- | ------ | ------------ | ----- | --------- | --------- | ---- | ---- | ---- | ---- | ---------- |
| **\(w\)** | 6 | 5 | 0 | 6 | 4 | 5 | 2 | 9 | 5 | 6 | 6 | 8 |
| **Training epochs** | 1000 | 1000 | 1750 | 1250 | 1000 | 500 | 500 | 500 | 500 | 500 | 500 | 500 |
| **FE epochs** | 1 | 3 | 0 | 0 | 3 | 7 | 11 | 8 | 5 | 1 | 1 | 6 |
## Citation
```
@article{mousakhan2023anomaly,
title={Anomaly Detection with Conditioned Denoising Diffusion Models},
author={Mousakhan, Arian and Brox, Thomas and Tayyub, Jawad},
journal={arXiv preprint arXiv:2305.15956},
year={2023}
}
```
## Feedback
For any feedback or inquiries, please contact arian.mousakhan@gmail.com
================================================
FILE: anomaly_map.py
================================================
import torch
import torch.nn.functional as F
from kornia.filters import gaussian_blur2d
from torchvision.transforms import transforms
import math
from dataset import *
from visualize import *
from feature_extractor import *
import numpy as np
def heat_map(output, target, FE, config):
'''
Compute the anomaly map
:param output: the output of the reconstruction
:param target: the target image
:param FE: the feature extractor
:param sigma: the sigma of the gaussian kernel
:param i_d: the pixel distance
:param f_d: the feature distance
'''
sigma = 4
kernel_size = 2 * int(4 * sigma + 0.5) +1
anomaly_map = 0
output = output.to(config.model.device)
target = target.to(config.model.device)
i_d = pixel_distance(output, target)
f_d = feature_distance((output), (target), FE, config)
f_d = torch.Tensor(f_d).to(config.model.device)
anomaly_map += f_d + config.model.v * (torch.max(f_d)/ torch.max(i_d)) * i_d
anomaly_map = gaussian_blur2d(
anomaly_map , kernel_size=(kernel_size,kernel_size), sigma=(sigma,sigma)
)
anomaly_map = torch.sum(anomaly_map, dim=1).unsqueeze(1)
return anomaly_map
def pixel_distance(output, target):
'''
Pixel distance between image1 and image2
'''
distance_map = torch.mean(torch.abs(output - target), dim=1).unsqueeze(1)
return distance_map
def feature_distance(output, target, FE, config):
'''
Feature distance between output and target
'''
FE.eval()
transform = transforms.Compose([
transforms.Lambda(lambda t: (t + 1) / (2)),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
target = transform(target)
output = transform(output)
inputs_features = FE(target)
output_features = FE(output)
out_size = config.data.image_size
anomaly_map = torch.zeros([inputs_features[0].shape[0] ,1 ,out_size, out_size]).to(config.model.device)
for i in range(len(inputs_features)):
if i == 0:
continue
a_map = 1 - F.cosine_similarity(patchify(inputs_features[i]), patchify(output_features[i]))
a_map = torch.unsqueeze(a_map, dim=1)
a_map = F.interpolate(a_map, size=out_size, mode='bilinear', align_corners=True)
anomaly_map += a_map
return anomaly_map
#https://github.com/amazon-science/patchcore-inspection
def patchify(features, return_spatial_info=False):
"""Convert a tensor into a tensor of respective patches.
Args:
x: [torch.Tensor, bs x c x w x h]
Returns:
x: [torch.Tensor, bs * w//stride * h//stride, c, patchsize,
patchsize]
"""
patchsize = 3
stride = 1
padding = int((patchsize - 1) / 2)
unfolder = torch.nn.Unfold(
kernel_size=patchsize, stride=stride, padding=padding, dilation=1
)
unfolded_features = unfolder(features)
number_of_total_patches = []
for s in features.shape[-2:]:
n_patches = (
s + 2 * padding - 1 * (patchsize - 1) - 1
) / stride + 1
number_of_total_patches.append(int(n_patches))
unfolded_features = unfolded_features.reshape(
*features.shape[:2], patchsize, patchsize, -1
)
unfolded_features = unfolded_features.permute(0, 4, 1, 2, 3)
max_features = torch.mean(unfolded_features, dim=(3,4))
features = max_features.reshape(features.shape[0], int(math.sqrt(max_features.shape[1])) , int(math.sqrt(max_features.shape[1])), max_features.shape[-1]).permute(0,3,1,2)
if return_spatial_info:
return unfolded_features, number_of_total_patches
return features
================================================
FILE: config.yaml
================================================
data :
name: MVTec #MVTec #MTD #VisA
data_dir: datasets/MVTec #MVTec #VisA #MTD
category: screw #['carpet', 'bottle', 'hazelnut', 'leather', 'cable', 'capsule', 'grid', 'pill', 'transistor', 'metal_nut', 'screw','toothbrush', 'zipper', 'tile', 'wood']
# ['candle', 'capsules', 'cashew', 'chewinggum', 'fryum', 'macaroni1', 'macaroni2', 'pcb1', 'pcb2' ,'pcb3', 'pcb4', 'pipe_fryum']
image_size: 256
batch_size: 32 # 32 for DDAD and 16 for DDADS
DA_batch_size: 16 #16 for MVTec and [macaroni2, pcb1] in VisA, and 32 for other categories in VisA
test_batch_size: 16 #16 for MVTec, 32 for VisA
mask : True
input_channel : 3
model:
DDADS: False
checkpoint_dir: checkpoints/MVTec #MTD #MVTec #VisA
checkpoint_name: weights
exp_name: default
feature_extractor: wide_resnet101_2 #wide_resnet101_2 # wide_resnet50_2 #resnet50
learning_rate: 3e-4
weight_decay: 0.05
epochs: 3000
load_chp : 2000 # From this epoch checkpoint will be loaded. Every 250 epochs a checkpoint is saved. Try to load 750 or 1000 epochs for Visa and 1000-1500-2000 for MVTec.
DA_epochs: 4 # Number of epochs for Domain adaptation.
DA_chp: 4
v : 1 #7 # 1 for MVTec and cashew in VisA, and 7 for VisA (1.5 for cashew). Control parameter for pixel-wise and feature-wise comparison. v * D_p + D_f
w : 2 # Conditionig parameter. The higher the value, the more the model is conditioned on the target image. "Fine tuninig this parameter results in better performance".
w_DA : 3 #3 # Conditionig parameter for domain adaptation. The higher the value, the more the model is conditioned on the target image.
DLlambda : 0.1 # 0.1 for MVTec and 0.01 for VisA
trajectory_steps: 1000
test_trajectoy_steps: 250 # Starting point for denoining trajectory.
test_trajectoy_steps_DA: 250 # Starting point for denoining trajectory for domain adaptation.
skip : 25 # Number of steps to skip for denoising trajectory.
skip_DA : 25
eta : 1 # Stochasticity parameter for denoising process.
beta_start : 0.0001
beta_end : 0.02
device: 'cuda' #<"cpu", "gpu", "tpu", "ipu">
save_model: True
num_workers : 2
seed : 42
metrics:
auroc: True
pro: True
misclassifications: False
visualisation: False
================================================
FILE: dataset.py
================================================
import os
from glob import glob
from pathlib import Path
import shutil
import numpy as np
import csv
import torch
import torch.utils.data
from PIL import Image
from torchvision import transforms
import torch.nn.functional as F
import torchvision.datasets as datasets
from torchvision.datasets import CIFAR10
class Dataset_maker(torch.utils.data.Dataset):
def __init__(self, root, category, config, is_train=True):
self.image_transform = transforms.Compose(
[
transforms.Resize((config.data.image_size, config.data.image_size)),
transforms.ToTensor(), # Scales data into [0,1]
transforms.Lambda(lambda t: (t * 2) - 1) # Scale between [-1, 1]
]
)
self.config = config
self.mask_transform = transforms.Compose(
[
transforms.Resize((config.data.image_size, config.data.image_size)),
transforms.ToTensor(), # Scales data into [0,1]
]
)
if is_train:
if category:
self.image_files = glob(
os.path.join(root, category, "train", "good", "*.png")
)
else:
self.image_files = glob(
os.path.join(root, "train", "good", "*.png")
)
else:
if category:
self.image_files = glob(os.path.join(root, category, "test", "*", "*.png"))
else:
self.image_files = glob(os.path.join(root, "test", "*", "*.png"))
self.is_train = is_train
def __getitem__(self, index):
image_file = self.image_files[index]
image = Image.open(image_file)
image = self.image_transform(image)
if(image.shape[0] == 1):
image = image.expand(3, self.config.data.image_size, self.config.data.image_size)
if self.is_train:
label = 'good'
return image, label
else:
if self.config.data.mask:
if os.path.dirname(image_file).endswith("good"):
target = torch.zeros([1, image.shape[-2], image.shape[-1]])
label = 'good'
else :
if self.config.data.name == 'MVTec':
target = Image.open(
image_file.replace("/test/", "/ground_truth/").replace(
".png", "_mask.png"
)
)
else:
target = Image.open(
image_file.replace("/test/", "/ground_truth/"))
target = self.mask_transform(target)
label = 'defective'
else:
if os.path.dirname(image_file).endswith("good"):
target = torch.zeros([1, image.shape[-2], image.shape[-1]])
label = 'good'
else :
target = torch.zeros([1, image.shape[-2], image.shape[-1]])
label = 'defective'
return image, target, label
def __len__(self):
return len(self.image_files)
================================================
FILE: ddad.py
================================================
from asyncio import constants
from typing import Any
import torch
from unet import *
from dataset import *
from visualize import *
from anomaly_map import *
from metrics import *
from feature_extractor import *
from reconstruction import *
os.environ['CUDA_VISIBLE_DEVICES'] = "0,1,2"
class DDAD:
def __init__(self, unet, config) -> None:
self.test_dataset = Dataset_maker(
root= config.data.data_dir,
category=config.data.category,
config = config,
is_train=False,
)
self.testloader = torch.utils.data.DataLoader(
self.test_dataset,
batch_size= config.data.test_batch_size,
shuffle=False,
num_workers= config.model.num_workers,
drop_last=False,
)
self.unet = unet
self.config = config
self.reconstruction = Reconstruction(self.unet, self.config)
self.transform = transforms.Compose([
transforms.CenterCrop((224)),
])
def __call__(self) -> Any:
feature_extractor = domain_adaptation(self.unet, self.config, fine_tune=False)
feature_extractor.eval()
labels_list = []
predictions= []
anomaly_map_list = []
gt_list = []
reconstructed_list = []
forward_list = []
with torch.no_grad():
for input, gt, labels in self.testloader:
input = input.to(self.config.model.device)
x0 = self.reconstruction(input, input, self.config.model.w)[-1]
anomaly_map = heat_map(x0, input, feature_extractor, self.config)
anomaly_map = self.transform(anomaly_map)
gt = self.transform(gt)
forward_list.append(input)
anomaly_map_list.append(anomaly_map)
gt_list.append(gt)
reconstructed_list.append(x0)
for pred, label in zip(anomaly_map, labels):
labels_list.append(0 if label == 'good' else 1)
predictions.append(torch.max(pred).item())
metric = Metric(labels_list, predictions, anomaly_map_list, gt_list, self.config)
metric.optimal_threshold()
if self.config.metrics.auroc:
print('AUROC: ({:.1f},{:.1f})'.format(metric.image_auroc() * 100, metric.pixel_auroc() * 100))
if self.config.metrics.pro:
print('PRO: {:.1f}'.format(metric.pixel_pro() * 100))
if self.config.metrics.misclassifications:
metric.miscalssified()
reconstructed_list = torch.cat(reconstructed_list, dim=0)
forward_list = torch.cat(forward_list, dim=0)
anomaly_map_list = torch.cat(anomaly_map_list, dim=0)
pred_mask = (anomaly_map_list > metric.threshold).float()
gt_list = torch.cat(gt_list, dim=0)
if not os.path.exists('results'):
os.mkdir('results')
if self.config.metrics.visualisation:
visualize(forward_list, reconstructed_list, gt_list, pred_mask, anomaly_map_list, self.config.data.category)
================================================
FILE: feature_extractor.py
================================================
import logging
import torch
from dataset import *
from dataset import *
from unet import *
from visualize import *
from resnet import *
import torchvision.transforms as T
from reconstruction import *
os.environ['CUDA_VISIBLE_DEVICES'] = "0,1,2"
def loss_fucntion(a, b, c, d, config):
cos_loss = torch.nn.CosineSimilarity()
loss1 = 0
loss2 = 0
loss3 = 0
for item in range(len(a)):
loss1 += torch.mean(1-cos_loss(a[item].view(a[item].shape[0],-1),b[item].view(b[item].shape[0],-1)))
loss2 += torch.mean(1-cos_loss(b[item].view(b[item].shape[0],-1),c[item].view(c[item].shape[0],-1))) * config.model.DLlambda
loss3 += torch.mean(1-cos_loss(a[item].view(a[item].shape[0],-1),d[item].view(d[item].shape[0],-1))) * config.model.DLlambda
loss = loss1+loss2+loss3
return loss
def domain_adaptation(unet, config, fine_tune):
if config.model.feature_extractor == 'wide_resnet101_2':
feature_extractor = wide_resnet101_2(pretrained=True)
frozen_feature_extractor = wide_resnet101_2(pretrained=True)
elif config.model.feature_extractor == 'wide_resnet50_2':
feature_extractor = wide_resnet50_2(pretrained=True)
frozen_feature_extractor = wide_resnet50_2(pretrained=True)
elif config.model.feature_extractor == 'resnet50':
feature_extractor = resnet50(pretrained=True)
frozen_feature_extractor = resnet50(pretrained=True)
else:
logging.warning("Feature extractor is not correctly selected, Default: wide_resnet101_2")
feature_extractor = wide_resnet101_2(pretrained=True)
frozen_feature_extractor = wide_resnet101_2(pretrained=True)
feature_extractor.to(config.model.device)
frozen_feature_extractor.to(config.model.device)
frozen_feature_extractor.eval()
feature_extractor = torch.nn.DataParallel(feature_extractor)
frozen_feature_extractor = torch.nn.DataParallel(frozen_feature_extractor)
train_dataset = Dataset_maker(
root= config.data.data_dir,
category= config.data.category,
config = config,
is_train=True,
)
trainloader = torch.utils.data.DataLoader(
train_dataset,
batch_size=config.data.DA_batch_size,
shuffle=True,
num_workers=config.model.num_workers,
drop_last=True,
)
if fine_tune:
unet.eval()
feature_extractor.train()
transform = transforms.Compose([
transforms.Lambda(lambda t: (t + 1) / (2)),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
optimizer = torch.optim.AdamW(feature_extractor.parameters(),lr= 1e-4)
torch.save(frozen_feature_extractor.state_dict(), os.path.join(os.path.join(os.getcwd(), config.model.checkpoint_dir), config.data.category,f'feat0'))
reconstruction = Reconstruction(unet, config)
for epoch in range(config.model.DA_epochs):
for step, batch in enumerate(trainloader):
half_batch_size = batch[0].shape[0]//2
target = batch[0][:half_batch_size].to(config.model.device)
input = batch[0][half_batch_size:].to(config.model.device)
x0 = reconstruction(input, target, config.model.w_DA)[-1].to(config.model.device)
x0 = transform(x0)
target = transform(target)
reconst_fe = feature_extractor(x0)
target_fe = feature_extractor(target)
target_frozen_fe = frozen_feature_extractor(target)
reconst_frozen_fe = frozen_feature_extractor(x0)
loss = loss_fucntion(reconst_fe, target_fe, target_frozen_fe,reconst_frozen_fe, config)
optimizer.zero_grad()
loss.backward()
optimizer.step()
print(f"Epoch {epoch+1} | Loss: {loss.item()}")
# if (epoch+1) % 5 == 0:
torch.save(feature_extractor.state_dict(), os.path.join(os.path.join(os.getcwd(), config.model.checkpoint_dir), config.data.category,f'feat{epoch+1}'))
else:
checkpoint = torch.load(os.path.join(os.path.join(os.getcwd(), config.model.checkpoint_dir), config.data.category,f'feat{config.model.DA_chp}'))#{config.model.DA_chp}
feature_extractor.load_state_dict(checkpoint)
return feature_extractor
================================================
FILE: loss.py
================================================
import torch
import torch.nn as nn
import numpy as np
def get_loss(model, x_0, t, config):
x_0 = x_0.to(config.model.device)
betas = np.linspace(config.model.beta_start, config.model.beta_end, config.model.trajectory_steps, dtype=np.float64)
b = torch.tensor(betas).type(torch.float).to(config.model.device)
e = torch.randn_like(x_0, device = x_0.device)
at = (1-b).cumprod(dim=0).index_select(0, t).view(-1, 1, 1, 1)
x = at.sqrt() * x_0 + (1- at).sqrt() * e
output = model(x, t.float())
return (e - output).square().sum(dim=(1, 2, 3)).mean(dim=0)
================================================
FILE: main.py
================================================
import torch
import numpy as np
import os
import argparse
from unet import *
from omegaconf import OmegaConf
from train import trainer
from feature_extractor import *
from ddad import *
os.environ['CUDA_VISIBLE_DEVICES'] = "0,1,2"
def build_model(config):
if config.model.DDADS:
unet = UNetModel(config.data.image_size, 32, dropout=0.3, n_heads=2 ,in_channels=config.data.input_channel)
else:
unet = UNetModel(config.data.image_size, 64, dropout=0.0, n_heads=4 ,in_channels=config.data.input_channel)
return unet
def train(config):
torch.manual_seed(42)
np.random.seed(42)
unet = build_model(config)
print(" Num params: ", sum(p.numel() for p in unet.parameters()))
unet = unet.to(config.model.device)
unet.train()
unet = torch.nn.DataParallel(unet)
# checkpoint = torch.load(os.path.join(os.path.join(os.getcwd(), config.model.checkpoint_dir), config.data.category,'1000'))
# unet.load_state_dict(checkpoint)
trainer(unet, config.data.category, config)#config.data.category,
def detection(config):
unet = build_model(config)
checkpoint = torch.load(os.path.join(os.getcwd(), config.model.checkpoint_dir, config.data.category, str(config.model.load_chp)))
unet = torch.nn.DataParallel(unet)
unet.load_state_dict(checkpoint)
unet.to(config.model.device)
checkpoint = torch.load(os.path.join(os.getcwd(), config.model.checkpoint_dir, config.data.category, str(config.model.load_chp)))
unet.eval()
ddad = DDAD(unet, config)
ddad()
def finetuning(config):
unet = build_model(config)
checkpoint = torch.load(os.path.join(os.getcwd(), config.model.checkpoint_dir, config.data.category, str(config.model.load_chp)))
unet = torch.nn.DataParallel(unet)
unet.load_state_dict(checkpoint)
unet.to(config.model.device)
unet.eval()
domain_adaptation(unet, config, fine_tune=True)
def parse_args():
cmdline_parser = argparse.ArgumentParser('DDAD')
cmdline_parser.add_argument('-cfg', '--config',
default= os.path.join(os.path.dirname(os.path.abspath(__file__)),'config.yaml'),
help='config file')
cmdline_parser.add_argument('--train',
default= False,
help='Train the diffusion model')
cmdline_parser.add_argument('--detection',
default= False,
help='Detection anomalies')
cmdline_parser.add_argument('--domain_adaptation',
default= False,
help='Domain adaptation')
args, unknowns = cmdline_parser.parse_known_args()
return args
if __name__ == "__main__":
torch.cuda.empty_cache()
args = parse_args()
config = OmegaConf.load(args.config)
print("Class: ",config.data.category, " w:", config.model.w, " v:", config.model.v, " load_chp:", config.model.load_chp, " feature extractor:", config.model.feature_extractor," w_DA: ",config.model.w_DA," DLlambda: ",config.model.DLlambda)
print(f'{config.model.test_trajectoy_steps=} , {config.data.test_batch_size=}')
torch.manual_seed(42)
np.random.seed(42)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(42)
if args.train:
print('Training...')
train(config)
if args.domain_adaptation:
print('Domain Adaptation...')
finetuning(config)
if args.detection:
print('Detecting Anomalies...')
detection(config)
================================================
FILE: metrics.py
================================================
import torch
from torchmetrics import ROC, AUROC, F1Score
import os
from torchvision.transforms import transforms
from skimage import measure
import pandas as pd
from statistics import mean
import numpy as np
from sklearn.metrics import auc
from sklearn import metrics
from sklearn.metrics import roc_auc_score, roc_curve
class Metric:
def __init__(self,labels_list, predictions, anomaly_map_list, gt_list, config) -> None:
self.labels_list = labels_list
self.predictions = predictions
self.anomaly_map_list = anomaly_map_list
self.gt_list = gt_list
self.config = config
self.threshold = 0.5
def image_auroc(self):
auroc_image = roc_auc_score(self.labels_list, self.predictions)
return auroc_image
def pixel_auroc(self):
resutls_embeddings = self.anomaly_map_list[0]
for feature in self.anomaly_map_list[1:]:
resutls_embeddings = torch.cat((resutls_embeddings, feature), 0)
resutls_embeddings = ((resutls_embeddings - resutls_embeddings.min())/ (resutls_embeddings.max() - resutls_embeddings.min()))
gt_embeddings = self.gt_list[0]
for feature in self.gt_list[1:]:
gt_embeddings = torch.cat((gt_embeddings, feature), 0)
resutls_embeddings = resutls_embeddings.clone().detach().requires_grad_(False)
gt_embeddings = gt_embeddings.clone().detach().requires_grad_(False)
auroc_p = AUROC(task="binary")
gt_embeddings = torch.flatten(gt_embeddings).type(torch.bool).cpu().detach()
resutls_embeddings = torch.flatten(resutls_embeddings).cpu().detach()
auroc_pixel = auroc_p(resutls_embeddings, gt_embeddings)
return auroc_pixel
def optimal_threshold(self):
fpr, tpr, thresholds = roc_curve(self.labels_list, self.predictions)
# Calculate Youden's J statistic for each threshold
youden_j = tpr - fpr
# Find the optimal threshold that maximizes Youden's J statistic
optimal_threshold_index = np.argmax(youden_j)
optimal_threshold = thresholds[optimal_threshold_index]
self.threshold = optimal_threshold
return optimal_threshold
def pixel_pro(self):
#https://github.com/hq-deng/RD4AD/blob/main/test.py#L337
def _compute_pro(masks, amaps, num_th = 200):
resutls_embeddings = amaps[0]
for feature in amaps[1:]:
resutls_embeddings = torch.cat((resutls_embeddings, feature), 0)
amaps = ((resutls_embeddings - resutls_embeddings.min())/ (resutls_embeddings.max() - resutls_embeddings.min()))
amaps = amaps.squeeze(1)
amaps = amaps.cpu().detach().numpy()
gt_embeddings = masks[0]
for feature in masks[1:]:
gt_embeddings = torch.cat((gt_embeddings, feature), 0)
masks = gt_embeddings.squeeze(1).cpu().detach().numpy()
min_th = amaps.min()
max_th = amaps.max()
delta = (max_th - min_th) / num_th
binary_amaps = np.zeros_like(amaps)
df = pd.DataFrame([], columns=["pro", "fpr", "threshold"])
for th in np.arange(min_th, max_th, delta):
binary_amaps[amaps <= th] = 0
binary_amaps[amaps > th] = 1
pros = []
for binary_amap, mask in zip(binary_amaps, masks):
for region in measure.regionprops(measure.label(mask)):
axes0_ids = region.coords[:, 0]
axes1_ids = region.coords[:, 1]
tp_pixels = binary_amap[axes0_ids, axes1_ids].sum()
pros.append(tp_pixels / region.area)
inverse_masks = 1 - masks
fp_pixels = np.logical_and(inverse_masks , binary_amaps).sum()
fpr = fp_pixels / inverse_masks.sum()
# print(f"Threshold: {th}, FPR: {fpr}, PRO: {mean(pros)}")
df = pd.concat([df, pd.DataFrame({"pro": mean(pros), "fpr": fpr, "threshold": th}, index=[0])], ignore_index=True)
# df = df.concat({"pro": mean(pros), "fpr": fpr, "threshold": th}, ignore_index=True)
# Normalize FPR from 0 ~ 1 to 0 ~ 0.3
df = df[df["fpr"] < 0.3]
df["fpr"] = df["fpr"] / df["fpr"].max()
pro_auc = auc(df["fpr"], df["pro"])
return pro_auc
pro = _compute_pro(self.gt_list, self.anomaly_map_list, num_th = 200)
return pro
def miscalssified(self):
predictions = torch.tensor(self.predictions)
labels_list = torch.tensor(self.labels_list)
predictions0_1 = (predictions > self.threshold).int()
for i,(l,p) in enumerate(zip(labels_list, predictions0_1)):
print('Sample : ', i, ' predicted as: ',p.item() ,' label is: ',l.item(),'\n' ) if l != p else None
================================================
FILE: reconstruction.py
================================================
from typing import Any
import torch
# from forward_process import *
import numpy as np
import os
os.environ['CUDA_VISIBLE_DEVICES'] = "0,1,2"
class Reconstruction:
'''
The reconstruction process
:param y: the target image
:param x: the input image
:param seq: the sequence of denoising steps
:param unet: the UNet model
:param x0_t: the prediction of x0 at time step t
'''
def __init__(self, unet, config) -> None:
self.unet = unet
self.config = config
def __call__(self, x, y0, w) -> Any:
def _compute_alpha(t):
betas = np.linspace(self.config.model.beta_start, self.config.model.beta_end, self.config.model.trajectory_steps, dtype=np.float64)
betas = torch.tensor(betas).type(torch.float).to(self.config.model.device)
beta = torch.cat([torch.zeros(1).to(self.config.model.device), betas], dim=0)
beta = beta.to(self.config.model.device)
a = (1 - beta).cumprod(dim=0).index_select(0, t + 1).view(-1, 1, 1, 1)
return a
test_trajectoy_steps = torch.Tensor([self.config.model.test_trajectoy_steps]).type(torch.int64).to(self.config.model.device).long()
at = _compute_alpha(test_trajectoy_steps)
xt = at.sqrt() * x + (1- at).sqrt() * torch.randn_like(x).to(self.config.model.device)
seq = range(0 , self.config.model.test_trajectoy_steps, self.config.model.skip)
with torch.no_grad():
n = x.size(0)
seq_next = [-1] + list(seq[:-1])
xs = [xt]
for index, (i, j) in enumerate(zip(reversed(seq), reversed(seq_next))):
t = (torch.ones(n) * i).to(self.config.model.device)
next_t = (torch.ones(n) * j).to(self.config.model.device)
at = _compute_alpha(t.long())
at_next = _compute_alpha(next_t.long())
xt = xs[-1].to(self.config.model.device)
self.unet = self.unet.to(self.config.model.device)
et = self.unet(xt, t)
yt = at.sqrt() * y0 + (1- at).sqrt() * et
et_hat = et - (1 - at).sqrt() * w * (yt-xt)
x0_t = (xt - et_hat * (1 - at).sqrt()) / at.sqrt()
c1 = (
self.config.model.eta * ((1 - at / at_next) * (1 - at_next) / (1 - at)).sqrt()
)
c2 = ((1 - at_next) - c1 ** 2).sqrt()
xt_next = at_next.sqrt() * x0_t + c1 * torch.randn_like(x) + c2 * et_hat
xs.append(xt_next)
return xs
================================================
FILE: requirements.txt
================================================
kornia==0.6.12
matplotlib==3.7.1
numpy==1.24.3
omegaconf==2.1.2
opencv-python-headless==4.5.5.64
pandas==2.0.1
Pillow==9.5.0
scikit-image==0.19.2
scikit-learn==1.2.2
scipy==1.10.1
sklearn==0.0.post5
torch==2.0.1
torchmetrics==0.11.4
torchvision==0.15.2
================================================
FILE: resnet.py
================================================
import torch
from torch import Tensor
import torch.nn as nn
try:
from torch.hub import load_state_dict_from_url
except ImportError:
from torch.utils.model_zoo import load_url as load_state_dict_from_url
from typing import Type, Any, Callable, Union, List, Optional
__all__ = ['ResNet', 'resnet18', 'resnet34', 'resnet50', 'resnet101',
'resnet152', 'resnext50_32x4d', 'resnext101_32x8d',
'wide_resnet50_2', 'wide_resnet101_2']
model_urls = {
'resnet18': 'https://download.pytorch.org/models/resnet18-f37072fd.pth',
'resnet34': 'https://download.pytorch.org/models/resnet34-b627a593.pth',
'resnet50': 'https://download.pytorch.org/models/resnet50-0676ba61.pth',
'resnet101': 'https://download.pytorch.org/models/resnet101-63fe2227.pth',
'resnet152': 'https://download.pytorch.org/models/resnet152-394f9c45.pth',
'resnext50_32x4d': 'https://download.pytorch.org/models/resnext50_32x4d-7cdf4587.pth',
'resnext101_32x8d': 'https://download.pytorch.org/models/resnext101_32x8d-8ba56ff5.pth',
'wide_resnet50_2': 'https://download.pytorch.org/models/wide_resnet50_2-95faca4d.pth',
'wide_resnet101_2': 'https://download.pytorch.org/models/wide_resnet101_2-32ee1156.pth',
}
def conv3x3(in_planes: int, out_planes: int, stride: int = 1, groups: int = 1, dilation: int = 1) -> nn.Conv2d:
"""3x3 convolution with padding"""
return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
padding=dilation, groups=groups, bias=False, dilation=dilation)
def conv1x1(in_planes: int, out_planes: int, stride: int = 1) -> nn.Conv2d:
"""1x1 convolution"""
return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False)
class BasicBlock(nn.Module):
expansion: int = 1
def __init__(
self,
inplanes: int,
planes: int,
stride: int = 1,
downsample: Optional[nn.Module] = None,
groups: int = 1,
base_width: int = 64,
dilation: int = 1,
norm_layer: Optional[Callable[..., nn.Module]] = None
) -> None:
super(BasicBlock, self).__init__()
if norm_layer is None:
norm_layer = nn.BatchNorm2d
if groups != 1 or base_width != 64:
raise ValueError('BasicBlock only supports groups=1 and base_width=64')
if dilation > 1:
raise NotImplementedError("Dilation > 1 not supported in BasicBlock")
# Both self.conv1 and self.downsample layers downsample the input when stride != 1
self.conv1 = conv3x3(inplanes, planes, stride)
self.bn1 = norm_layer(planes)
self.relu = nn.ReLU(inplace=True)
self.conv2 = conv3x3(planes, planes)
self.bn2 = norm_layer(planes)
self.downsample = downsample
self.stride = stride
def forward(self, x: Tensor) -> Tensor:
identity = x
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
out = self.conv2(out)
out = self.bn2(out)
if self.downsample is not None:
identity = self.downsample(x)
out += identity
out = self.relu(out)
return out
class Bottleneck(nn.Module):
# Bottleneck in torchvision places the stride for downsampling at 3x3 convolution(self.conv2)
# while original implementation places the stride at the first 1x1 convolution(self.conv1)
# according to "Deep residual learning for image recognition"https://arxiv.org/abs/1512.03385.
# This variant is also known as ResNet V1.5 and improves accuracy according to
# https://ngc.nvidia.com/catalog/model-scripts/nvidia:resnet_50_v1_5_for_pytorch.
expansion: int = 4
def __init__(
self,
inplanes: int,
planes: int,
stride: int = 1,
downsample: Optional[nn.Module] = None,
groups: int = 1,
base_width: int = 64,
dilation: int = 1,
norm_layer: Optional[Callable[..., nn.Module]] = None
) -> None:
super(Bottleneck, self).__init__()
if norm_layer is None:
norm_layer = nn.BatchNorm2d
width = int(planes * (base_width / 64.)) * groups
# Both self.conv2 and self.downsample layers downsample the input when stride != 1
self.conv1 = conv1x1(inplanes, width)
self.bn1 = norm_layer(width)
self.conv2 = conv3x3(width, width, stride, groups, dilation)
self.bn2 = norm_layer(width)
self.conv3 = conv1x1(width, planes * self.expansion)
self.bn3 = norm_layer(planes * self.expansion)
self.relu = nn.ReLU(inplace=True)
self.downsample = downsample
self.stride = stride
def forward(self, x: Tensor) -> Tensor:
identity = x
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
out = self.conv2(out)
out = self.bn2(out)
out = self.relu(out)
out = self.conv3(out)
out = self.bn3(out)
if self.downsample is not None:
identity = self.downsample(x)
out += identity
out = self.relu(out)
return out
class ResNet(nn.Module):
def __init__(
self,
block: Type[Union[BasicBlock, Bottleneck]],
layers: List[int],
num_classes: int = 1000,
zero_init_residual: bool = False,
groups: int = 1,
width_per_group: int = 64,
replace_stride_with_dilation: Optional[List[bool]] = None,
norm_layer: Optional[Callable[..., nn.Module]] = None
) -> None:
super(ResNet, self).__init__()
if norm_layer is None:
norm_layer = nn.BatchNorm2d
self._norm_layer = norm_layer
self.inplanes = 64
self.dilation = 1
if replace_stride_with_dilation is None:
# each element in the tuple indicates if we should replace
# the 2x2 stride with a dilated convolution instead
replace_stride_with_dilation = [False, False, False]
if len(replace_stride_with_dilation) != 3:
raise ValueError("replace_stride_with_dilation should be None "
"or a 3-element tuple, got {}".format(replace_stride_with_dilation))
self.groups = groups
self.base_width = width_per_group
self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3,
bias=False)
self.bn1 = norm_layer(self.inplanes)
self.relu = nn.ReLU(inplace=True)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.layer1 = self._make_layer(block, 64, layers[0])
self.layer2 = self._make_layer(block, 128, layers[1], stride=2,
dilate=replace_stride_with_dilation[0])
self.layer3 = self._make_layer(block, 256, layers[2], stride=2,
dilate=replace_stride_with_dilation[1])
self.layer4 = self._make_layer(block, 512, layers[3], stride=2,
dilate=replace_stride_with_dilation[2])
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.fc = nn.Linear(512 * block.expansion, num_classes)
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
# Zero-initialize the last BN in each residual branch,
# so that the residual branch starts with zeros, and each residual block behaves like an identity.
# This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
if zero_init_residual:
for m in self.modules():
if isinstance(m, Bottleneck):
nn.init.constant_(m.bn3.weight, 0) # type: ignore[arg-type]
elif isinstance(m, BasicBlock):
nn.init.constant_(m.bn2.weight, 0) # type: ignore[arg-type]
def _make_layer(self, block: Type[Union[BasicBlock, Bottleneck]], planes: int, blocks: int,
stride: int = 1, dilate: bool = False) -> nn.Sequential:
norm_layer = self._norm_layer
downsample = None
previous_dilation = self.dilation
if dilate:
self.dilation *= stride
stride = 1
if stride != 1 or self.inplanes != planes * block.expansion:
downsample = nn.Sequential(
conv1x1(self.inplanes, planes * block.expansion, stride),
norm_layer(planes * block.expansion),
)
layers = []
layers.append(block(self.inplanes, planes, stride, downsample, self.groups,
self.base_width, previous_dilation, norm_layer))
self.inplanes = planes * block.expansion
for _ in range(1, blocks):
layers.append(block(self.inplanes, planes, groups=self.groups,
base_width=self.base_width, dilation=self.dilation,
norm_layer=norm_layer))
return nn.Sequential(*layers)
def _forward_impl(self, x: Tensor) -> Tensor:
# See note [TorchScript super()]
x = self.conv1(x)
x = self.bn1(x)
x = self.relu(x)
x = self.maxpool(x)
feature_a = self.layer1(x)
feature_b = self.layer2(feature_a)
feature_c = self.layer3(feature_b)
feature_d = self.layer4(feature_c)
return [feature_a, feature_b, feature_c]
def forward(self, x: Tensor) -> Tensor:
return self._forward_impl(x)
def _resnet(
arch: str,
block: Type[Union[BasicBlock, Bottleneck]],
layers: List[int],
pretrained: bool,
progress: bool,
**kwargs: Any
) -> ResNet:
model = ResNet(block, layers, **kwargs)
if pretrained:
state_dict = load_state_dict_from_url(model_urls[arch],
progress=progress)
#for k,v in list(state_dict.items()):
# if 'layer4' in k or 'fc' in k:
# state_dict.pop(k)
model.load_state_dict(state_dict)
return model
class AttnBasicBlock(nn.Module):
expansion: int = 1
def __init__(
self,
inplanes: int,
planes: int,
stride: int = 1,
downsample: Optional[nn.Module] = None,
groups: int = 1,
base_width: int = 64,
dilation: int = 1,
norm_layer: Optional[Callable[..., nn.Module]] = None,
attention: bool = True,
) -> None:
super(AttnBasicBlock, self).__init__()
self.attention = attention
#print("Attention:", self.attention)
if norm_layer is None:
norm_layer = nn.BatchNorm2d
if groups != 1 or base_width != 64:
raise ValueError('BasicBlock only supports groups=1 and base_width=64')
if dilation > 1:
raise NotImplementedError("Dilation > 1 not supported in BasicBlock")
# Both self.conv1 and self.downsample layers downsample the input when stride != 1
self.conv1 = conv3x3(inplanes, planes, stride)
self.bn1 = norm_layer(planes)
self.relu = nn.ReLU(inplace=True)
self.conv2 = conv3x3(planes, planes)
self.bn2 = norm_layer(planes)
#self.cbam = GLEAM(planes, 16)
self.downsample = downsample
self.stride = stride
def forward(self, x: Tensor) -> Tensor:
#if self.attention:
# x = self.cbam(x)
identity = x
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
out = self.conv2(out)
out = self.bn2(out)
if self.downsample is not None:
identity = self.downsample(x)
out += identity
out = self.relu(out)
return out
class AttnBottleneck(nn.Module):
expansion: int = 4
def __init__(
self,
inplanes: int,
planes: int,
stride: int = 1,
downsample: Optional[nn.Module] = None,
groups: int = 1,
base_width: int = 64,
dilation: int = 1,
norm_layer: Optional[Callable[..., nn.Module]] = None,
attention: bool = True,
) -> None:
super(AttnBottleneck, self).__init__()
self.attention = attention
#print("Attention:",self.attention)
if norm_layer is None:
norm_layer = nn.BatchNorm2d
width = int(planes * (base_width / 64.)) * groups
# Both self.conv2 and self.downsample layers downsample the input when stride != 1
self.conv1 = conv1x1(inplanes, width)
self.bn1 = norm_layer(width)
self.conv2 = conv3x3(width, width, stride, groups, dilation)
self.bn2 = norm_layer(width)
self.conv3 = conv1x1(width, planes * self.expansion)
self.bn3 = norm_layer(planes * self.expansion)
self.relu = nn.ReLU(inplace=True)
#self.cbam = GLEAM([int(planes * self.expansion/4),
# int(planes * self.expansion//2),
# planes * self.expansion], 16)
self.downsample = downsample
self.stride = stride
def forward(self, x: Tensor) -> Tensor:
#if self.attention:
# x = self.cbam(x)
identity = x
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
out = self.conv2(out)
out = self.bn2(out)
out = self.relu(out)
out = self.conv3(out)
out = self.bn3(out)
if self.downsample is not None:
identity = self.downsample(x)
out += identity
out = self.relu(out)
return out
def __init__(self,
block: Type[Union[BasicBlock, Bottleneck]],
layers: int,
groups: int = 1,
width_per_group: int = 64,
norm_layer: Optional[Callable[..., nn.Module]] = None,
):
super(BN_layer, self).__init__()
if norm_layer is None:
norm_layer = nn.BatchNorm2d
self._norm_layer = norm_layer
self.groups = groups
self.base_width = width_per_group
self.inplanes = 256 * block.expansion
self.dilation = 1
self.bn_layer = self._make_layer(block, 512, layers, stride=2)
self.conv1 = conv3x3(64 * block.expansion, 128 * block.expansion, 2)
self.bn1 = norm_layer(128 * block.expansion)
self.relu = nn.ReLU(inplace=True)
self.conv2 = conv3x3(128 * block.expansion, 256 * block.expansion, 2)
self.bn2 = norm_layer(256 * block.expansion)
self.conv3 = conv3x3(128 * block.expansion, 256 * block.expansion, 2)
self.bn3 = norm_layer(256 * block.expansion)
self.conv4 = conv1x1(1024 * block.expansion, 512 * block.expansion, 1)
self.bn4 = norm_layer(512 * block.expansion)
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
def _make_layer(self, block: Type[Union[BasicBlock, Bottleneck]], planes: int, blocks: int,
stride: int = 1, dilate: bool = False) -> nn.Sequential:
norm_layer = self._norm_layer
downsample = None
previous_dilation = self.dilation
if dilate:
self.dilation *= stride
stride = 1
if stride != 1 or self.inplanes != planes * block.expansion:
downsample = nn.Sequential(
conv1x1(self.inplanes*3, planes * block.expansion, stride),
norm_layer(planes * block.expansion),
)
layers = []
layers.append(block(self.inplanes*3, planes, stride, downsample, self.groups,
self.base_width, previous_dilation, norm_layer))
self.inplanes = planes * block.expansion
for _ in range(1, blocks):
layers.append(block(self.inplanes, planes, groups=self.groups,
base_width=self.base_width, dilation=self.dilation,
norm_layer=norm_layer))
return nn.Sequential(*layers)
def _forward_impl(self, x: Tensor) -> Tensor:
# See note [TorchScript super()]
#x = self.cbam(x)
l1 = self.relu(self.bn2(self.conv2(self.relu(self.bn1(self.conv1(x[0]))))))
l2 = self.relu(self.bn3(self.conv3(x[1])))
feature = torch.cat([l1,l2,x[2]],1)
output = self.bn_layer(feature)
#x = self.avgpool(feature_d)
#x = torch.flatten(x, 1)
#x = self.fc(x)
return output.contiguous()
def forward(self, x: Tensor) -> Tensor:
return self._forward_impl(x)
def resnet18(pretrained: bool = False, progress: bool = True,**kwargs: Any) -> ResNet:
r"""ResNet-18 model from
`"Deep Residual Learning for Image Recognition" <https://arxiv.org/pdf/1512.03385.pdf>`_.
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
progress (bool): If True, displays a progress bar of the download to stderr
"""
return _resnet('resnet18', BasicBlock, [2, 2, 2, 2], pretrained, progress,
**kwargs)
def resnet34(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet:
r"""ResNet-34 model from
`"Deep Residual Learning for Image Recognition" <https://arxiv.org/pdf/1512.03385.pdf>`_.
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
progress (bool): If True, displays a progress bar of the download to stderr
"""
return _resnet('resnet34', BasicBlock, [3, 4, 6, 3], pretrained, progress,
**kwargs)
def resnet50(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet:
r"""ResNet-50 model from
`"Deep Residual Learning for Image Recognition" <https://arxiv.org/pdf/1512.03385.pdf>`_.
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
progress (bool): If True, displays a progress bar of the download to stderr
"""
return _resnet('resnet50', Bottleneck, [3, 4, 6, 3], pretrained, progress,
**kwargs)
def resnet101(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet:
r"""ResNet-101 model from
`"Deep Residual Learning for Image Recognition" <https://arxiv.org/pdf/1512.03385.pdf>`_.
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
progress (bool): If True, displays a progress bar of the download to stderr
"""
return _resnet('resnet101', Bottleneck, [3, 4, 23, 3], pretrained, progress,
**kwargs)
def resnet152(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet:
r"""ResNet-152 model from
`"Deep Residual Learning for Image Recognition" <https://arxiv.org/pdf/1512.03385.pdf>`_.
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
progress (bool): If True, displays a progress bar of the download to stderr
"""
return _resnet('resnet152', Bottleneck, [3, 8, 36, 3], pretrained, progress,
**kwargs)
def resnext50_32x4d(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet:
r"""ResNeXt-50 32x4d model from
`"Aggregated Residual Transformation for Deep Neural Networks" <https://arxiv.org/pdf/1611.05431.pdf>`_.
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
progress (bool): If True, displays a progress bar of the download to stderr
"""
kwargs['groups'] = 32
kwargs['width_per_group'] = 4
return _resnet('resnext50_32x4d', Bottleneck, [3, 4, 6, 3],
pretrained, progress, **kwargs)
def resnext101_32x8d(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet:
r"""ResNeXt-101 32x8d model from
`"Aggregated Residual Transformation for Deep Neural Networks" <https://arxiv.org/pdf/1611.05431.pdf>`_.
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
progress (bool): If True, displays a progress bar of the download to stderr
"""
kwargs['groups'] = 32
kwargs['width_per_group'] = 8
return _resnet('resnext101_32x8d', Bottleneck, [3, 4, 23, 3],
pretrained, progress, **kwargs)
def wide_resnet50_2(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet:
r"""Wide ResNet-50-2 model from
`"Wide Residual Networks" <https://arxiv.org/pdf/1605.07146.pdf>`_.
The model is the same as ResNet except for the bottleneck number of channels
which is twice larger in every block. The number of channels in outer 1x1
convolutions is the same, e.g. last block in ResNet-50 has 2048-512-2048
channels, and in Wide ResNet-50-2 has 2048-1024-2048.
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
progress (bool): If True, displays a progress bar of the download to stderr
"""
kwargs['width_per_group'] = 64 * 2
return _resnet('wide_resnet50_2', Bottleneck, [3, 4, 6, 3],
pretrained, progress, **kwargs)
def wide_resnet101_2(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet:
r"""Wide ResNet-101-2 model from
`"Wide Residual Networks" <https://arxiv.org/pdf/1605.07146.pdf>`_.
The model is the same as ResNet except for the bottleneck number of channels
which is twice larger in every block. The number of channels in outer 1x1
convolutions is the same, e.g. last block in ResNet-50 has 2048-512-2048
channels, and in Wide ResNet-50-2 has 2048-1024-2048.
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
progress (bool): If True, displays a progress bar of the download to stderr
"""
kwargs['width_per_group'] = 64 * 2
return _resnet('wide_resnet101_2', Bottleneck, [3, 4, 23, 3],
pretrained, progress, **kwargs)
================================================
FILE: train.py
================================================
import torch
import os
import torch.nn as nn
from dataset import *
from dataset import *
from loss import *
def trainer(model, category, config):
'''
Training the UNet model
:param model: the UNet model
:param category: the category of the dataset
'''
# optimizer = torch.optim.AdamW(
# model.parameters(), lr=config.model.learning_rate)
optimizer = torch.optim.Adam(
model.parameters(), lr=config.model.learning_rate, weight_decay=config.model.weight_decay
)
train_dataset = Dataset_maker(
root= config.data.data_dir,
category=category,
config = config,
is_train=True,
)
trainloader = torch.utils.data.DataLoader(
train_dataset,
batch_size=config.data.batch_size,
shuffle=True,
num_workers=config.model.num_workers,
drop_last=True,
)
if not os.path.exists('checkpoints'):
os.mkdir('checkpoints')
if not os.path.exists(config.model.checkpoint_dir):
os.mkdir(config.model.checkpoint_dir)
for epoch in range(config.model.epochs):
for step, batch in enumerate(trainloader):
optimizer.zero_grad()
# loss = 0
# for _ in range(2):
t = torch.randint(0, config.model.trajectory_steps, (batch[0].shape[0],), device=config.model.device).long()
loss = get_loss(model, batch[0], t, config)
loss.backward()
optimizer.step()
if (epoch+1) % 25 == 0 and step == 0:
print(f"Epoch {epoch+1} | Loss: {loss.item()}")
if (epoch+1) %250 == 0 and epoch>0 and step ==0:
if config.model.save_model:
model_save_dir = os.path.join(os.getcwd(), config.model.checkpoint_dir, category)
if not os.path.exists(model_save_dir):
os.mkdir(model_save_dir)
torch.save(model.state_dict(), os.path.join(model_save_dir, str(epoch+1)))
================================================
FILE: unet.py
================================================
# https://github.com/openai/guided-diffusion/tree/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924
import math
from abc import abstractmethod
import numpy as np
import torch
import torch.nn.functional as F
from torch import nn
class TimestepBlock(nn.Module):
"""
Any module where forward() takes timestep embeddings as a second argument.
"""
@abstractmethod
def forward(self, x, emb):
"""
Apply the module to `x` given `emb` timestep embeddings.
"""
class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
"""
A sequential module that passes timestep embeddings to the children that
support it as an extra input.
"""
def forward(self, x, emb):
for layer in self:
if isinstance(layer, TimestepBlock):
x = layer(x, emb)
else:
x = layer(x)
return x
class PositionalEmbedding(nn.Module):
# PositionalEmbedding
"""
Computes Positional Embedding of the timestep
"""
def __init__(self, dim, scale=1):
super().__init__()
assert dim % 2 == 0
self.dim = dim
self.scale = scale
def forward(self, x):
device = x.device
half_dim = self.dim // 2
emb = np.log(10000) / half_dim
emb = torch.exp(torch.arange(half_dim, device=device) * -emb)
emb = torch.outer(x * self.scale, emb)
emb = torch.cat((emb.sin(), emb.cos()), dim=-1)
return emb
class Downsample(nn.Module):
def __init__(self, in_channels, use_conv, out_channels=None):
super().__init__()
self.channels = in_channels
out_channels = out_channels or in_channels
if use_conv:
# downsamples by 1/2
self.downsample = nn.Conv2d(in_channels, out_channels, 3, stride=2, padding=1)
else:
assert in_channels == out_channels
self.downsample = nn.AvgPool2d(kernel_size=2, stride=2)
def forward(self, x, time_embed=None):
assert x.shape[1] == self.channels
return self.downsample(x)
class Upsample(nn.Module):
def __init__(self, in_channels, use_conv, out_channels=None):
super().__init__()
self.channels = in_channels
self.use_conv = use_conv
# uses upsample then conv to avoid checkerboard artifacts
# self.upsample = nn.Upsample(scale_factor=2, mode="nearest")
if use_conv:
self.conv = nn.Conv2d(in_channels, out_channels, 3, padding=1)
def forward(self, x, time_embed=None):
assert x.shape[1] == self.channels
x = F.interpolate(x, scale_factor=2, mode="nearest")
if self.use_conv:
x = self.conv(x)
return x
class AttentionBlock(nn.Module):
"""
An attention block that allows spatial positions to attend to each other.
Originally ported from here, but adapted to the N-d case.
https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66.
"""
def __init__(self, in_channels, n_heads=1, n_head_channels=-1):
super().__init__()
self.in_channels = in_channels
self.norm = GroupNorm32(32, self.in_channels)
if n_head_channels == -1:
self.num_heads = n_heads
else:
assert (
in_channels % n_head_channels == 0
), f"q,k,v channels {in_channels} is not divisible by num_head_channels {n_head_channels}"
self.num_heads = in_channels // n_head_channels
# query, key, value for attention
self.to_qkv = nn.Conv1d(in_channels, in_channels * 3, 1)
self.attention = QKVAttention(self.num_heads)
self.proj_out = zero_module(nn.Conv1d(in_channels, in_channels, 1))
def forward(self, x, time=None):
b, c, *spatial = x.shape
x = x.reshape(b, c, -1)
qkv = self.to_qkv(self.norm(x))
h = self.attention(qkv)
h = self.proj_out(h)
return (x + h).reshape(b, c, *spatial)
class QKVAttention(nn.Module):
"""
A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping
"""
def __init__(self, n_heads):
super().__init__()
self.n_heads = n_heads
def forward(self, qkv, time=None):
"""
Apply QKV attention.
:param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs.
:return: an [N x (H * C) x T] tensor after attention.
"""
bs, width, length = qkv.shape
assert width % (3 * self.n_heads) == 0
ch = width // (3 * self.n_heads)
q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch, dim=1)
scale = 1 / math.sqrt(math.sqrt(ch))
weight = torch.einsum(
"bct,bcs->bts", q * scale, k * scale
) # More stable with f16 than dividing afterwards
weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
a = torch.einsum("bts,bcs->bct", weight, v)
return a.reshape(bs, -1, length)
class ResBlock(TimestepBlock):
def __init__(
self,
in_channels,
time_embed_dim,
dropout,
out_channels=None,
use_conv=False,
up=False,
down=False
):
super().__init__()
out_channels = out_channels or in_channels
self.in_layers = nn.Sequential(
GroupNorm32(32, in_channels),
nn.SiLU(),
nn.Conv2d(in_channels, out_channels, 3, padding=1)
)
self.updown = up or down
if up:
self.h_upd = Upsample(in_channels, False)
self.x_upd = Upsample(in_channels, False)
elif down:
self.h_upd = Downsample(in_channels, False)
self.x_upd = Downsample(in_channels, False)
else:
self.h_upd = self.x_upd = nn.Identity()
self.embed_layers = nn.Sequential(
nn.SiLU(),
nn.Linear(time_embed_dim, out_channels)
)
self.out_layers = nn.Sequential(
GroupNorm32(32, out_channels),
nn.SiLU(),
nn.Dropout(p=dropout),
zero_module(nn.Conv2d(out_channels, out_channels, 3, padding=1))
)
if out_channels == in_channels:
self.skip_connection = nn.Identity()
elif use_conv:
self.skip_connection = nn.Conv2d(in_channels, out_channels, 3, padding=1)
else:
self.skip_connection = nn.Conv2d(in_channels, out_channels, 1)
def forward(self, x, time_embed):
if self.updown:
in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
h = in_rest(x)
h = self.h_upd(h)
x = self.x_upd(x)
h = in_conv(h)
else:
h = self.in_layers(x)
emb_out = self.embed_layers(time_embed).type(h.dtype)
while len(emb_out.shape) < len(h.shape):
emb_out = emb_out[..., None]
h = h + emb_out
h = self.out_layers(h)
return self.skip_connection(x) + h
class UNetModel(nn.Module):
# UNet model
def __init__(
self,
img_size,
base_channels,
conv_resample=True,
n_heads=1,
n_head_channels=-1,
channel_mults="",
num_res_blocks=2,
dropout=0,
attention_resolutions="32,16,8",
biggan_updown=True,
in_channels=1
):
self.dtype = torch.float32
super().__init__()
if channel_mults == "":
if img_size == 512:
channel_mults = (0.5, 1, 1, 2, 2, 4, 4)
elif img_size == 256:
channel_mults = (1, 1, 2, 2, 4, 4)#
elif img_size == 128:
channel_mults = (1, 1, 2, 3, 4)
elif img_size == 64:
channel_mults = (1, 2, 3, 4)
elif img_size == 32:
channel_mults = (1, 2, 3, 4)
else:
raise ValueError(f"unsupported image size: {img_size}")
attention_ds = []
for res in attention_resolutions.split(","):
attention_ds.append(img_size // int(res))
self.image_size = img_size
self.in_channels = in_channels
self.model_channels = base_channels
self.out_channels = in_channels
self.num_res_blocks = num_res_blocks
self.attention_resolutions = attention_resolutions
self.dropout = dropout
self.channel_mult = channel_mults
self.conv_resample = conv_resample
self.dtype = torch.float32
self.num_heads = n_heads
self.num_head_channels = n_head_channels
time_embed_dim = base_channels * 4
self.time_embedding = nn.Sequential(
PositionalEmbedding(base_channels, 1),
nn.Linear(base_channels, time_embed_dim),
nn.SiLU(),
nn.Linear(time_embed_dim, time_embed_dim),
)
ch = int(channel_mults[0] * base_channels)
self.down = nn.ModuleList(
[TimestepEmbedSequential(nn.Conv2d(self.in_channels, base_channels, 3, padding=1))]
)
channels = [ch]
ds = 1
for i, mult in enumerate(channel_mults):
# out_channels = base_channels * mult
for _ in range(num_res_blocks):
layers = [ResBlock(
ch,
time_embed_dim=time_embed_dim,
out_channels=base_channels * mult,
dropout=dropout,
)]
ch = base_channels * mult
# channels.append(ch)
if ds in attention_ds:
layers.append(
AttentionBlock(
ch,
n_heads=n_heads,
n_head_channels=n_head_channels,
)
)
self.down.append(TimestepEmbedSequential(*layers))
channels.append(ch)
if i != len(channel_mults) - 1:
out_channels = ch
self.down.append(
TimestepEmbedSequential(
ResBlock(
ch,
time_embed_dim=time_embed_dim,
out_channels=out_channels,
dropout=dropout,
down=True
)
if biggan_updown
else
Downsample(ch, conv_resample, out_channels=out_channels)
)
)
ds *= 2
ch = out_channels
channels.append(ch)
self.middle = TimestepEmbedSequential(
ResBlock(
ch,
time_embed_dim=time_embed_dim,
dropout=dropout
),
AttentionBlock(
ch,
n_heads=n_heads,
n_head_channels=n_head_channels
),
ResBlock(
ch,
time_embed_dim=time_embed_dim,
dropout=dropout
)
)
self.up = nn.ModuleList([])
for i, mult in reversed(list(enumerate(channel_mults))):
for j in range(num_res_blocks + 1):
inp_chs = channels.pop()
layers = [
ResBlock(
ch + inp_chs,
time_embed_dim=time_embed_dim,
out_channels=base_channels * mult,
dropout=dropout
)
]
ch = base_channels * mult
if ds in attention_ds:
layers.append(
AttentionBlock(
ch,
n_heads=n_heads,
n_head_channels=n_head_channels
),
)
if i and j == num_res_blocks:
out_channels = ch
layers.append(
ResBlock(
ch,
time_embed_dim=time_embed_dim,
out_channels=out_channels,
dropout=dropout,
up=True
)
if biggan_updown
else
Upsample(ch, conv_resample, out_channels=out_channels)
)
ds //= 2
self.up.append(TimestepEmbedSequential(*layers))
self.out = nn.Sequential(
GroupNorm32(32, ch),
nn.SiLU(),
zero_module(nn.Conv2d(base_channels * channel_mults[0], self.out_channels, 3, padding=1))
)
def forward(self, x, time):
time_embed = self.time_embedding(time)
skips = []
h = x.type(self.dtype)
for i, module in enumerate(self.down):
h = module(h, time_embed)
skips.append(h)
h = self.middle(h, time_embed)
for i, module in enumerate(self.up):
h = torch.cat([h, skips.pop()], dim=1)
h = module(h, time_embed)
h = h.type(x.dtype)
h = self.out(h)
return h
class GroupNorm32(nn.GroupNorm):
def forward(self, x):
return super().forward(x.float()).type(x.dtype)
def zero_module(module):
"""
Zero out the parameters of a module and return it.
"""
for p in module.parameters():
p.detach().zero_()
return module
def update_ema_params(target, source, decay_rate=0.9999):
targParams = dict(target.named_parameters())
srcParams = dict(source.named_parameters())
for k in targParams:
targParams[k].data.mul_(decay_rate).add_(srcParams[k].data, alpha=1 - decay_rate)
================================================
FILE: visualize.py
================================================
import matplotlib.pyplot as plt
from torchvision.transforms import transforms
import numpy as np
import torch
import os
from dataset import *
def visualalize_reconstruction(input, recon, target):
plt.figure(figsize=(11,11))
plt.subplot(1, 3, 1).axis('off')
plt.subplot(1, 3, 2).axis('off')
plt.subplot(1, 3, 3).axis('off')
plt.subplot(1, 3, 1)
plt.imshow(show_tensor_image(input))
plt.title('input image')
plt.subplot(1, 3, 2)
plt.imshow(show_tensor_mask(recon))
plt.title('recon image')
plt.subplot(1, 3, 3)
plt.imshow(show_tensor_mask(target))
plt.title('target image')
k = 0
while os.path.exists('results/heatmap{}.png'.format(k)):
k += 1
plt.savefig('results/heatmap{}.png'.format(k))
plt.close()
# def visualize_reconstructed(input, data,s):
# fig, axs = plt.subplots(int(len(data)/5),6)
# row = 0
# col = 1
# axs[0,0].imshow(show_tensor_image(input))
# axs[0, 0].get_xaxis().set_visible(False)
# axs[0, 0].get_yaxis().set_visible(False)
# axs[0,0].set_title('input')
# for i, img in enumerate(data):
# axs[row, col].imshow(show_tensor_image(img))
# axs[row, col].get_xaxis().set_visible(False)
# axs[row, col].get_yaxis().set_visible(False)
# axs[row, col].set_title(str(i))
# col += 1
# if col == 6:
# row += 1
# col = 0
# col = 6
# row = int(len(data)/5)
# remain = col * row - len(data) -1
# for j in range(remain):
# col -= 1
# axs[row-1, col].remove()
# axs[row-1, col].get_xaxis().set_visible(False)
# axs[row-1, col].get_yaxis().set_visible(False)
# plt.subplots_adjust(left=0.1,
# bottom=0.1,
# right=0.9,
# top=0.9,
# wspace=0.4,
# hspace=0.4)
# k = 0
# while os.path.exists(f'results/reconstructed{k}{s}.png'):
# k += 1
# plt.savefig(f'results/reconstructed{k}{s}.png')
# plt.close()
def visualize(image, noisy_image, GT, pred_mask, anomaly_map, category) :
for idx, img in enumerate(image):
plt.figure(figsize=(11,11))
plt.subplot(1, 2, 1).axis('off')
plt.subplot(1, 2, 2).axis('off')
plt.subplot(1, 2, 1)
plt.imshow(show_tensor_image(image[idx]))
plt.title('clear image')
plt.subplot(1, 2, 2)
plt.imshow(show_tensor_image(noisy_image[idx]))
plt.title('reconstructed image')
plt.savefig('results/{}sample{}.png'.format(category,idx))
plt.close()
plt.figure(figsize=(11,11))
plt.subplot(1, 3, 1).axis('off')
plt.subplot(1, 3, 2).axis('off')
plt.subplot(1, 3, 3).axis('off')
plt.subplot(1, 3, 1)
plt.imshow(show_tensor_mask(GT[idx]))
plt.title('ground truth')
plt.subplot(1, 3, 2)
plt.imshow(show_tensor_mask(pred_mask[idx]))
plt.title('normal' if torch.max(pred_mask[idx]) == 0 else 'abnormal', color="g" if torch.max(pred_mask[idx]) == 0 else "r")
plt.subplot(1, 3, 3)
plt.imshow(show_tensor_image(anomaly_map[idx]))
plt.title('heat map')
plt.savefig('results/{}sample{}heatmap.png'.format(category,idx))
plt.close()
def show_tensor_image(image):
reverse_transforms = transforms.Compose([
transforms.Lambda(lambda t: (t + 1) / (2)),
transforms.Lambda(lambda t: t.permute(1, 2, 0)), # CHW to HWC
transforms.Lambda(lambda t: t * 255.),
transforms.Lambda(lambda t: t.cpu().numpy().astype(np.uint8)),
])
# Takes the first image of batch
if len(image.shape) == 4:
image = image[0, :, :, :]
return reverse_transforms(image)
def show_tensor_mask(image):
reverse_transforms = transforms.Compose([
# transforms.Lambda(lambda t: (t + 1) / (2)),
transforms.Lambda(lambda t: t.permute(1, 2, 0)), # CHW to HWC
transforms.Lambda(lambda t: t.cpu().numpy().astype(np.int8)),
])
# Takes the first image of batch
if len(image.shape) == 4:
image = image[0, :, :, :]
return reverse_transforms(image)
gitextract_ujrg7voc/ ├── LICENSE ├── README.md ├── anomaly_map.py ├── config.yaml ├── dataset.py ├── ddad.py ├── feature_extractor.py ├── loss.py ├── main.py ├── metrics.py ├── reconstruction.py ├── requirements.txt ├── resnet.py ├── train.py ├── unet.py └── visualize.py
SYMBOL INDEX (96 symbols across 12 files)
FILE: anomaly_map.py
function heat_map (line 12) | def heat_map(output, target, FE, config):
function pixel_distance (line 42) | def pixel_distance(output, target):
function feature_distance (line 52) | def feature_distance(output, target, FE, config):
function patchify (line 78) | def patchify(features, return_spatial_info=False):
FILE: dataset.py
class Dataset_maker (line 17) | class Dataset_maker(torch.utils.data.Dataset):
method __init__ (line 18) | def __init__(self, root, category, config, is_train=True):
method __getitem__ (line 49) | def __getitem__(self, index):
method __len__ (line 85) | def __len__(self):
FILE: ddad.py
class DDAD (line 13) | class DDAD:
method __init__ (line 14) | def __init__(self, unet, config) -> None:
method __call__ (line 35) | def __call__(self) -> Any:
FILE: feature_extractor.py
function loss_fucntion (line 13) | def loss_fucntion(a, b, c, d, config):
function domain_adaptation (line 27) | def domain_adaptation(unet, config, fine_tune):
FILE: loss.py
function get_loss (line 6) | def get_loss(model, x_0, t, config):
FILE: main.py
function build_model (line 12) | def build_model(config):
function train (line 19) | def train(config):
function detection (line 32) | def detection(config):
function finetuning (line 44) | def finetuning(config):
function parse_args (line 57) | def parse_args():
FILE: metrics.py
class Metric (line 14) | class Metric:
method __init__ (line 15) | def __init__(self,labels_list, predictions, anomaly_map_list, gt_list,...
method image_auroc (line 23) | def image_auroc(self):
method pixel_auroc (line 27) | def pixel_auroc(self):
method optimal_threshold (line 47) | def optimal_threshold(self):
method pixel_pro (line 60) | def pixel_pro(self):
method miscalssified (line 110) | def miscalssified(self):
FILE: reconstruction.py
class Reconstruction (line 8) | class Reconstruction:
method __init__ (line 17) | def __init__(self, unet, config) -> None:
method __call__ (line 23) | def __call__(self, x, y0, w) -> Any:
FILE: resnet.py
function conv3x3 (line 29) | def conv3x3(in_planes: int, out_planes: int, stride: int = 1, groups: in...
function conv1x1 (line 35) | def conv1x1(in_planes: int, out_planes: int, stride: int = 1) -> nn.Conv2d:
class BasicBlock (line 40) | class BasicBlock(nn.Module):
method __init__ (line 43) | def __init__(
method forward (line 70) | def forward(self, x: Tensor) -> Tensor:
class Bottleneck (line 89) | class Bottleneck(nn.Module):
method __init__ (line 98) | def __init__(
method forward (line 124) | def forward(self, x: Tensor) -> Tensor:
class ResNet (line 147) | class ResNet(nn.Module):
method __init__ (line 149) | def __init__(
method _make_layer (line 208) | def _make_layer(self, block: Type[Union[BasicBlock, Bottleneck]], plan...
method _forward_impl (line 233) | def _forward_impl(self, x: Tensor) -> Tensor:
method forward (line 248) | def forward(self, x: Tensor) -> Tensor:
function _resnet (line 252) | def _resnet(
class AttnBasicBlock (line 270) | class AttnBasicBlock(nn.Module):
method __init__ (line 273) | def __init__(
method forward (line 304) | def forward(self, x: Tensor) -> Tensor:
class AttnBottleneck (line 325) | class AttnBottleneck(nn.Module):
method __init__ (line 329) | def __init__(
method forward (line 361) | def forward(self, x: Tensor) -> Tensor:
method __init__ (line 387) | def __init__(self,
method _make_layer (line 423) | def _make_layer(self, block: Type[Union[BasicBlock, Bottleneck]], plan...
method _forward_impl (line 448) | def _forward_impl(self, x: Tensor) -> Tensor:
method forward (line 461) | def forward(self, x: Tensor) -> Tensor:
function resnet18 (line 465) | def resnet18(pretrained: bool = False, progress: bool = True,**kwargs: A...
function resnet34 (line 476) | def resnet34(pretrained: bool = False, progress: bool = True, **kwargs: ...
function resnet50 (line 487) | def resnet50(pretrained: bool = False, progress: bool = True, **kwargs: ...
function resnet101 (line 498) | def resnet101(pretrained: bool = False, progress: bool = True, **kwargs:...
function resnet152 (line 509) | def resnet152(pretrained: bool = False, progress: bool = True, **kwargs:...
function resnext50_32x4d (line 520) | def resnext50_32x4d(pretrained: bool = False, progress: bool = True, **k...
function resnext101_32x8d (line 533) | def resnext101_32x8d(pretrained: bool = False, progress: bool = True, **...
function wide_resnet50_2 (line 546) | def wide_resnet50_2(pretrained: bool = False, progress: bool = True, **k...
function wide_resnet101_2 (line 562) | def wide_resnet101_2(pretrained: bool = False, progress: bool = True, **...
FILE: train.py
function trainer (line 10) | def trainer(model, category, config):
FILE: unet.py
class TimestepBlock (line 11) | class TimestepBlock(nn.Module):
method forward (line 17) | def forward(self, x, emb):
class TimestepEmbedSequential (line 23) | class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
method forward (line 29) | def forward(self, x, emb):
class PositionalEmbedding (line 38) | class PositionalEmbedding(nn.Module):
method __init__ (line 44) | def __init__(self, dim, scale=1):
method forward (line 50) | def forward(self, x):
class Downsample (line 60) | class Downsample(nn.Module):
method __init__ (line 61) | def __init__(self, in_channels, use_conv, out_channels=None):
method forward (line 72) | def forward(self, x, time_embed=None):
class Upsample (line 77) | class Upsample(nn.Module):
method __init__ (line 78) | def __init__(self, in_channels, use_conv, out_channels=None):
method forward (line 87) | def forward(self, x, time_embed=None):
class AttentionBlock (line 95) | class AttentionBlock(nn.Module):
method __init__ (line 102) | def __init__(self, in_channels, n_heads=1, n_head_channels=-1):
method forward (line 119) | def forward(self, x, time=None):
class QKVAttention (line 128) | class QKVAttention(nn.Module):
method __init__ (line 133) | def __init__(self, n_heads):
method forward (line 137) | def forward(self, qkv, time=None):
class ResBlock (line 156) | class ResBlock(TimestepBlock):
method __init__ (line 157) | def __init__(
method forward (line 202) | def forward(self, x, time_embed):
class UNetModel (line 220) | class UNetModel(nn.Module):
method __init__ (line 222) | def __init__(
method forward (line 390) | def forward(self, x, time):
class GroupNorm32 (line 409) | class GroupNorm32(nn.GroupNorm):
method forward (line 410) | def forward(self, x):
function zero_module (line 414) | def zero_module(module):
function update_ema_params (line 423) | def update_ema_params(target, source, decay_rate=0.9999):
FILE: visualize.py
function visualalize_reconstruction (line 9) | def visualalize_reconstruction(input, recon, target):
function visualize (line 79) | def visualize(image, noisy_image, GT, pred_mask, anomaly_map, category) :
function show_tensor_image (line 116) | def show_tensor_image(image):
function show_tensor_mask (line 129) | def show_tensor_mask(image):
Condensed preview — 16 files, each showing path, character count, and a content snippet. Download the .json file or copy for the full structured content (82K chars).
[
{
"path": "LICENSE",
"chars": 1072,
"preview": "MIT License\n\nCopyright (c) 2023 Arian Mousakhan\n\nPermission is hereby granted, free of charge, to any person obtaining a"
},
{
"path": "README.md",
"chars": 5768,
"preview": "# Anomaly Detection with Conditioned Denoising Diffusion Models.\n\nOfficial implementation of [DDAD](https://arxiv.org/ab"
},
{
"path": "anomaly_map.py",
"chars": 3653,
"preview": "import torch\nimport torch.nn.functional as F\nfrom kornia.filters import gaussian_blur2d\nfrom torchvision.transforms impo"
},
{
"path": "config.yaml",
"chars": 2259,
"preview": "data :\n name: MVTec #MVTec #MTD #VisA \n data_dir: datasets/MVTec #MVTec #VisA #MTD \n category: screw #['carpet', "
},
{
"path": "dataset.py",
"chars": 3224,
"preview": "import os\nfrom glob import glob\nfrom pathlib import Path\nimport shutil\nimport numpy as np\nimport csv\nimport torch\nimport"
},
{
"path": "ddad.py",
"chars": 3151,
"preview": "from asyncio import constants\nfrom typing import Any\nimport torch\nfrom unet import *\nfrom dataset import *\nfrom visualiz"
},
{
"path": "feature_extractor.py",
"chars": 4456,
"preview": "import logging\nimport torch\nfrom dataset import *\nfrom dataset import *\nfrom unet import *\nfrom visualize import *\nfrom "
},
{
"path": "loss.py",
"chars": 586,
"preview": "import torch\nimport torch.nn as nn\nimport numpy as np\n\n\ndef get_loss(model, x_0, t, config):\n x_0 = x_0.to(config.mod"
},
{
"path": "main.py",
"chars": 3638,
"preview": "import torch\nimport numpy as np\nimport os\nimport argparse\nfrom unet import *\nfrom omegaconf import OmegaConf\nfrom train "
},
{
"path": "metrics.py",
"chars": 4945,
"preview": "import torch\nfrom torchmetrics import ROC, AUROC, F1Score\nimport os\nfrom torchvision.transforms import transforms\nfrom s"
},
{
"path": "reconstruction.py",
"chars": 2601,
"preview": "from typing import Any\nimport torch\n# from forward_process import *\nimport numpy as np\nimport os\nos.environ['CUDA_VISIBL"
},
{
"path": "requirements.txt",
"chars": 253,
"preview": "\nkornia==0.6.12\nmatplotlib==3.7.1\nnumpy==1.24.3\nomegaconf==2.1.2\nopencv-python-headless==4.5.5.64\npandas==2.0.1\nPillow=="
},
{
"path": "resnet.py",
"chars": 22419,
"preview": "import torch\nfrom torch import Tensor\nimport torch.nn as nn\ntry:\n from torch.hub import load_state_dict_from_url\nexce"
},
{
"path": "train.py",
"chars": 2014,
"preview": "import torch\nimport os\nimport torch.nn as nn\nfrom dataset import *\n\nfrom dataset import *\nfrom loss import *\n\n\ndef train"
},
{
"path": "unet.py",
"chars": 14715,
"preview": "# https://github.com/openai/guided-diffusion/tree/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924\nimport math\nfrom abc import a"
},
{
"path": "visualize.py",
"chars": 4242,
"preview": "import matplotlib.pyplot as plt\nfrom torchvision.transforms import transforms\nimport numpy as np\nimport torch\nimport os\n"
}
]
About this extraction
This page contains the full source code of the arimousa/DDAD GitHub repository, extracted and formatted as plain text for AI agents and large language models (LLMs). The extraction includes 16 files (77.1 KB), approximately 20.2k tokens, and a symbol index with 96 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.