ZhengPeng7 commited on
Commit
f648b6e
1 Parent(s): 4ed6b0d

Initial uploading of BiRefNet model for COD.

Browse files
Files changed (5) hide show
  1. BiRefNet_config.py +11 -0
  2. README.md +58 -3
  3. birefnet.py +2244 -0
  4. config.json +20 -0
  5. model.safetensors +3 -0
BiRefNet_config.py ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from transformers import PretrainedConfig
2
+
3
+ class BiRefNetConfig(PretrainedConfig):
4
+ model_type = "SegformerForSemanticSegmentation"
5
+ def __init__(
6
+ self,
7
+ bb_pretrained=False,
8
+ **kwargs
9
+ ):
10
+ self.bb_pretrained = bb_pretrained
11
+ super().__init__(**kwargs)
README.md CHANGED
@@ -1,3 +1,58 @@
1
- ---
2
- license: mit
3
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: BiRefNet
3
+ tags:
4
+ - background-removal
5
+ - mask-generation
6
+ - Dichotomous Image Segmentation
7
+ - Camouflaged Object Detection
8
+ - pytorch_model_hub_mixin
9
+ - model_hub_mixin
10
+ repo_url: https://github.com/ZhengPeng7/BiRefNet
11
+ pipeline_tag: image-segmentation
12
+ ---
13
+ <h1 align="center">Bilateral Reference for High-Resolution Dichotomous Image Segmentation</h1>
14
+
15
+ <div align='center'>
16
+ <a href='https://scholar.google.com/citations?user=TZRzWOsAAAAJ' target='_blank'><strong>Peng Zheng</strong></a><sup> 1,4,5,6</sup>,&thinsp;
17
+ <a href='https://scholar.google.com/citations?user=0uPb8MMAAAAJ' target='_blank'><strong>Dehong Gao</strong></a><sup> 2</sup>,&thinsp;
18
+ <a href='https://scholar.google.com/citations?user=kakwJ5QAAAAJ' target='_blank'><strong>Deng-Ping Fan</strong></a><sup> 1*</sup>,&thinsp;
19
+ <a href='https://scholar.google.com/citations?user=9cMQrVsAAAAJ' target='_blank'><strong>Li Liu</strong></a><sup> 3</sup>,&thinsp;
20
+ <a href='https://scholar.google.com/citations?user=qQP6WXIAAAAJ' target='_blank'><strong>Jorma Laaksonen</strong></a><sup> 4</sup>,&thinsp;
21
+ <a href='https://scholar.google.com/citations?user=pw_0Z_UAAAAJ' target='_blank'><strong>Wanli Ouyang</strong></a><sup> 5</sup>,&thinsp;
22
+ <a href='https://scholar.google.com/citations?user=stFCYOAAAAAJ' target='_blank'><strong>Nicu Sebe</strong></a><sup> 6</sup>
23
+ </div>
24
+
25
+ <div align='center'>
26
+ <sup>1 </sup>Nankai University&ensp; <sup>2 </sup>Northwestern Polytechnical University&ensp; <sup>3 </sup>National University of Defense Technology&ensp; <sup>4 </sup>Aalto University&ensp; <sup>5 </sup>Shanghai AI Laboratory&ensp; <sup>6 </sup>University of Trento&ensp;
27
+ </div>
28
+
29
+ <div align="center" style="display: flex; justify-content: center; flex-wrap: wrap;">
30
+ <a href='https://arxiv.org/pdf/2401.03407'><img src='https://img.shields.io/badge/arXiv-BiRefNet-red'></a>&ensp;
31
+ <a href='https://drive.google.com/file/d/1aBnJ_R9lbnC2dm8dqD0-pzP2Cu-U1Xpt/view?usp=drive_link'><img src='https://img.shields.io/badge/中文版-BiRefNet-red'></a>&ensp;
32
+ <a href='https://www.birefnet.top'><img src='https://img.shields.io/badge/Page-BiRefNet-red'></a>&ensp;
33
+ <a href='https://drive.google.com/drive/folders/1s2Xe0cjq-2ctnJBR24563yMSCOu4CcxM'><img src='https://img.shields.io/badge/Drive-Stuff-green'></a>&ensp;
34
+ <a href='LICENSE'><img src='https://img.shields.io/badge/License-MIT-yellow'></a>&ensp;
35
+ <a href='https://huggingface.co/spaces/ZhengPeng7/BiRefNet_demo'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20HF%20Spaces-BiRefNet-blue'></a>&ensp;
36
+ <a href='https://huggingface.co/ZhengPeng7/BiRefNet'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20HF%20Models-BiRefNet-blue'></a>&ensp;
37
+ <a href='https://colab.research.google.com/drive/14Dqg7oeBkFEtchaHLNpig2BcdkZEogba?usp=drive_link'><img src='https://img.shields.io/badge/Single_Image_Inference-F9AB00?style=for-the-badge&logo=googlecolab&color=525252'></a>&ensp;
38
+ <a href='https://colab.research.google.com/drive/1MaEiBfJ4xIaZZn0DqKrhydHB8X97hNXl#scrollTo=DJ4meUYjia6S'><img src='https://img.shields.io/badge/Inference_&_Evaluation-F9AB00?style=for-the-badge&logo=googlecolab&color=525252'></a>&ensp;
39
+ </div>
40
+
41
+ ## This repo holds the official weights of BiRefNet trained for Camouflaged Object Detection (COD) in the original paper.
42
+
43
+ **Check the main BiRefNet model repo for how to use it:**
44
+ https://huggingface.co/ZhengPeng7/BiRefNet/blob/main/README.md
45
+
46
+ **Also check the GitHub repo of BiRefNet for all things you may want:**
47
+ https://github.com/ZhengPeng7/BiRefNet
48
+
49
+ ## Citation
50
+
51
+ ```
52
+ @article{BiRefNet,
53
+ title={Bilateral Reference for High-Resolution Dichotomous Image Segmentation},
54
+ author={Zheng, Peng and Gao, Dehong and Fan, Deng-Ping and Liu, Li and Laaksonen, Jorma and Ouyang, Wanli and Sebe, Nicu},
55
+ journal={CAAI Artificial Intelligence Research},
56
+ year={2024}
57
+ }
58
+ ```
birefnet.py ADDED
@@ -0,0 +1,2244 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ### config.py
2
+
3
+ import os
4
+ import math
5
+
6
+
7
+ class Config():
8
+ def __init__(self) -> None:
9
+ # PATH settings
10
+ self.sys_home_dir = os.environ['HOME'] # Make up your file system as: SYS_HOME_DIR/codes/dis/BiRefNet, SYS_HOME_DIR/datasets/dis/xx, SYS_HOME_DIR/weights/xx
11
+
12
+ # TASK settings
13
+ self.task = ['DIS5K', 'COD', 'HRSOD', 'DIS5K+HRSOD+HRS10K', 'P3M-10k'][0]
14
+ self.training_set = {
15
+ 'DIS5K': ['DIS-TR', 'DIS-TR+DIS-TE1+DIS-TE2+DIS-TE3+DIS-TE4'][0],
16
+ 'COD': 'TR-COD10K+TR-CAMO',
17
+ 'HRSOD': ['TR-DUTS', 'TR-HRSOD', 'TR-UHRSD', 'TR-DUTS+TR-HRSOD', 'TR-DUTS+TR-UHRSD', 'TR-HRSOD+TR-UHRSD', 'TR-DUTS+TR-HRSOD+TR-UHRSD'][5],
18
+ 'DIS5K+HRSOD+HRS10K': 'DIS-TE1+DIS-TE2+DIS-TE3+DIS-TE4+DIS-TR+TE-HRS10K+TE-HRSOD+TE-UHRSD+TR-HRS10K+TR-HRSOD+TR-UHRSD', # leave DIS-VD for evaluation.
19
+ 'P3M-10k': 'TR-P3M-10k',
20
+ }[self.task]
21
+ self.prompt4loc = ['dense', 'sparse'][0]
22
+
23
+ # Faster-Training settings
24
+ self.load_all = True
25
+ self.compile = True # 1. Trigger CPU memory leak in some extend, which is an inherent problem of PyTorch.
26
+ # Machines with > 70GB CPU memory can run the whole training on DIS5K with default setting.
27
+ # 2. Higher PyTorch version may fix it: https://github.com/pytorch/pytorch/issues/119607.
28
+ # 3. But compile in Pytorch > 2.0.1 seems to bring no acceleration for training.
29
+ self.precisionHigh = True
30
+
31
+ # MODEL settings
32
+ self.ms_supervision = True
33
+ self.out_ref = self.ms_supervision and True
34
+ self.dec_ipt = True
35
+ self.dec_ipt_split = True
36
+ self.cxt_num = [0, 3][1] # multi-scale skip connections from encoder
37
+ self.mul_scl_ipt = ['', 'add', 'cat'][2]
38
+ self.dec_att = ['', 'ASPP', 'ASPPDeformable'][2]
39
+ self.squeeze_block = ['', 'BasicDecBlk_x1', 'ResBlk_x4', 'ASPP_x3', 'ASPPDeformable_x3'][1]
40
+ self.dec_blk = ['BasicDecBlk', 'ResBlk', 'HierarAttDecBlk'][0]
41
+
42
+ # TRAINING settings
43
+ self.batch_size = 4
44
+ self.IoU_finetune_last_epochs = [
45
+ 0,
46
+ {
47
+ 'DIS5K': -50,
48
+ 'COD': -20,
49
+ 'HRSOD': -20,
50
+ 'DIS5K+HRSOD+HRS10K': -20,
51
+ 'P3M-10k': -20,
52
+ }[self.task]
53
+ ][1] # choose 0 to skip
54
+ self.lr = (1e-4 if 'DIS5K' in self.task else 1e-5) * math.sqrt(self.batch_size / 4) # DIS needs high lr to converge faster. Adapt the lr linearly
55
+ self.size = 1024
56
+ self.num_workers = max(4, self.batch_size) # will be decrease to min(it, batch_size) at the initialization of the data_loader
57
+
58
+ # Backbone settings
59
+ self.bb = [
60
+ 'vgg16', 'vgg16bn', 'resnet50', # 0, 1, 2
61
+ 'swin_v1_t', 'swin_v1_s', # 3, 4
62
+ 'swin_v1_b', 'swin_v1_l', # 5-bs9, 6-bs4
63
+ 'pvt_v2_b0', 'pvt_v2_b1', # 7, 8
64
+ 'pvt_v2_b2', 'pvt_v2_b5', # 9-bs10, 10-bs5
65
+ ][6]
66
+ self.lateral_channels_in_collection = {
67
+ 'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
68
+ 'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
69
+ 'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
70
+ 'swin_v1_t': [768, 384, 192, 96], 'swin_v1_s': [768, 384, 192, 96],
71
+ 'pvt_v2_b0': [256, 160, 64, 32], 'pvt_v2_b1': [512, 320, 128, 64],
72
+ }[self.bb]
73
+ if self.mul_scl_ipt == 'cat':
74
+ self.lateral_channels_in_collection = [channel * 2 for channel in self.lateral_channels_in_collection]
75
+ self.cxt = self.lateral_channels_in_collection[1:][::-1][-self.cxt_num:] if self.cxt_num else []
76
+
77
+ # MODEL settings - inactive
78
+ self.lat_blk = ['BasicLatBlk'][0]
79
+ self.dec_channels_inter = ['fixed', 'adap'][0]
80
+ self.refine = ['', 'itself', 'RefUNet', 'Refiner', 'RefinerPVTInChannels4'][0]
81
+ self.progressive_ref = self.refine and True
82
+ self.ender = self.progressive_ref and False
83
+ self.scale = self.progressive_ref and 2
84
+ self.auxiliary_classification = False # Only for DIS5K, where class labels are saved in `dataset.py`.
85
+ self.refine_iteration = 1
86
+ self.freeze_bb = False
87
+ self.model = [
88
+ 'BiRefNet',
89
+ ][0]
90
+ if self.dec_blk == 'HierarAttDecBlk':
91
+ self.batch_size = 2 ** [0, 1, 2, 3, 4][2]
92
+
93
+ # TRAINING settings - inactive
94
+ self.preproc_methods = ['flip', 'enhance', 'rotate', 'pepper', 'crop'][:4]
95
+ self.optimizer = ['Adam', 'AdamW'][1]
96
+ self.lr_decay_epochs = [1e5] # Set to negative N to decay the lr in the last N-th epoch.
97
+ self.lr_decay_rate = 0.5
98
+ # Loss
99
+ self.lambdas_pix_last = {
100
+ # not 0 means opening this loss
101
+ # original rate -- 1 : 30 : 1.5 : 0.2, bce x 30
102
+ 'bce': 30 * 1, # high performance
103
+ 'iou': 0.5 * 1, # 0 / 255
104
+ 'iou_patch': 0.5 * 0, # 0 / 255, win_size = (64, 64)
105
+ 'mse': 150 * 0, # can smooth the saliency map
106
+ 'triplet': 3 * 0,
107
+ 'reg': 100 * 0,
108
+ 'ssim': 10 * 1, # help contours,
109
+ 'cnt': 5 * 0, # help contours
110
+ 'structure': 5 * 0, # structure loss from codes of MVANet. A little improvement on DIS-TE[1,2,3], a bit more decrease on DIS-TE4.
111
+ }
112
+ self.lambdas_cls = {
113
+ 'ce': 5.0
114
+ }
115
+ # Adv
116
+ self.lambda_adv_g = 10. * 0 # turn to 0 to avoid adv training
117
+ self.lambda_adv_d = 3. * (self.lambda_adv_g > 0)
118
+
119
+ # PATH settings - inactive
120
+ self.data_root_dir = os.path.join(self.sys_home_dir, 'datasets/dis')
121
+ self.weights_root_dir = os.path.join(self.sys_home_dir, 'weights')
122
+ self.weights = {
123
+ 'pvt_v2_b2': os.path.join(self.weights_root_dir, 'pvt_v2_b2.pth'),
124
+ 'pvt_v2_b5': os.path.join(self.weights_root_dir, ['pvt_v2_b5.pth', 'pvt_v2_b5_22k.pth'][0]),
125
+ 'swin_v1_b': os.path.join(self.weights_root_dir, ['swin_base_patch4_window12_384_22kto1k.pth', 'swin_base_patch4_window12_384_22k.pth'][0]),
126
+ 'swin_v1_l': os.path.join(self.weights_root_dir, ['swin_large_patch4_window12_384_22kto1k.pth', 'swin_large_patch4_window12_384_22k.pth'][0]),
127
+ 'swin_v1_t': os.path.join(self.weights_root_dir, ['swin_tiny_patch4_window7_224_22kto1k_finetune.pth'][0]),
128
+ 'swin_v1_s': os.path.join(self.weights_root_dir, ['swin_small_patch4_window7_224_22kto1k_finetune.pth'][0]),
129
+ 'pvt_v2_b0': os.path.join(self.weights_root_dir, ['pvt_v2_b0.pth'][0]),
130
+ 'pvt_v2_b1': os.path.join(self.weights_root_dir, ['pvt_v2_b1.pth'][0]),
131
+ }
132
+
133
+ # Callbacks - inactive
134
+ self.verbose_eval = True
135
+ self.only_S_MAE = False
136
+ self.use_fp16 = False # Bugs. It may cause nan in training.
137
+ self.SDPA_enabled = False # Bugs. Slower and errors occur in multi-GPUs
138
+
139
+ # others
140
+ self.device = [0, 'cpu'][0] # .to(0) == .to('cuda:0')
141
+
142
+ self.batch_size_valid = 1
143
+ self.rand_seed = 7
144
+ # run_sh_file = [f for f in os.listdir('.') if 'train.sh' == f] + [os.path.join('..', f) for f in os.listdir('..') if 'train.sh' == f]
145
+ # with open(run_sh_file[0], 'r') as f:
146
+ # lines = f.readlines()
147
+ # self.save_last = int([l.strip() for l in lines if '"{}")'.format(self.task) in l and 'val_last=' in l][0].split('val_last=')[-1].split()[0])
148
+ # self.save_step = int([l.strip() for l in lines if '"{}")'.format(self.task) in l and 'step=' in l][0].split('step=')[-1].split()[0])
149
+ # self.val_step = [0, self.save_step][0]
150
+
151
+ def print_task(self) -> None:
152
+ # Return task for choosing settings in shell scripts.
153
+ print(self.task)
154
+
155
+
156
+
157
+ ### models/backbones/pvt_v2.py
158
+
159
+ import torch
160
+ import torch.nn as nn
161
+ from functools import partial
162
+
163
+ from timm.models.layers import DropPath, to_2tuple, trunc_normal_
164
+ from timm.models.registry import register_model
165
+
166
+ import math
167
+
168
+ # from config import Config
169
+
170
+ # config = Config()
171
+
172
+ class Mlp(nn.Module):
173
+ def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
174
+ super().__init__()
175
+ out_features = out_features or in_features
176
+ hidden_features = hidden_features or in_features
177
+ self.fc1 = nn.Linear(in_features, hidden_features)
178
+ self.dwconv = DWConv(hidden_features)
179
+ self.act = act_layer()
180
+ self.fc2 = nn.Linear(hidden_features, out_features)
181
+ self.drop = nn.Dropout(drop)
182
+
183
+ self.apply(self._init_weights)
184
+
185
+ def _init_weights(self, m):
186
+ if isinstance(m, nn.Linear):
187
+ trunc_normal_(m.weight, std=.02)
188
+ if isinstance(m, nn.Linear) and m.bias is not None:
189
+ nn.init.constant_(m.bias, 0)
190
+ elif isinstance(m, nn.LayerNorm):
191
+ nn.init.constant_(m.bias, 0)
192
+ nn.init.constant_(m.weight, 1.0)
193
+ elif isinstance(m, nn.Conv2d):
194
+ fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
195
+ fan_out //= m.groups
196
+ m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
197
+ if m.bias is not None:
198
+ m.bias.data.zero_()
199
+
200
+ def forward(self, x, H, W):
201
+ x = self.fc1(x)
202
+ x = self.dwconv(x, H, W)
203
+ x = self.act(x)
204
+ x = self.drop(x)
205
+ x = self.fc2(x)
206
+ x = self.drop(x)
207
+ return x
208
+
209
+
210
+ class Attention(nn.Module):
211
+ def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., sr_ratio=1):
212
+ super().__init__()
213
+ assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
214
+
215
+ self.dim = dim
216
+ self.num_heads = num_heads
217
+ head_dim = dim // num_heads
218
+ self.scale = qk_scale or head_dim ** -0.5
219
+
220
+ self.q = nn.Linear(dim, dim, bias=qkv_bias)
221
+ self.kv = nn.Linear(dim, dim * 2, bias=qkv_bias)
222
+ self.attn_drop_prob = attn_drop
223
+ self.attn_drop = nn.Dropout(attn_drop)
224
+ self.proj = nn.Linear(dim, dim)
225
+ self.proj_drop = nn.Dropout(proj_drop)
226
+
227
+ self.sr_ratio = sr_ratio
228
+ if sr_ratio > 1:
229
+ self.sr = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)
230
+ self.norm = nn.LayerNorm(dim)
231
+
232
+ self.apply(self._init_weights)
233
+
234
+ def _init_weights(self, m):
235
+ if isinstance(m, nn.Linear):
236
+ trunc_normal_(m.weight, std=.02)
237
+ if isinstance(m, nn.Linear) and m.bias is not None:
238
+ nn.init.constant_(m.bias, 0)
239
+ elif isinstance(m, nn.LayerNorm):
240
+ nn.init.constant_(m.bias, 0)
241
+ nn.init.constant_(m.weight, 1.0)
242
+ elif isinstance(m, nn.Conv2d):
243
+ fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
244
+ fan_out //= m.groups
245
+ m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
246
+ if m.bias is not None:
247
+ m.bias.data.zero_()
248
+
249
+ def forward(self, x, H, W):
250
+ B, N, C = x.shape
251
+ q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
252
+
253
+ if self.sr_ratio > 1:
254
+ x_ = x.permute(0, 2, 1).reshape(B, C, H, W)
255
+ x_ = self.sr(x_).reshape(B, C, -1).permute(0, 2, 1)
256
+ x_ = self.norm(x_)
257
+ kv = self.kv(x_).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
258
+ else:
259
+ kv = self.kv(x).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
260
+ k, v = kv[0], kv[1]
261
+
262
+ if config.SDPA_enabled:
263
+ x = torch.nn.functional.scaled_dot_product_attention(
264
+ q, k, v,
265
+ attn_mask=None, dropout_p=self.attn_drop_prob, is_causal=False
266
+ ).transpose(1, 2).reshape(B, N, C)
267
+ else:
268
+ attn = (q @ k.transpose(-2, -1)) * self.scale
269
+ attn = attn.softmax(dim=-1)
270
+ attn = self.attn_drop(attn)
271
+
272
+ x = (attn @ v).transpose(1, 2).reshape(B, N, C)
273
+ x = self.proj(x)
274
+ x = self.proj_drop(x)
275
+
276
+ return x
277
+
278
+
279
+ class Block(nn.Module):
280
+
281
+ def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
282
+ drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, sr_ratio=1):
283
+ super().__init__()
284
+ self.norm1 = norm_layer(dim)
285
+ self.attn = Attention(
286
+ dim,
287
+ num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
288
+ attn_drop=attn_drop, proj_drop=drop, sr_ratio=sr_ratio)
289
+ # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
290
+ self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
291
+ self.norm2 = norm_layer(dim)
292
+ mlp_hidden_dim = int(dim * mlp_ratio)
293
+ self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
294
+
295
+ self.apply(self._init_weights)
296
+
297
+ def _init_weights(self, m):
298
+ if isinstance(m, nn.Linear):
299
+ trunc_normal_(m.weight, std=.02)
300
+ if isinstance(m, nn.Linear) and m.bias is not None:
301
+ nn.init.constant_(m.bias, 0)
302
+ elif isinstance(m, nn.LayerNorm):
303
+ nn.init.constant_(m.bias, 0)
304
+ nn.init.constant_(m.weight, 1.0)
305
+ elif isinstance(m, nn.Conv2d):
306
+ fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
307
+ fan_out //= m.groups
308
+ m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
309
+ if m.bias is not None:
310
+ m.bias.data.zero_()
311
+
312
+ def forward(self, x, H, W):
313
+ x = x + self.drop_path(self.attn(self.norm1(x), H, W))
314
+ x = x + self.drop_path(self.mlp(self.norm2(x), H, W))
315
+
316
+ return x
317
+
318
+
319
+ class OverlapPatchEmbed(nn.Module):
320
+ """ Image to Patch Embedding
321
+ """
322
+
323
+ def __init__(self, img_size=224, patch_size=7, stride=4, in_channels=3, embed_dim=768):
324
+ super().__init__()
325
+ img_size = to_2tuple(img_size)
326
+ patch_size = to_2tuple(patch_size)
327
+
328
+ self.img_size = img_size
329
+ self.patch_size = patch_size
330
+ self.H, self.W = img_size[0] // patch_size[0], img_size[1] // patch_size[1]
331
+ self.num_patches = self.H * self.W
332
+ self.proj = nn.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=stride,
333
+ padding=(patch_size[0] // 2, patch_size[1] // 2))
334
+ self.norm = nn.LayerNorm(embed_dim)
335
+
336
+ self.apply(self._init_weights)
337
+
338
+ def _init_weights(self, m):
339
+ if isinstance(m, nn.Linear):
340
+ trunc_normal_(m.weight, std=.02)
341
+ if isinstance(m, nn.Linear) and m.bias is not None:
342
+ nn.init.constant_(m.bias, 0)
343
+ elif isinstance(m, nn.LayerNorm):
344
+ nn.init.constant_(m.bias, 0)
345
+ nn.init.constant_(m.weight, 1.0)
346
+ elif isinstance(m, nn.Conv2d):
347
+ fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
348
+ fan_out //= m.groups
349
+ m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
350
+ if m.bias is not None:
351
+ m.bias.data.zero_()
352
+
353
+ def forward(self, x):
354
+ x = self.proj(x)
355
+ _, _, H, W = x.shape
356
+ x = x.flatten(2).transpose(1, 2)
357
+ x = self.norm(x)
358
+
359
+ return x, H, W
360
+
361
+
362
+ class PyramidVisionTransformerImpr(nn.Module):
363
+ def __init__(self, img_size=224, patch_size=16, in_channels=3, num_classes=1000, embed_dims=[64, 128, 256, 512],
364
+ num_heads=[1, 2, 4, 8], mlp_ratios=[4, 4, 4, 4], qkv_bias=False, qk_scale=None, drop_rate=0.,
365
+ attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm,
366
+ depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1]):
367
+ super().__init__()
368
+ self.num_classes = num_classes
369
+ self.depths = depths
370
+
371
+ # patch_embed
372
+ self.patch_embed1 = OverlapPatchEmbed(img_size=img_size, patch_size=7, stride=4, in_channels=in_channels,
373
+ embed_dim=embed_dims[0])
374
+ self.patch_embed2 = OverlapPatchEmbed(img_size=img_size // 4, patch_size=3, stride=2, in_channels=embed_dims[0],
375
+ embed_dim=embed_dims[1])
376
+ self.patch_embed3 = OverlapPatchEmbed(img_size=img_size // 8, patch_size=3, stride=2, in_channels=embed_dims[1],
377
+ embed_dim=embed_dims[2])
378
+ self.patch_embed4 = OverlapPatchEmbed(img_size=img_size // 16, patch_size=3, stride=2, in_channels=embed_dims[2],
379
+ embed_dim=embed_dims[3])
380
+
381
+ # transformer encoder
382
+ dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule
383
+ cur = 0
384
+ self.block1 = nn.ModuleList([Block(
385
+ dim=embed_dims[0], num_heads=num_heads[0], mlp_ratio=mlp_ratios[0], qkv_bias=qkv_bias, qk_scale=qk_scale,
386
+ drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
387
+ sr_ratio=sr_ratios[0])
388
+ for i in range(depths[0])])
389
+ self.norm1 = norm_layer(embed_dims[0])
390
+
391
+ cur += depths[0]
392
+ self.block2 = nn.ModuleList([Block(
393
+ dim=embed_dims[1], num_heads=num_heads[1], mlp_ratio=mlp_ratios[1], qkv_bias=qkv_bias, qk_scale=qk_scale,
394
+ drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
395
+ sr_ratio=sr_ratios[1])
396
+ for i in range(depths[1])])
397
+ self.norm2 = norm_layer(embed_dims[1])
398
+
399
+ cur += depths[1]
400
+ self.block3 = nn.ModuleList([Block(
401
+ dim=embed_dims[2], num_heads=num_heads[2], mlp_ratio=mlp_ratios[2], qkv_bias=qkv_bias, qk_scale=qk_scale,
402
+ drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
403
+ sr_ratio=sr_ratios[2])
404
+ for i in range(depths[2])])
405
+ self.norm3 = norm_layer(embed_dims[2])
406
+
407
+ cur += depths[2]
408
+ self.block4 = nn.ModuleList([Block(
409
+ dim=embed_dims[3], num_heads=num_heads[3], mlp_ratio=mlp_ratios[3], qkv_bias=qkv_bias, qk_scale=qk_scale,
410
+ drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
411
+ sr_ratio=sr_ratios[3])
412
+ for i in range(depths[3])])
413
+ self.norm4 = norm_layer(embed_dims[3])
414
+
415
+ # classification head
416
+ # self.head = nn.Linear(embed_dims[3], num_classes) if num_classes > 0 else nn.Identity()
417
+
418
+ self.apply(self._init_weights)
419
+
420
+ def _init_weights(self, m):
421
+ if isinstance(m, nn.Linear):
422
+ trunc_normal_(m.weight, std=.02)
423
+ if isinstance(m, nn.Linear) and m.bias is not None:
424
+ nn.init.constant_(m.bias, 0)
425
+ elif isinstance(m, nn.LayerNorm):
426
+ nn.init.constant_(m.bias, 0)
427
+ nn.init.constant_(m.weight, 1.0)
428
+ elif isinstance(m, nn.Conv2d):
429
+ fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
430
+ fan_out //= m.groups
431
+ m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
432
+ if m.bias is not None:
433
+ m.bias.data.zero_()
434
+
435
+ def init_weights(self, pretrained=None):
436
+ if isinstance(pretrained, str):
437
+ logger = 1
438
+ #load_checkpoint(self, pretrained, map_location='cpu', strict=False, logger=logger)
439
+
440
+ def reset_drop_path(self, drop_path_rate):
441
+ dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(self.depths))]
442
+ cur = 0
443
+ for i in range(self.depths[0]):
444
+ self.block1[i].drop_path.drop_prob = dpr[cur + i]
445
+
446
+ cur += self.depths[0]
447
+ for i in range(self.depths[1]):
448
+ self.block2[i].drop_path.drop_prob = dpr[cur + i]
449
+
450
+ cur += self.depths[1]
451
+ for i in range(self.depths[2]):
452
+ self.block3[i].drop_path.drop_prob = dpr[cur + i]
453
+
454
+ cur += self.depths[2]
455
+ for i in range(self.depths[3]):
456
+ self.block4[i].drop_path.drop_prob = dpr[cur + i]
457
+
458
+ def freeze_patch_emb(self):
459
+ self.patch_embed1.requires_grad = False
460
+
461
+ @torch.jit.ignore
462
+ def no_weight_decay(self):
463
+ return {'pos_embed1', 'pos_embed2', 'pos_embed3', 'pos_embed4', 'cls_token'} # has pos_embed may be better
464
+
465
+ def get_classifier(self):
466
+ return self.head
467
+
468
+ def reset_classifier(self, num_classes, global_pool=''):
469
+ self.num_classes = num_classes
470
+ self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
471
+
472
+ def forward_features(self, x):
473
+ B = x.shape[0]
474
+ outs = []
475
+
476
+ # stage 1
477
+ x, H, W = self.patch_embed1(x)
478
+ for i, blk in enumerate(self.block1):
479
+ x = blk(x, H, W)
480
+ x = self.norm1(x)
481
+ x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
482
+ outs.append(x)
483
+
484
+ # stage 2
485
+ x, H, W = self.patch_embed2(x)
486
+ for i, blk in enumerate(self.block2):
487
+ x = blk(x, H, W)
488
+ x = self.norm2(x)
489
+ x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
490
+ outs.append(x)
491
+
492
+ # stage 3
493
+ x, H, W = self.patch_embed3(x)
494
+ for i, blk in enumerate(self.block3):
495
+ x = blk(x, H, W)
496
+ x = self.norm3(x)
497
+ x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
498
+ outs.append(x)
499
+
500
+ # stage 4
501
+ x, H, W = self.patch_embed4(x)
502
+ for i, blk in enumerate(self.block4):
503
+ x = blk(x, H, W)
504
+ x = self.norm4(x)
505
+ x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
506
+ outs.append(x)
507
+
508
+ return outs
509
+
510
+ # return x.mean(dim=1)
511
+
512
+ def forward(self, x):
513
+ x = self.forward_features(x)
514
+ # x = self.head(x)
515
+
516
+ return x
517
+
518
+
519
+ class DWConv(nn.Module):
520
+ def __init__(self, dim=768):
521
+ super(DWConv, self).__init__()
522
+ self.dwconv = nn.Conv2d(dim, dim, 3, 1, 1, bias=True, groups=dim)
523
+
524
+ def forward(self, x, H, W):
525
+ B, N, C = x.shape
526
+ x = x.transpose(1, 2).view(B, C, H, W).contiguous()
527
+ x = self.dwconv(x)
528
+ x = x.flatten(2).transpose(1, 2)
529
+
530
+ return x
531
+
532
+
533
+ def _conv_filter(state_dict, patch_size=16):
534
+ """ convert patch embedding weight from manual patchify + linear proj to conv"""
535
+ out_dict = {}
536
+ for k, v in state_dict.items():
537
+ if 'patch_embed.proj.weight' in k:
538
+ v = v.reshape((v.shape[0], 3, patch_size, patch_size))
539
+ out_dict[k] = v
540
+
541
+ return out_dict
542
+
543
+
544
+ ## @register_model
545
+ class pvt_v2_b0(PyramidVisionTransformerImpr):
546
+ def __init__(self, **kwargs):
547
+ super(pvt_v2_b0, self).__init__(
548
+ patch_size=4, embed_dims=[32, 64, 160, 256], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
549
+ qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1],
550
+ drop_rate=0.0, drop_path_rate=0.1)
551
+
552
+
553
+
554
+ ## @register_model
555
+ class pvt_v2_b1(PyramidVisionTransformerImpr):
556
+ def __init__(self, **kwargs):
557
+ super(pvt_v2_b1, self).__init__(
558
+ patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
559
+ qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1],
560
+ drop_rate=0.0, drop_path_rate=0.1)
561
+
562
+ ## @register_model
563
+ class pvt_v2_b2(PyramidVisionTransformerImpr):
564
+ def __init__(self, in_channels=3, **kwargs):
565
+ super(pvt_v2_b2, self).__init__(
566
+ patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
567
+ qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1],
568
+ drop_rate=0.0, drop_path_rate=0.1, in_channels=in_channels)
569
+
570
+ ## @register_model
571
+ class pvt_v2_b3(PyramidVisionTransformerImpr):
572
+ def __init__(self, **kwargs):
573
+ super(pvt_v2_b3, self).__init__(
574
+ patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
575
+ qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 18, 3], sr_ratios=[8, 4, 2, 1],
576
+ drop_rate=0.0, drop_path_rate=0.1)
577
+
578
+ ## @register_model
579
+ class pvt_v2_b4(PyramidVisionTransformerImpr):
580
+ def __init__(self, **kwargs):
581
+ super(pvt_v2_b4, self).__init__(
582
+ patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
583
+ qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 8, 27, 3], sr_ratios=[8, 4, 2, 1],
584
+ drop_rate=0.0, drop_path_rate=0.1)
585
+
586
+
587
+ ## @register_model
588
+ class pvt_v2_b5(PyramidVisionTransformerImpr):
589
+ def __init__(self, **kwargs):
590
+ super(pvt_v2_b5, self).__init__(
591
+ patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],
592
+ qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 6, 40, 3], sr_ratios=[8, 4, 2, 1],
593
+ drop_rate=0.0, drop_path_rate=0.1)
594
+
595
+
596
+
597
+ ### models/backbones/swin_v1.py
598
+
599
+ # --------------------------------------------------------
600
+ # Swin Transformer
601
+ # Copyright (c) 2021 Microsoft
602
+ # Licensed under The MIT License [see LICENSE for details]
603
+ # Written by Ze Liu, Yutong Lin, Yixuan Wei
604
+ # --------------------------------------------------------
605
+
606
+ import torch
607
+ import torch.nn as nn
608
+ import torch.nn.functional as F
609
+ import torch.utils.checkpoint as checkpoint
610
+ import numpy as np
611
+ from timm.models.layers import DropPath, to_2tuple, trunc_normal_
612
+
613
+ # from config import Config
614
+
615
+
616
+ # config = Config()
617
+
618
+ class Mlp(nn.Module):
619
+ """ Multilayer perceptron."""
620
+
621
+ def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
622
+ super().__init__()
623
+ out_features = out_features or in_features
624
+ hidden_features = hidden_features or in_features
625
+ self.fc1 = nn.Linear(in_features, hidden_features)
626
+ self.act = act_layer()
627
+ self.fc2 = nn.Linear(hidden_features, out_features)
628
+ self.drop = nn.Dropout(drop)
629
+
630
+ def forward(self, x):
631
+ x = self.fc1(x)
632
+ x = self.act(x)
633
+ x = self.drop(x)
634
+ x = self.fc2(x)
635
+ x = self.drop(x)
636
+ return x
637
+
638
+
639
+ def window_partition(x, window_size):
640
+ """
641
+ Args:
642
+ x: (B, H, W, C)
643
+ window_size (int): window size
644
+
645
+ Returns:
646
+ windows: (num_windows*B, window_size, window_size, C)
647
+ """
648
+ B, H, W, C = x.shape
649
+ x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
650
+ windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
651
+ return windows
652
+
653
+
654
+ def window_reverse(windows, window_size, H, W):
655
+ """
656
+ Args:
657
+ windows: (num_windows*B, window_size, window_size, C)
658
+ window_size (int): Window size
659
+ H (int): Height of image
660
+ W (int): Width of image
661
+
662
+ Returns:
663
+ x: (B, H, W, C)
664
+ """
665
+ B = int(windows.shape[0] / (H * W / window_size / window_size))
666
+ x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
667
+ x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
668
+ return x
669
+
670
+
671
+ class WindowAttention(nn.Module):
672
+ """ Window based multi-head self attention (W-MSA) module with relative position bias.
673
+ It supports both of shifted and non-shifted window.
674
+
675
+ Args:
676
+ dim (int): Number of input channels.
677
+ window_size (tuple[int]): The height and width of the window.
678
+ num_heads (int): Number of attention heads.
679
+ qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
680
+ qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
681
+ attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
682
+ proj_drop (float, optional): Dropout ratio of output. Default: 0.0
683
+ """
684
+
685
+ def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
686
+
687
+ super().__init__()
688
+ self.dim = dim
689
+ self.window_size = window_size # Wh, Ww
690
+ self.num_heads = num_heads
691
+ head_dim = dim // num_heads
692
+ self.scale = qk_scale or head_dim ** -0.5
693
+
694
+ # define a parameter table of relative position bias
695
+ self.relative_position_bias_table = nn.Parameter(
696
+ torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)) # 2*Wh-1 * 2*Ww-1, nH
697
+
698
+ # get pair-wise relative position index for each token inside the window
699
+ coords_h = torch.arange(self.window_size[0])
700
+ coords_w = torch.arange(self.window_size[1])
701
+ coords = torch.stack(torch.meshgrid([coords_h, coords_w], indexing='ij')) # 2, Wh, Ww
702
+ coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww
703
+ relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww
704
+ relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2
705
+ relative_coords[:, :, 0] += self.window_size[0] - 1 # shift to start from 0
706
+ relative_coords[:, :, 1] += self.window_size[1] - 1
707
+ relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
708
+ relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww
709
+ self.register_buffer("relative_position_index", relative_position_index)
710
+
711
+ self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
712
+ self.attn_drop_prob = attn_drop
713
+ self.attn_drop = nn.Dropout(attn_drop)
714
+ self.proj = nn.Linear(dim, dim)
715
+ self.proj_drop = nn.Dropout(proj_drop)
716
+
717
+ trunc_normal_(self.relative_position_bias_table, std=.02)
718
+ self.softmax = nn.Softmax(dim=-1)
719
+
720
+ def forward(self, x, mask=None):
721
+ """ Forward function.
722
+
723
+ Args:
724
+ x: input features with shape of (num_windows*B, N, C)
725
+ mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
726
+ """
727
+ B_, N, C = x.shape
728
+ qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
729
+ q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple)
730
+
731
+ q = q * self.scale
732
+
733
+ if config.SDPA_enabled:
734
+ x = torch.nn.functional.scaled_dot_product_attention(
735
+ q, k, v,
736
+ attn_mask=None, dropout_p=self.attn_drop_prob, is_causal=False
737
+ ).transpose(1, 2).reshape(B_, N, C)
738
+ else:
739
+ attn = (q @ k.transpose(-2, -1))
740
+
741
+ relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
742
+ self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1) # Wh*Ww,Wh*Ww,nH
743
+ relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww
744
+ attn = attn + relative_position_bias.unsqueeze(0)
745
+
746
+ if mask is not None:
747
+ nW = mask.shape[0]
748
+ attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
749
+ attn = attn.view(-1, self.num_heads, N, N)
750
+ attn = self.softmax(attn)
751
+ else:
752
+ attn = self.softmax(attn)
753
+
754
+ attn = self.attn_drop(attn)
755
+
756
+ x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
757
+ x = self.proj(x)
758
+ x = self.proj_drop(x)
759
+ return x
760
+
761
+
762
+ class SwinTransformerBlock(nn.Module):
763
+ """ Swin Transformer Block.
764
+
765
+ Args:
766
+ dim (int): Number of input channels.
767
+ num_heads (int): Number of attention heads.
768
+ window_size (int): Window size.
769
+ shift_size (int): Shift size for SW-MSA.
770
+ mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
771
+ qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
772
+ qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
773
+ drop (float, optional): Dropout rate. Default: 0.0
774
+ attn_drop (float, optional): Attention dropout rate. Default: 0.0
775
+ drop_path (float, optional): Stochastic depth rate. Default: 0.0
776
+ act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
777
+ norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
778
+ """
779
+
780
+ def __init__(self, dim, num_heads, window_size=7, shift_size=0,
781
+ mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
782
+ act_layer=nn.GELU, norm_layer=nn.LayerNorm):
783
+ super().__init__()
784
+ self.dim = dim
785
+ self.num_heads = num_heads
786
+ self.window_size = window_size
787
+ self.shift_size = shift_size
788
+ self.mlp_ratio = mlp_ratio
789
+ assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
790
+
791
+ self.norm1 = norm_layer(dim)
792
+ self.attn = WindowAttention(
793
+ dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
794
+ qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
795
+
796
+ self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
797
+ self.norm2 = norm_layer(dim)
798
+ mlp_hidden_dim = int(dim * mlp_ratio)
799
+ self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
800
+
801
+ self.H = None
802
+ self.W = None
803
+
804
+ def forward(self, x, mask_matrix):
805
+ """ Forward function.
806
+
807
+ Args:
808
+ x: Input feature, tensor size (B, H*W, C).
809
+ H, W: Spatial resolution of the input feature.
810
+ mask_matrix: Attention mask for cyclic shift.
811
+ """
812
+ B, L, C = x.shape
813
+ H, W = self.H, self.W
814
+ assert L == H * W, "input feature has wrong size"
815
+
816
+ shortcut = x
817
+ x = self.norm1(x)
818
+ x = x.view(B, H, W, C)
819
+
820
+ # pad feature maps to multiples of window size
821
+ pad_l = pad_t = 0
822
+ pad_r = (self.window_size - W % self.window_size) % self.window_size
823
+ pad_b = (self.window_size - H % self.window_size) % self.window_size
824
+ x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
825
+ _, Hp, Wp, _ = x.shape
826
+
827
+ # cyclic shift
828
+ if self.shift_size > 0:
829
+ shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
830
+ attn_mask = mask_matrix
831
+ else:
832
+ shifted_x = x
833
+ attn_mask = None
834
+
835
+ # partition windows
836
+ x_windows = window_partition(shifted_x, self.window_size) # nW*B, window_size, window_size, C
837
+ x_windows = x_windows.view(-1, self.window_size * self.window_size, C) # nW*B, window_size*window_size, C
838
+
839
+ # W-MSA/SW-MSA
840
+ attn_windows = self.attn(x_windows, mask=attn_mask) # nW*B, window_size*window_size, C
841
+
842
+ # merge windows
843
+ attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
844
+ shifted_x = window_reverse(attn_windows, self.window_size, Hp, Wp) # B H' W' C
845
+
846
+ # reverse cyclic shift
847
+ if self.shift_size > 0:
848
+ x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
849
+ else:
850
+ x = shifted_x
851
+
852
+ if pad_r > 0 or pad_b > 0:
853
+ x = x[:, :H, :W, :].contiguous()
854
+
855
+ x = x.view(B, H * W, C)
856
+
857
+ # FFN
858
+ x = shortcut + self.drop_path(x)
859
+ x = x + self.drop_path(self.mlp(self.norm2(x)))
860
+
861
+ return x
862
+
863
+
864
+ class PatchMerging(nn.Module):
865
+ """ Patch Merging Layer
866
+
867
+ Args:
868
+ dim (int): Number of input channels.
869
+ norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
870
+ """
871
+ def __init__(self, dim, norm_layer=nn.LayerNorm):
872
+ super().__init__()
873
+ self.dim = dim
874
+ self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
875
+ self.norm = norm_layer(4 * dim)
876
+
877
+ def forward(self, x, H, W):
878
+ """ Forward function.
879
+
880
+ Args:
881
+ x: Input feature, tensor size (B, H*W, C).
882
+ H, W: Spatial resolution of the input feature.
883
+ """
884
+ B, L, C = x.shape
885
+ assert L == H * W, "input feature has wrong size"
886
+
887
+ x = x.view(B, H, W, C)
888
+
889
+ # padding
890
+ pad_input = (H % 2 == 1) or (W % 2 == 1)
891
+ if pad_input:
892
+ x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2))
893
+
894
+ x0 = x[:, 0::2, 0::2, :] # B H/2 W/2 C
895
+ x1 = x[:, 1::2, 0::2, :] # B H/2 W/2 C
896
+ x2 = x[:, 0::2, 1::2, :] # B H/2 W/2 C
897
+ x3 = x[:, 1::2, 1::2, :] # B H/2 W/2 C
898
+ x = torch.cat([x0, x1, x2, x3], -1) # B H/2 W/2 4*C
899
+ x = x.view(B, -1, 4 * C) # B H/2*W/2 4*C
900
+
901
+ x = self.norm(x)
902
+ x = self.reduction(x)
903
+
904
+ return x
905
+
906
+
907
+ class BasicLayer(nn.Module):
908
+ """ A basic Swin Transformer layer for one stage.
909
+
910
+ Args:
911
+ dim (int): Number of feature channels
912
+ depth (int): Depths of this stage.
913
+ num_heads (int): Number of attention head.
914
+ window_size (int): Local window size. Default: 7.
915
+ mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
916
+ qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
917
+ qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
918
+ drop (float, optional): Dropout rate. Default: 0.0
919
+ attn_drop (float, optional): Attention dropout rate. Default: 0.0
920
+ drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
921
+ norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
922
+ downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
923
+ use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
924
+ """
925
+
926
+ def __init__(self,
927
+ dim,
928
+ depth,
929
+ num_heads,
930
+ window_size=7,
931
+ mlp_ratio=4.,
932
+ qkv_bias=True,
933
+ qk_scale=None,
934
+ drop=0.,
935
+ attn_drop=0.,
936
+ drop_path=0.,
937
+ norm_layer=nn.LayerNorm,
938
+ downsample=None,
939
+ use_checkpoint=False):
940
+ super().__init__()
941
+ self.window_size = window_size
942
+ self.shift_size = window_size // 2
943
+ self.depth = depth
944
+ self.use_checkpoint = use_checkpoint
945
+
946
+ # build blocks
947
+ self.blocks = nn.ModuleList([
948
+ SwinTransformerBlock(
949
+ dim=dim,
950
+ num_heads=num_heads,
951
+ window_size=window_size,
952
+ shift_size=0 if (i % 2 == 0) else window_size // 2,
953
+ mlp_ratio=mlp_ratio,
954
+ qkv_bias=qkv_bias,
955
+ qk_scale=qk_scale,
956
+ drop=drop,
957
+ attn_drop=attn_drop,
958
+ drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
959
+ norm_layer=norm_layer)
960
+ for i in range(depth)])
961
+
962
+ # patch merging layer
963
+ if downsample is not None:
964
+ self.downsample = downsample(dim=dim, norm_layer=norm_layer)
965
+ else:
966
+ self.downsample = None
967
+
968
+ def forward(self, x, H, W):
969
+ """ Forward function.
970
+
971
+ Args:
972
+ x: Input feature, tensor size (B, H*W, C).
973
+ H, W: Spatial resolution of the input feature.
974
+ """
975
+
976
+ # calculate attention mask for SW-MSA
977
+ Hp = int(np.ceil(H / self.window_size)) * self.window_size
978
+ Wp = int(np.ceil(W / self.window_size)) * self.window_size
979
+ img_mask = torch.zeros((1, Hp, Wp, 1), device=x.device) # 1 Hp Wp 1
980
+ h_slices = (slice(0, -self.window_size),
981
+ slice(-self.window_size, -self.shift_size),
982
+ slice(-self.shift_size, None))
983
+ w_slices = (slice(0, -self.window_size),
984
+ slice(-self.window_size, -self.shift_size),
985
+ slice(-self.shift_size, None))
986
+ cnt = 0
987
+ for h in h_slices:
988
+ for w in w_slices:
989
+ img_mask[:, h, w, :] = cnt
990
+ cnt += 1
991
+
992
+ mask_windows = window_partition(img_mask, self.window_size) # nW, window_size, window_size, 1
993
+ mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
994
+ attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
995
+ attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
996
+
997
+ for blk in self.blocks:
998
+ blk.H, blk.W = H, W
999
+ if self.use_checkpoint:
1000
+ x = checkpoint.checkpoint(blk, x, attn_mask)
1001
+ else:
1002
+ x = blk(x, attn_mask)
1003
+ if self.downsample is not None:
1004
+ x_down = self.downsample(x, H, W)
1005
+ Wh, Ww = (H + 1) // 2, (W + 1) // 2
1006
+ return x, H, W, x_down, Wh, Ww
1007
+ else:
1008
+ return x, H, W, x, H, W
1009
+
1010
+
1011
+ class PatchEmbed(nn.Module):
1012
+ """ Image to Patch Embedding
1013
+
1014
+ Args:
1015
+ patch_size (int): Patch token size. Default: 4.
1016
+ in_channels (int): Number of input image channels. Default: 3.
1017
+ embed_dim (int): Number of linear projection output channels. Default: 96.
1018
+ norm_layer (nn.Module, optional): Normalization layer. Default: None
1019
+ """
1020
+
1021
+ def __init__(self, patch_size=4, in_channels=3, embed_dim=96, norm_layer=None):
1022
+ super().__init__()
1023
+ patch_size = to_2tuple(patch_size)
1024
+ self.patch_size = patch_size
1025
+
1026
+ self.in_channels = in_channels
1027
+ self.embed_dim = embed_dim
1028
+
1029
+ self.proj = nn.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=patch_size)
1030
+ if norm_layer is not None:
1031
+ self.norm = norm_layer(embed_dim)
1032
+ else:
1033
+ self.norm = None
1034
+
1035
+ def forward(self, x):
1036
+ """Forward function."""
1037
+ # padding
1038
+ _, _, H, W = x.size()
1039
+ if W % self.patch_size[1] != 0:
1040
+ x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1]))
1041
+ if H % self.patch_size[0] != 0:
1042
+ x = F.pad(x, (0, 0, 0, self.patch_size[0] - H % self.patch_size[0]))
1043
+
1044
+ x = self.proj(x) # B C Wh Ww
1045
+ if self.norm is not None:
1046
+ Wh, Ww = x.size(2), x.size(3)
1047
+ x = x.flatten(2).transpose(1, 2)
1048
+ x = self.norm(x)
1049
+ x = x.transpose(1, 2).view(-1, self.embed_dim, Wh, Ww)
1050
+
1051
+ return x
1052
+
1053
+
1054
+ class SwinTransformer(nn.Module):
1055
+ """ Swin Transformer backbone.
1056
+ A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows` -
1057
+ https://arxiv.org/pdf/2103.14030
1058
+
1059
+ Args:
1060
+ pretrain_img_size (int): Input image size for training the pretrained model,
1061
+ used in absolute postion embedding. Default 224.
1062
+ patch_size (int | tuple(int)): Patch size. Default: 4.
1063
+ in_channels (int): Number of input image channels. Default: 3.
1064
+ embed_dim (int): Number of linear projection output channels. Default: 96.
1065
+ depths (tuple[int]): Depths of each Swin Transformer stage.
1066
+ num_heads (tuple[int]): Number of attention head of each stage.
1067
+ window_size (int): Window size. Default: 7.
1068
+ mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
1069
+ qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
1070
+ qk_scale (float): Override default qk scale of head_dim ** -0.5 if set.
1071
+ drop_rate (float): Dropout rate.
1072
+ attn_drop_rate (float): Attention dropout rate. Default: 0.
1073
+ drop_path_rate (float): Stochastic depth rate. Default: 0.2.
1074
+ norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
1075
+ ape (bool): If True, add absolute position embedding to the patch embedding. Default: False.
1076
+ patch_norm (bool): If True, add normalization after patch embedding. Default: True.
1077
+ out_indices (Sequence[int]): Output from which stages.
1078
+ frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
1079
+ -1 means not freezing any parameters.
1080
+ use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
1081
+ """
1082
+
1083
+ def __init__(self,
1084
+ pretrain_img_size=224,
1085
+ patch_size=4,
1086
+ in_channels=3,
1087
+ embed_dim=96,
1088
+ depths=[2, 2, 6, 2],
1089
+ num_heads=[3, 6, 12, 24],
1090
+ window_size=7,
1091
+ mlp_ratio=4.,
1092
+ qkv_bias=True,
1093
+ qk_scale=None,
1094
+ drop_rate=0.,
1095
+ attn_drop_rate=0.,
1096
+ drop_path_rate=0.2,
1097
+ norm_layer=nn.LayerNorm,
1098
+ ape=False,
1099
+ patch_norm=True,
1100
+ out_indices=(0, 1, 2, 3),
1101
+ frozen_stages=-1,
1102
+ use_checkpoint=False):
1103
+ super().__init__()
1104
+
1105
+ self.pretrain_img_size = pretrain_img_size
1106
+ self.num_layers = len(depths)
1107
+ self.embed_dim = embed_dim
1108
+ self.ape = ape
1109
+ self.patch_norm = patch_norm
1110
+ self.out_indices = out_indices
1111
+ self.frozen_stages = frozen_stages
1112
+
1113
+ # split image into non-overlapping patches
1114
+ self.patch_embed = PatchEmbed(
1115
+ patch_size=patch_size, in_channels=in_channels, embed_dim=embed_dim,
1116
+ norm_layer=norm_layer if self.patch_norm else None)
1117
+
1118
+ # absolute position embedding
1119
+ if self.ape:
1120
+ pretrain_img_size = to_2tuple(pretrain_img_size)
1121
+ patch_size = to_2tuple(patch_size)
1122
+ patches_resolution = [pretrain_img_size[0] // patch_size[0], pretrain_img_size[1] // patch_size[1]]
1123
+
1124
+ self.absolute_pos_embed = nn.Parameter(torch.zeros(1, embed_dim, patches_resolution[0], patches_resolution[1]))
1125
+ trunc_normal_(self.absolute_pos_embed, std=.02)
1126
+
1127
+ self.pos_drop = nn.Dropout(p=drop_rate)
1128
+
1129
+ # stochastic depth
1130
+ dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule
1131
+
1132
+ # build layers
1133
+ self.layers = nn.ModuleList()
1134
+ for i_layer in range(self.num_layers):
1135
+ layer = BasicLayer(
1136
+ dim=int(embed_dim * 2 ** i_layer),
1137
+ depth=depths[i_layer],
1138
+ num_heads=num_heads[i_layer],
1139
+ window_size=window_size,
1140
+ mlp_ratio=mlp_ratio,
1141
+ qkv_bias=qkv_bias,
1142
+ qk_scale=qk_scale,
1143
+ drop=drop_rate,
1144
+ attn_drop=attn_drop_rate,
1145
+ drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],
1146
+ norm_layer=norm_layer,
1147
+ downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
1148
+ use_checkpoint=use_checkpoint)
1149
+ self.layers.append(layer)
1150
+
1151
+ num_features = [int(embed_dim * 2 ** i) for i in range(self.num_layers)]
1152
+ self.num_features = num_features
1153
+
1154
+ # add a norm layer for each output
1155
+ for i_layer in out_indices:
1156
+ layer = norm_layer(num_features[i_layer])
1157
+ layer_name = f'norm{i_layer}'
1158
+ self.add_module(layer_name, layer)
1159
+
1160
+ self._freeze_stages()
1161
+
1162
+ def _freeze_stages(self):
1163
+ if self.frozen_stages >= 0:
1164
+ self.patch_embed.eval()
1165
+ for param in self.patch_embed.parameters():
1166
+ param.requires_grad = False
1167
+
1168
+ if self.frozen_stages >= 1 and self.ape:
1169
+ self.absolute_pos_embed.requires_grad = False
1170
+
1171
+ if self.frozen_stages >= 2:
1172
+ self.pos_drop.eval()
1173
+ for i in range(0, self.frozen_stages - 1):
1174
+ m = self.layers[i]
1175
+ m.eval()
1176
+ for param in m.parameters():
1177
+ param.requires_grad = False
1178
+
1179
+
1180
+ def forward(self, x):
1181
+ """Forward function."""
1182
+ x = self.patch_embed(x)
1183
+
1184
+ Wh, Ww = x.size(2), x.size(3)
1185
+ if self.ape:
1186
+ # interpolate the position embedding to the corresponding size
1187
+ absolute_pos_embed = F.interpolate(self.absolute_pos_embed, size=(Wh, Ww), mode='bicubic')
1188
+ x = (x + absolute_pos_embed) # B Wh*Ww C
1189
+
1190
+ outs = []#x.contiguous()]
1191
+ x = x.flatten(2).transpose(1, 2)
1192
+ x = self.pos_drop(x)
1193
+ for i in range(self.num_layers):
1194
+ layer = self.layers[i]
1195
+ x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww)
1196
+
1197
+ if i in self.out_indices:
1198
+ norm_layer = getattr(self, f'norm{i}')
1199
+ x_out = norm_layer(x_out)
1200
+
1201
+ out = x_out.view(-1, H, W, self.num_features[i]).permute(0, 3, 1, 2).contiguous()
1202
+ outs.append(out)
1203
+
1204
+ return tuple(outs)
1205
+
1206
+ def train(self, mode=True):
1207
+ """Convert the model into training mode while keep layers freezed."""
1208
+ super(SwinTransformer, self).train(mode)
1209
+ self._freeze_stages()
1210
+
1211
+ def swin_v1_t():
1212
+ model = SwinTransformer(embed_dim=96, depths=[2, 2, 6, 2], num_heads=[3, 6, 12, 24], window_size=7)
1213
+ return model
1214
+
1215
+ def swin_v1_s():
1216
+ model = SwinTransformer(embed_dim=96, depths=[2, 2, 18, 2], num_heads=[3, 6, 12, 24], window_size=7)
1217
+ return model
1218
+
1219
+ def swin_v1_b():
1220
+ model = SwinTransformer(embed_dim=128, depths=[2, 2, 18, 2], num_heads=[4, 8, 16, 32], window_size=12)
1221
+ return model
1222
+
1223
+ def swin_v1_l():
1224
+ model = SwinTransformer(embed_dim=192, depths=[2, 2, 18, 2], num_heads=[6, 12, 24, 48], window_size=12)
1225
+ return model
1226
+
1227
+
1228
+
1229
+ ### models/modules/deform_conv.py
1230
+
1231
+ import torch
1232
+ import torch.nn as nn
1233
+ from torchvision.ops import deform_conv2d
1234
+
1235
+
1236
+ class DeformableConv2d(nn.Module):
1237
+ def __init__(self,
1238
+ in_channels,
1239
+ out_channels,
1240
+ kernel_size=3,
1241
+ stride=1,
1242
+ padding=1,
1243
+ bias=False):
1244
+
1245
+ super(DeformableConv2d, self).__init__()
1246
+
1247
+ assert type(kernel_size) == tuple or type(kernel_size) == int
1248
+
1249
+ kernel_size = kernel_size if type(kernel_size) == tuple else (kernel_size, kernel_size)
1250
+ self.stride = stride if type(stride) == tuple else (stride, stride)
1251
+ self.padding = padding
1252
+
1253
+ self.offset_conv = nn.Conv2d(in_channels,
1254
+ 2 * kernel_size[0] * kernel_size[1],
1255
+ kernel_size=kernel_size,
1256
+ stride=stride,
1257
+ padding=self.padding,
1258
+ bias=True)
1259
+
1260
+ nn.init.constant_(self.offset_conv.weight, 0.)
1261
+ nn.init.constant_(self.offset_conv.bias, 0.)
1262
+
1263
+ self.modulator_conv = nn.Conv2d(in_channels,
1264
+ 1 * kernel_size[0] * kernel_size[1],
1265
+ kernel_size=kernel_size,
1266
+ stride=stride,
1267
+ padding=self.padding,
1268
+ bias=True)
1269
+
1270
+ nn.init.constant_(self.modulator_conv.weight, 0.)
1271
+ nn.init.constant_(self.modulator_conv.bias, 0.)
1272
+
1273
+ self.regular_conv = nn.Conv2d(in_channels,
1274
+ out_channels=out_channels,
1275
+ kernel_size=kernel_size,
1276
+ stride=stride,
1277
+ padding=self.padding,
1278
+ bias=bias)
1279
+
1280
+ def forward(self, x):
1281
+ #h, w = x.shape[2:]
1282
+ #max_offset = max(h, w)/4.
1283
+
1284
+ offset = self.offset_conv(x)#.clamp(-max_offset, max_offset)
1285
+ modulator = 2. * torch.sigmoid(self.modulator_conv(x))
1286
+
1287
+ x = deform_conv2d(
1288
+ input=x,
1289
+ offset=offset,
1290
+ weight=self.regular_conv.weight,
1291
+ bias=self.regular_conv.bias,
1292
+ padding=self.padding,
1293
+ mask=modulator,
1294
+ stride=self.stride,
1295
+ )
1296
+ return x
1297
+
1298
+
1299
+
1300
+
1301
+ ### utils.py
1302
+
1303
+ import torch.nn as nn
1304
+
1305
+
1306
+ def build_act_layer(act_layer):
1307
+ if act_layer == 'ReLU':
1308
+ return nn.ReLU(inplace=True)
1309
+ elif act_layer == 'SiLU':
1310
+ return nn.SiLU(inplace=True)
1311
+ elif act_layer == 'GELU':
1312
+ return nn.GELU()
1313
+
1314
+ raise NotImplementedError(f'build_act_layer does not support {act_layer}')
1315
+
1316
+
1317
+ def build_norm_layer(dim,
1318
+ norm_layer,
1319
+ in_format='channels_last',
1320
+ out_format='channels_last',
1321
+ eps=1e-6):
1322
+ layers = []
1323
+ if norm_layer == 'BN':
1324
+ if in_format == 'channels_last':
1325
+ layers.append(to_channels_first())
1326
+ layers.append(nn.BatchNorm2d(dim))
1327
+ if out_format == 'channels_last':
1328
+ layers.append(to_channels_last())
1329
+ elif norm_layer == 'LN':
1330
+ if in_format == 'channels_first':
1331
+ layers.append(to_channels_last())
1332
+ layers.append(nn.LayerNorm(dim, eps=eps))
1333
+ if out_format == 'channels_first':
1334
+ layers.append(to_channels_first())
1335
+ else:
1336
+ raise NotImplementedError(
1337
+ f'build_norm_layer does not support {norm_layer}')
1338
+ return nn.Sequential(*layers)
1339
+
1340
+
1341
+ class to_channels_first(nn.Module):
1342
+
1343
+ def __init__(self):
1344
+ super().__init__()
1345
+
1346
+ def forward(self, x):
1347
+ return x.permute(0, 3, 1, 2)
1348
+
1349
+
1350
+ class to_channels_last(nn.Module):
1351
+
1352
+ def __init__(self):
1353
+ super().__init__()
1354
+
1355
+ def forward(self, x):
1356
+ return x.permute(0, 2, 3, 1)
1357
+
1358
+
1359
+
1360
+ ### dataset.py
1361
+
1362
+ _class_labels_TR_sorted = (
1363
+ 'Airplane, Ant, Antenna, Archery, Axe, BabyCarriage, Bag, BalanceBeam, Balcony, Balloon, Basket, BasketballHoop, Beatle, Bed, Bee, Bench, Bicycle, '
1364
+ 'BicycleFrame, BicycleStand, Boat, Bonsai, BoomLift, Bridge, BunkBed, Butterfly, Button, Cable, CableLift, Cage, Camcorder, Cannon, Canoe, Car, '
1365
+ 'CarParkDropArm, Carriage, Cart, Caterpillar, CeilingLamp, Centipede, Chair, Clip, Clock, Clothes, CoatHanger, Comb, ConcretePumpTruck, Crack, Crane, '
1366
+ 'Cup, DentalChair, Desk, DeskChair, Diagram, DishRack, DoorHandle, Dragonfish, Dragonfly, Drum, Earphone, Easel, ElectricIron, Excavator, Eyeglasses, '
1367
+ 'Fan, Fence, Fencing, FerrisWheel, FireExtinguisher, Fishing, Flag, FloorLamp, Forklift, GasStation, Gate, Gear, Goal, Golf, GymEquipment, Hammock, '
1368
+ 'Handcart, Handcraft, Handrail, HangGlider, Harp, Harvester, Headset, Helicopter, Helmet, Hook, HorizontalBar, Hydrovalve, IroningTable, Jewelry, Key, '
1369
+ 'KidsPlayground, Kitchenware, Kite, Knife, Ladder, LaundryRack, Lightning, Lobster, Locust, Machine, MachineGun, MagazineRack, Mantis, Medal, MemorialArchway, '
1370
+ 'Microphone, Missile, MobileHolder, Monitor, Mosquito, Motorcycle, MovingTrolley, Mower, MusicPlayer, MusicStand, ObservationTower, Octopus, OilWell, '
1371
+ 'OlympicLogo, OperatingTable, OutdoorFitnessEquipment, Parachute, Pavilion, Piano, Pipe, PlowHarrow, PoleVault, Punchbag, Rack, Racket, Rifle, Ring, Robot, '
1372
+ 'RockClimbing, Rope, Sailboat, Satellite, Scaffold, Scale, Scissor, Scooter, Sculpture, Seadragon, Seahorse, Seal, SewingMachine, Ship, Shoe, ShoppingCart, '
1373
+ 'ShoppingTrolley, Shower, Shrimp, Signboard, Skateboarding, Skeleton, Skiing, Spade, SpeedBoat, Spider, Spoon, Stair, Stand, Stationary, SteeringWheel, '
1374
+ 'Stethoscope, Stool, Stove, StreetLamp, SweetStand, Swing, Sword, TV, Table, TableChair, TableLamp, TableTennis, Tank, Tapeline, Teapot, Telescope, Tent, '
1375
+ 'TobaccoPipe, Toy, Tractor, TrafficLight, TrafficSign, Trampoline, TransmissionTower, Tree, Tricycle, TrimmerCover, Tripod, Trombone, Truck, Trumpet, Tuba, '
1376
+ 'UAV, Umbrella, UnevenBars, UtilityPole, VacuumCleaner, Violin, Wakesurfing, Watch, WaterTower, WateringPot, Well, WellLid, Wheel, Wheelchair, WindTurbine, Windmill, WineGlass, WireWhisk, Yacht'
1377
+ )
1378
+ class_labels_TR_sorted = _class_labels_TR_sorted.split(', ')
1379
+
1380
+
1381
+ ### models/backbones/build_backbones.py
1382
+
1383
+ import torch
1384
+ import torch.nn as nn
1385
+ from collections import OrderedDict
1386
+ from torchvision.models import vgg16, vgg16_bn, VGG16_Weights, VGG16_BN_Weights, resnet50, ResNet50_Weights
1387
+ # from models.pvt_v2 import pvt_v2_b0, pvt_v2_b1, pvt_v2_b2, pvt_v2_b5
1388
+ # from models.swin_v1 import swin_v1_t, swin_v1_s, swin_v1_b, swin_v1_l
1389
+ # from config import Config
1390
+
1391
+
1392
+ config = Config()
1393
+
1394
+ def build_backbone(bb_name, pretrained=True, params_settings=''):
1395
+ if bb_name == 'vgg16':
1396
+ bb_net = list(vgg16(pretrained=VGG16_Weights.DEFAULT if pretrained else None).children())[0]
1397
+ bb = nn.Sequential(OrderedDict({'conv1': bb_net[:4], 'conv2': bb_net[4:9], 'conv3': bb_net[9:16], 'conv4': bb_net[16:23]}))
1398
+ elif bb_name == 'vgg16bn':
1399
+ bb_net = list(vgg16_bn(pretrained=VGG16_BN_Weights.DEFAULT if pretrained else None).children())[0]
1400
+ bb = nn.Sequential(OrderedDict({'conv1': bb_net[:6], 'conv2': bb_net[6:13], 'conv3': bb_net[13:23], 'conv4': bb_net[23:33]}))
1401
+ elif bb_name == 'resnet50':
1402
+ bb_net = list(resnet50(pretrained=ResNet50_Weights.DEFAULT if pretrained else None).children())
1403
+ bb = nn.Sequential(OrderedDict({'conv1': nn.Sequential(*bb_net[0:3]), 'conv2': bb_net[4], 'conv3': bb_net[5], 'conv4': bb_net[6]}))
1404
+ else:
1405
+ bb = eval('{}({})'.format(bb_name, params_settings))
1406
+ if pretrained:
1407
+ bb = load_weights(bb, bb_name)
1408
+ return bb
1409
+
1410
+ def load_weights(model, model_name):
1411
+ save_model = torch.load(config.weights[model_name], map_location='cpu')
1412
+ model_dict = model.state_dict()
1413
+ state_dict = {k: v if v.size() == model_dict[k].size() else model_dict[k] for k, v in save_model.items() if k in model_dict.keys()}
1414
+ # to ignore the weights with mismatched size when I modify the backbone itself.
1415
+ if not state_dict:
1416
+ save_model_keys = list(save_model.keys())
1417
+ sub_item = save_model_keys[0] if len(save_model_keys) == 1 else None
1418
+ state_dict = {k: v if v.size() == model_dict[k].size() else model_dict[k] for k, v in save_model[sub_item].items() if k in model_dict.keys()}
1419
+ if not state_dict or not sub_item:
1420
+ print('Weights are not successully loaded. Check the state dict of weights file.')
1421
+ return None
1422
+ else:
1423
+ print('Found correct weights in the "{}" item of loaded state_dict.'.format(sub_item))
1424
+ model_dict.update(state_dict)
1425
+ model.load_state_dict(model_dict)
1426
+ return model
1427
+
1428
+
1429
+
1430
+ ### models/modules/decoder_blocks.py
1431
+
1432
+ import torch
1433
+ import torch.nn as nn
1434
+ # from models.aspp import ASPP, ASPPDeformable
1435
+ # from config import Config
1436
+
1437
+
1438
+ # config = Config()
1439
+
1440
+
1441
+ class BasicDecBlk(nn.Module):
1442
+ def __init__(self, in_channels=64, out_channels=64, inter_channels=64):
1443
+ super(BasicDecBlk, self).__init__()
1444
+ inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
1445
+ self.conv_in = nn.Conv2d(in_channels, inter_channels, 3, 1, padding=1)
1446
+ self.relu_in = nn.ReLU(inplace=True)
1447
+ if config.dec_att == 'ASPP':
1448
+ self.dec_att = ASPP(in_channels=inter_channels)
1449
+ elif config.dec_att == 'ASPPDeformable':
1450
+ self.dec_att = ASPPDeformable(in_channels=inter_channels)
1451
+ self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, padding=1)
1452
+ self.bn_in = nn.BatchNorm2d(inter_channels) if config.batch_size > 1 else nn.Identity()
1453
+ self.bn_out = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
1454
+
1455
+ def forward(self, x):
1456
+ x = self.conv_in(x)
1457
+ x = self.bn_in(x)
1458
+ x = self.relu_in(x)
1459
+ if hasattr(self, 'dec_att'):
1460
+ x = self.dec_att(x)
1461
+ x = self.conv_out(x)
1462
+ x = self.bn_out(x)
1463
+ return x
1464
+
1465
+
1466
+ class ResBlk(nn.Module):
1467
+ def __init__(self, in_channels=64, out_channels=None, inter_channels=64):
1468
+ super(ResBlk, self).__init__()
1469
+ if out_channels is None:
1470
+ out_channels = in_channels
1471
+ inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
1472
+
1473
+ self.conv_in = nn.Conv2d(in_channels, inter_channels, 3, 1, padding=1)
1474
+ self.bn_in = nn.BatchNorm2d(inter_channels) if config.batch_size > 1 else nn.Identity()
1475
+ self.relu_in = nn.ReLU(inplace=True)
1476
+
1477
+ if config.dec_att == 'ASPP':
1478
+ self.dec_att = ASPP(in_channels=inter_channels)
1479
+ elif config.dec_att == 'ASPPDeformable':
1480
+ self.dec_att = ASPPDeformable(in_channels=inter_channels)
1481
+
1482
+ self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, padding=1)
1483
+ self.bn_out = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
1484
+
1485
+ self.conv_resi = nn.Conv2d(in_channels, out_channels, 1, 1, 0)
1486
+
1487
+ def forward(self, x):
1488
+ _x = self.conv_resi(x)
1489
+ x = self.conv_in(x)
1490
+ x = self.bn_in(x)
1491
+ x = self.relu_in(x)
1492
+ if hasattr(self, 'dec_att'):
1493
+ x = self.dec_att(x)
1494
+ x = self.conv_out(x)
1495
+ x = self.bn_out(x)
1496
+ return x + _x
1497
+
1498
+
1499
+
1500
+ ### models/modules/lateral_blocks.py
1501
+
1502
+ import numpy as np
1503
+ import torch
1504
+ import torch.nn as nn
1505
+ import torch.nn.functional as F
1506
+ from functools import partial
1507
+
1508
+ # from config import Config
1509
+
1510
+
1511
+ # config = Config()
1512
+
1513
+
1514
+ class BasicLatBlk(nn.Module):
1515
+ def __init__(self, in_channels=64, out_channels=64, inter_channels=64):
1516
+ super(BasicLatBlk, self).__init__()
1517
+ inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
1518
+ self.conv = nn.Conv2d(in_channels, out_channels, 1, 1, 0)
1519
+
1520
+ def forward(self, x):
1521
+ x = self.conv(x)
1522
+ return x
1523
+
1524
+
1525
+
1526
+ ### models/modules/aspp.py
1527
+
1528
+ import torch
1529
+ import torch.nn as nn
1530
+ import torch.nn.functional as F
1531
+ # from models.deform_conv import DeformableConv2d
1532
+ # from config import Config
1533
+
1534
+
1535
+ # config = Config()
1536
+
1537
+
1538
+ class _ASPPModule(nn.Module):
1539
+ def __init__(self, in_channels, planes, kernel_size, padding, dilation):
1540
+ super(_ASPPModule, self).__init__()
1541
+ self.atrous_conv = nn.Conv2d(in_channels, planes, kernel_size=kernel_size,
1542
+ stride=1, padding=padding, dilation=dilation, bias=False)
1543
+ self.bn = nn.BatchNorm2d(planes) if config.batch_size > 1 else nn.Identity()
1544
+ self.relu = nn.ReLU(inplace=True)
1545
+
1546
+ def forward(self, x):
1547
+ x = self.atrous_conv(x)
1548
+ x = self.bn(x)
1549
+
1550
+ return self.relu(x)
1551
+
1552
+
1553
+ class ASPP(nn.Module):
1554
+ def __init__(self, in_channels=64, out_channels=None, output_stride=16):
1555
+ super(ASPP, self).__init__()
1556
+ self.down_scale = 1
1557
+ if out_channels is None:
1558
+ out_channels = in_channels
1559
+ self.in_channelster = 256 // self.down_scale
1560
+ if output_stride == 16:
1561
+ dilations = [1, 6, 12, 18]
1562
+ elif output_stride == 8:
1563
+ dilations = [1, 12, 24, 36]
1564
+ else:
1565
+ raise NotImplementedError
1566
+
1567
+ self.aspp1 = _ASPPModule(in_channels, self.in_channelster, 1, padding=0, dilation=dilations[0])
1568
+ self.aspp2 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[1], dilation=dilations[1])
1569
+ self.aspp3 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[2], dilation=dilations[2])
1570
+ self.aspp4 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[3], dilation=dilations[3])
1571
+
1572
+ self.global_avg_pool = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)),
1573
+ nn.Conv2d(in_channels, self.in_channelster, 1, stride=1, bias=False),
1574
+ nn.BatchNorm2d(self.in_channelster) if config.batch_size > 1 else nn.Identity(),
1575
+ nn.ReLU(inplace=True))
1576
+ self.conv1 = nn.Conv2d(self.in_channelster * 5, out_channels, 1, bias=False)
1577
+ self.bn1 = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
1578
+ self.relu = nn.ReLU(inplace=True)
1579
+ self.dropout = nn.Dropout(0.5)
1580
+
1581
+ def forward(self, x):
1582
+ x1 = self.aspp1(x)
1583
+ x2 = self.aspp2(x)
1584
+ x3 = self.aspp3(x)
1585
+ x4 = self.aspp4(x)
1586
+ x5 = self.global_avg_pool(x)
1587
+ x5 = F.interpolate(x5, size=x1.size()[2:], mode='bilinear', align_corners=True)
1588
+ x = torch.cat((x1, x2, x3, x4, x5), dim=1)
1589
+
1590
+ x = self.conv1(x)
1591
+ x = self.bn1(x)
1592
+ x = self.relu(x)
1593
+
1594
+ return self.dropout(x)
1595
+
1596
+
1597
+ ##################### Deformable
1598
+ class _ASPPModuleDeformable(nn.Module):
1599
+ def __init__(self, in_channels, planes, kernel_size, padding):
1600
+ super(_ASPPModuleDeformable, self).__init__()
1601
+ self.atrous_conv = DeformableConv2d(in_channels, planes, kernel_size=kernel_size,
1602
+ stride=1, padding=padding, bias=False)
1603
+ self.bn = nn.BatchNorm2d(planes) if config.batch_size > 1 else nn.Identity()
1604
+ self.relu = nn.ReLU(inplace=True)
1605
+
1606
+ def forward(self, x):
1607
+ x = self.atrous_conv(x)
1608
+ x = self.bn(x)
1609
+
1610
+ return self.relu(x)
1611
+
1612
+
1613
+ class ASPPDeformable(nn.Module):
1614
+ def __init__(self, in_channels, out_channels=None, parallel_block_sizes=[1, 3, 7]):
1615
+ super(ASPPDeformable, self).__init__()
1616
+ self.down_scale = 1
1617
+ if out_channels is None:
1618
+ out_channels = in_channels
1619
+ self.in_channelster = 256 // self.down_scale
1620
+
1621
+ self.aspp1 = _ASPPModuleDeformable(in_channels, self.in_channelster, 1, padding=0)
1622
+ self.aspp_deforms = nn.ModuleList([
1623
+ _ASPPModuleDeformable(in_channels, self.in_channelster, conv_size, padding=int(conv_size//2)) for conv_size in parallel_block_sizes
1624
+ ])
1625
+
1626
+ self.global_avg_pool = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)),
1627
+ nn.Conv2d(in_channels, self.in_channelster, 1, stride=1, bias=False),
1628
+ nn.BatchNorm2d(self.in_channelster) if config.batch_size > 1 else nn.Identity(),
1629
+ nn.ReLU(inplace=True))
1630
+ self.conv1 = nn.Conv2d(self.in_channelster * (2 + len(self.aspp_deforms)), out_channels, 1, bias=False)
1631
+ self.bn1 = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
1632
+ self.relu = nn.ReLU(inplace=True)
1633
+ self.dropout = nn.Dropout(0.5)
1634
+
1635
+ def forward(self, x):
1636
+ x1 = self.aspp1(x)
1637
+ x_aspp_deforms = [aspp_deform(x) for aspp_deform in self.aspp_deforms]
1638
+ x5 = self.global_avg_pool(x)
1639
+ x5 = F.interpolate(x5, size=x1.size()[2:], mode='bilinear', align_corners=True)
1640
+ x = torch.cat((x1, *x_aspp_deforms, x5), dim=1)
1641
+
1642
+ x = self.conv1(x)
1643
+ x = self.bn1(x)
1644
+ x = self.relu(x)
1645
+
1646
+ return self.dropout(x)
1647
+
1648
+
1649
+
1650
+ ### models/refinement/refiner.py
1651
+
1652
+ import torch
1653
+ import torch.nn as nn
1654
+ from collections import OrderedDict
1655
+ import torch
1656
+ import torch.nn as nn
1657
+ import torch.nn.functional as F
1658
+ from torchvision.models import vgg16, vgg16_bn
1659
+ from torchvision.models import resnet50
1660
+
1661
+ # from config import Config
1662
+ # from dataset import class_labels_TR_sorted
1663
+ # from models.build_backbone import build_backbone
1664
+ # from models.decoder_blocks import BasicDecBlk
1665
+ # from models.lateral_blocks import BasicLatBlk
1666
+ # from models.ing import *
1667
+ # from models.stem_layer import StemLayer
1668
+
1669
+
1670
+ class RefinerPVTInChannels4(nn.Module):
1671
+ def __init__(self, in_channels=3+1):
1672
+ super(RefinerPVTInChannels4, self).__init__()
1673
+ self.config = Config()
1674
+ self.epoch = 1
1675
+ self.bb = build_backbone(self.config.bb, params_settings='in_channels=4')
1676
+
1677
+ lateral_channels_in_collection = {
1678
+ 'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
1679
+ 'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
1680
+ 'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
1681
+ }
1682
+ channels = lateral_channels_in_collection[self.config.bb]
1683
+ self.squeeze_module = BasicDecBlk(channels[0], channels[0])
1684
+
1685
+ self.decoder = Decoder(channels)
1686
+
1687
+ if 0:
1688
+ for key, value in self.named_parameters():
1689
+ if 'bb.' in key:
1690
+ value.requires_grad = False
1691
+
1692
+ def forward(self, x):
1693
+ if isinstance(x, list):
1694
+ x = torch.cat(x, dim=1)
1695
+ ########## Encoder ##########
1696
+ if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
1697
+ x1 = self.bb.conv1(x)
1698
+ x2 = self.bb.conv2(x1)
1699
+ x3 = self.bb.conv3(x2)
1700
+ x4 = self.bb.conv4(x3)
1701
+ else:
1702
+ x1, x2, x3, x4 = self.bb(x)
1703
+
1704
+ x4 = self.squeeze_module(x4)
1705
+
1706
+ ########## Decoder ##########
1707
+
1708
+ features = [x, x1, x2, x3, x4]
1709
+ scaled_preds = self.decoder(features)
1710
+
1711
+ return scaled_preds
1712
+
1713
+
1714
+ class Refiner(nn.Module):
1715
+ def __init__(self, in_channels=3+1):
1716
+ super(Refiner, self).__init__()
1717
+ self.config = Config()
1718
+ self.epoch = 1
1719
+ self.stem_layer = StemLayer(in_channels=in_channels, inter_channels=48, out_channels=3, norm_layer='BN' if self.config.batch_size > 1 else 'LN')
1720
+ self.bb = build_backbone(self.config.bb)
1721
+
1722
+ lateral_channels_in_collection = {
1723
+ 'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
1724
+ 'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
1725
+ 'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
1726
+ }
1727
+ channels = lateral_channels_in_collection[self.config.bb]
1728
+ self.squeeze_module = BasicDecBlk(channels[0], channels[0])
1729
+
1730
+ self.decoder = Decoder(channels)
1731
+
1732
+ if 0:
1733
+ for key, value in self.named_parameters():
1734
+ if 'bb.' in key:
1735
+ value.requires_grad = False
1736
+
1737
+ def forward(self, x):
1738
+ if isinstance(x, list):
1739
+ x = torch.cat(x, dim=1)
1740
+ x = self.stem_layer(x)
1741
+ ########## Encoder ##########
1742
+ if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
1743
+ x1 = self.bb.conv1(x)
1744
+ x2 = self.bb.conv2(x1)
1745
+ x3 = self.bb.conv3(x2)
1746
+ x4 = self.bb.conv4(x3)
1747
+ else:
1748
+ x1, x2, x3, x4 = self.bb(x)
1749
+
1750
+ x4 = self.squeeze_module(x4)
1751
+
1752
+ ########## Decoder ##########
1753
+
1754
+ features = [x, x1, x2, x3, x4]
1755
+ scaled_preds = self.decoder(features)
1756
+
1757
+ return scaled_preds
1758
+
1759
+
1760
+ class Decoder(nn.Module):
1761
+ def __init__(self, channels):
1762
+ super(Decoder, self).__init__()
1763
+ self.config = Config()
1764
+ DecoderBlock = eval('BasicDecBlk')
1765
+ LateralBlock = eval('BasicLatBlk')
1766
+
1767
+ self.decoder_block4 = DecoderBlock(channels[0], channels[1])
1768
+ self.decoder_block3 = DecoderBlock(channels[1], channels[2])
1769
+ self.decoder_block2 = DecoderBlock(channels[2], channels[3])
1770
+ self.decoder_block1 = DecoderBlock(channels[3], channels[3]//2)
1771
+
1772
+ self.lateral_block4 = LateralBlock(channels[1], channels[1])
1773
+ self.lateral_block3 = LateralBlock(channels[2], channels[2])
1774
+ self.lateral_block2 = LateralBlock(channels[3], channels[3])
1775
+
1776
+ if self.config.ms_supervision:
1777
+ self.conv_ms_spvn_4 = nn.Conv2d(channels[1], 1, 1, 1, 0)
1778
+ self.conv_ms_spvn_3 = nn.Conv2d(channels[2], 1, 1, 1, 0)
1779
+ self.conv_ms_spvn_2 = nn.Conv2d(channels[3], 1, 1, 1, 0)
1780
+ self.conv_out1 = nn.Sequential(nn.Conv2d(channels[3]//2, 1, 1, 1, 0))
1781
+
1782
+ def forward(self, features):
1783
+ x, x1, x2, x3, x4 = features
1784
+ outs = []
1785
+ p4 = self.decoder_block4(x4)
1786
+ _p4 = F.interpolate(p4, size=x3.shape[2:], mode='bilinear', align_corners=True)
1787
+ _p3 = _p4 + self.lateral_block4(x3)
1788
+
1789
+ p3 = self.decoder_block3(_p3)
1790
+ _p3 = F.interpolate(p3, size=x2.shape[2:], mode='bilinear', align_corners=True)
1791
+ _p2 = _p3 + self.lateral_block3(x2)
1792
+
1793
+ p2 = self.decoder_block2(_p2)
1794
+ _p2 = F.interpolate(p2, size=x1.shape[2:], mode='bilinear', align_corners=True)
1795
+ _p1 = _p2 + self.lateral_block2(x1)
1796
+
1797
+ _p1 = self.decoder_block1(_p1)
1798
+ _p1 = F.interpolate(_p1, size=x.shape[2:], mode='bilinear', align_corners=True)
1799
+ p1_out = self.conv_out1(_p1)
1800
+
1801
+ if self.config.ms_supervision:
1802
+ outs.append(self.conv_ms_spvn_4(p4))
1803
+ outs.append(self.conv_ms_spvn_3(p3))
1804
+ outs.append(self.conv_ms_spvn_2(p2))
1805
+ outs.append(p1_out)
1806
+ return outs
1807
+
1808
+
1809
+ class RefUNet(nn.Module):
1810
+ # Refinement
1811
+ def __init__(self, in_channels=3+1):
1812
+ super(RefUNet, self).__init__()
1813
+ self.encoder_1 = nn.Sequential(
1814
+ nn.Conv2d(in_channels, 64, 3, 1, 1),
1815
+ nn.Conv2d(64, 64, 3, 1, 1),
1816
+ nn.BatchNorm2d(64),
1817
+ nn.ReLU(inplace=True)
1818
+ )
1819
+
1820
+ self.encoder_2 = nn.Sequential(
1821
+ nn.MaxPool2d(2, 2, ceil_mode=True),
1822
+ nn.Conv2d(64, 64, 3, 1, 1),
1823
+ nn.BatchNorm2d(64),
1824
+ nn.ReLU(inplace=True)
1825
+ )
1826
+
1827
+ self.encoder_3 = nn.Sequential(
1828
+ nn.MaxPool2d(2, 2, ceil_mode=True),
1829
+ nn.Conv2d(64, 64, 3, 1, 1),
1830
+ nn.BatchNorm2d(64),
1831
+ nn.ReLU(inplace=True)
1832
+ )
1833
+
1834
+ self.encoder_4 = nn.Sequential(
1835
+ nn.MaxPool2d(2, 2, ceil_mode=True),
1836
+ nn.Conv2d(64, 64, 3, 1, 1),
1837
+ nn.BatchNorm2d(64),
1838
+ nn.ReLU(inplace=True)
1839
+ )
1840
+
1841
+ self.pool4 = nn.MaxPool2d(2, 2, ceil_mode=True)
1842
+ #####
1843
+ self.decoder_5 = nn.Sequential(
1844
+ nn.Conv2d(64, 64, 3, 1, 1),
1845
+ nn.BatchNorm2d(64),
1846
+ nn.ReLU(inplace=True)
1847
+ )
1848
+ #####
1849
+ self.decoder_4 = nn.Sequential(
1850
+ nn.Conv2d(128, 64, 3, 1, 1),
1851
+ nn.BatchNorm2d(64),
1852
+ nn.ReLU(inplace=True)
1853
+ )
1854
+
1855
+ self.decoder_3 = nn.Sequential(
1856
+ nn.Conv2d(128, 64, 3, 1, 1),
1857
+ nn.BatchNorm2d(64),
1858
+ nn.ReLU(inplace=True)
1859
+ )
1860
+
1861
+ self.decoder_2 = nn.Sequential(
1862
+ nn.Conv2d(128, 64, 3, 1, 1),
1863
+ nn.BatchNorm2d(64),
1864
+ nn.ReLU(inplace=True)
1865
+ )
1866
+
1867
+ self.decoder_1 = nn.Sequential(
1868
+ nn.Conv2d(128, 64, 3, 1, 1),
1869
+ nn.BatchNorm2d(64),
1870
+ nn.ReLU(inplace=True)
1871
+ )
1872
+
1873
+ self.conv_d0 = nn.Conv2d(64, 1, 3, 1, 1)
1874
+
1875
+ self.upscore2 = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)
1876
+
1877
+ def forward(self, x):
1878
+ outs = []
1879
+ if isinstance(x, list):
1880
+ x = torch.cat(x, dim=1)
1881
+ hx = x
1882
+
1883
+ hx1 = self.encoder_1(hx)
1884
+ hx2 = self.encoder_2(hx1)
1885
+ hx3 = self.encoder_3(hx2)
1886
+ hx4 = self.encoder_4(hx3)
1887
+
1888
+ hx = self.decoder_5(self.pool4(hx4))
1889
+ hx = torch.cat((self.upscore2(hx), hx4), 1)
1890
+
1891
+ d4 = self.decoder_4(hx)
1892
+ hx = torch.cat((self.upscore2(d4), hx3), 1)
1893
+
1894
+ d3 = self.decoder_3(hx)
1895
+ hx = torch.cat((self.upscore2(d3), hx2), 1)
1896
+
1897
+ d2 = self.decoder_2(hx)
1898
+ hx = torch.cat((self.upscore2(d2), hx1), 1)
1899
+
1900
+ d1 = self.decoder_1(hx)
1901
+
1902
+ x = self.conv_d0(d1)
1903
+ outs.append(x)
1904
+ return outs
1905
+
1906
+
1907
+
1908
+ ### models/stem_layer.py
1909
+
1910
+ import torch.nn as nn
1911
+ # from utils import build_act_layer, build_norm_layer
1912
+
1913
+
1914
+ class StemLayer(nn.Module):
1915
+ r""" Stem layer of InternImage
1916
+ Args:
1917
+ in_channels (int): number of input channels
1918
+ out_channels (int): number of output channels
1919
+ act_layer (str): activation layer
1920
+ norm_layer (str): normalization layer
1921
+ """
1922
+
1923
+ def __init__(self,
1924
+ in_channels=3+1,
1925
+ inter_channels=48,
1926
+ out_channels=96,
1927
+ act_layer='GELU',
1928
+ norm_layer='BN'):
1929
+ super().__init__()
1930
+ self.conv1 = nn.Conv2d(in_channels,
1931
+ inter_channels,
1932
+ kernel_size=3,
1933
+ stride=1,
1934
+ padding=1)
1935
+ self.norm1 = build_norm_layer(
1936
+ inter_channels, norm_layer, 'channels_first', 'channels_first'
1937
+ )
1938
+ self.act = build_act_layer(act_layer)
1939
+ self.conv2 = nn.Conv2d(inter_channels,
1940
+ out_channels,
1941
+ kernel_size=3,
1942
+ stride=1,
1943
+ padding=1)
1944
+ self.norm2 = build_norm_layer(
1945
+ out_channels, norm_layer, 'channels_first', 'channels_first'
1946
+ )
1947
+
1948
+ def forward(self, x):
1949
+ x = self.conv1(x)
1950
+ x = self.norm1(x)
1951
+ x = self.act(x)
1952
+ x = self.conv2(x)
1953
+ x = self.norm2(x)
1954
+ return x
1955
+
1956
+
1957
+ ### models/birefnet.py
1958
+
1959
+ import torch
1960
+ import torch.nn as nn
1961
+ import torch.nn.functional as F
1962
+ from kornia.filters import laplacian
1963
+ from transformers import PreTrainedModel
1964
+
1965
+ # from config import Config
1966
+ # from dataset import class_labels_TR_sorted
1967
+ # from models.build_backbone import build_backbone
1968
+ # from models.decoder_blocks import BasicDecBlk, ResBlk, HierarAttDecBlk
1969
+ # from models.lateral_blocks import BasicLatBlk
1970
+ # from models.aspp import ASPP, ASPPDeformable
1971
+ # from models.ing import *
1972
+ # from models.refiner import Refiner, RefinerPVTInChannels4, RefUNet
1973
+ # from models.stem_layer import StemLayer
1974
+ from .BiRefNet_config import BiRefNetConfig
1975
+
1976
+
1977
+ class BiRefNet(
1978
+ PreTrainedModel
1979
+ ):
1980
+ config_class = BiRefNetConfig
1981
+ def __init__(self, bb_pretrained=True, config=BiRefNetConfig()):
1982
+ super(BiRefNet, self).__init__(config)
1983
+ bb_pretrained = config.bb_pretrained
1984
+ self.config = Config()
1985
+ self.epoch = 1
1986
+ self.bb = build_backbone(self.config.bb, pretrained=bb_pretrained)
1987
+
1988
+ channels = self.config.lateral_channels_in_collection
1989
+
1990
+ if self.config.auxiliary_classification:
1991
+ self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
1992
+ self.cls_head = nn.Sequential(
1993
+ nn.Linear(channels[0], len(class_labels_TR_sorted))
1994
+ )
1995
+
1996
+ if self.config.squeeze_block:
1997
+ self.squeeze_module = nn.Sequential(*[
1998
+ eval(self.config.squeeze_block.split('_x')[0])(channels[0]+sum(self.config.cxt), channels[0])
1999
+ for _ in range(eval(self.config.squeeze_block.split('_x')[1]))
2000
+ ])
2001
+
2002
+ self.decoder = Decoder(channels)
2003
+
2004
+ if self.config.ender:
2005
+ self.dec_end = nn.Sequential(
2006
+ nn.Conv2d(1, 16, 3, 1, 1),
2007
+ nn.Conv2d(16, 1, 3, 1, 1),
2008
+ nn.ReLU(inplace=True),
2009
+ )
2010
+
2011
+ # refine patch-level segmentation
2012
+ if self.config.refine:
2013
+ if self.config.refine == 'itself':
2014
+ self.stem_layer = StemLayer(in_channels=3+1, inter_channels=48, out_channels=3, norm_layer='BN' if self.config.batch_size > 1 else 'LN')
2015
+ else:
2016
+ self.refiner = eval('{}({})'.format(self.config.refine, 'in_channels=3+1'))
2017
+
2018
+ if self.config.freeze_bb:
2019
+ # Freeze the backbone...
2020
+ print(self.named_parameters())
2021
+ for key, value in self.named_parameters():
2022
+ if 'bb.' in key and 'refiner.' not in key:
2023
+ value.requires_grad = False
2024
+
2025
+ def forward_enc(self, x):
2026
+ if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
2027
+ x1 = self.bb.conv1(x); x2 = self.bb.conv2(x1); x3 = self.bb.conv3(x2); x4 = self.bb.conv4(x3)
2028
+ else:
2029
+ x1, x2, x3, x4 = self.bb(x)
2030
+ if self.config.mul_scl_ipt == 'cat':
2031
+ B, C, H, W = x.shape
2032
+ x1_, x2_, x3_, x4_ = self.bb(F.interpolate(x, size=(H//2, W//2), mode='bilinear', align_corners=True))
2033
+ x1 = torch.cat([x1, F.interpolate(x1_, size=x1.shape[2:], mode='bilinear', align_corners=True)], dim=1)
2034
+ x2 = torch.cat([x2, F.interpolate(x2_, size=x2.shape[2:], mode='bilinear', align_corners=True)], dim=1)
2035
+ x3 = torch.cat([x3, F.interpolate(x3_, size=x3.shape[2:], mode='bilinear', align_corners=True)], dim=1)
2036
+ x4 = torch.cat([x4, F.interpolate(x4_, size=x4.shape[2:], mode='bilinear', align_corners=True)], dim=1)
2037
+ elif self.config.mul_scl_ipt == 'add':
2038
+ B, C, H, W = x.shape
2039
+ x1_, x2_, x3_, x4_ = self.bb(F.interpolate(x, size=(H//2, W//2), mode='bilinear', align_corners=True))
2040
+ x1 = x1 + F.interpolate(x1_, size=x1.shape[2:], mode='bilinear', align_corners=True)
2041
+ x2 = x2 + F.interpolate(x2_, size=x2.shape[2:], mode='bilinear', align_corners=True)
2042
+ x3 = x3 + F.interpolate(x3_, size=x3.shape[2:], mode='bilinear', align_corners=True)
2043
+ x4 = x4 + F.interpolate(x4_, size=x4.shape[2:], mode='bilinear', align_corners=True)
2044
+ class_preds = self.cls_head(self.avgpool(x4).view(x4.shape[0], -1)) if self.training and self.config.auxiliary_classification else None
2045
+ if self.config.cxt:
2046
+ x4 = torch.cat(
2047
+ (
2048
+ *[
2049
+ F.interpolate(x1, size=x4.shape[2:], mode='bilinear', align_corners=True),
2050
+ F.interpolate(x2, size=x4.shape[2:], mode='bilinear', align_corners=True),
2051
+ F.interpolate(x3, size=x4.shape[2:], mode='bilinear', align_corners=True),
2052
+ ][-len(self.config.cxt):],
2053
+ x4
2054
+ ),
2055
+ dim=1
2056
+ )
2057
+ return (x1, x2, x3, x4), class_preds
2058
+
2059
+ def forward_ori(self, x):
2060
+ ########## Encoder ##########
2061
+ (x1, x2, x3, x4), class_preds = self.forward_enc(x)
2062
+ if self.config.squeeze_block:
2063
+ x4 = self.squeeze_module(x4)
2064
+ ########## Decoder ##########
2065
+ features = [x, x1, x2, x3, x4]
2066
+ if self.training and self.config.out_ref:
2067
+ features.append(laplacian(torch.mean(x, dim=1).unsqueeze(1), kernel_size=5))
2068
+ scaled_preds = self.decoder(features)
2069
+ return scaled_preds, class_preds
2070
+
2071
+ def forward(self, x):
2072
+ scaled_preds, class_preds = self.forward_ori(x)
2073
+ class_preds_lst = [class_preds]
2074
+ return [scaled_preds, class_preds_lst] if self.training else scaled_preds
2075
+
2076
+
2077
+ class Decoder(nn.Module):
2078
+ def __init__(self, channels):
2079
+ super(Decoder, self).__init__()
2080
+ self.config = Config()
2081
+ DecoderBlock = eval(self.config.dec_blk)
2082
+ LateralBlock = eval(self.config.lat_blk)
2083
+
2084
+ if self.config.dec_ipt:
2085
+ self.split = self.config.dec_ipt_split
2086
+ N_dec_ipt = 64
2087
+ DBlock = SimpleConvs
2088
+ ic = 64
2089
+ ipt_cha_opt = 1
2090
+ self.ipt_blk5 = DBlock(2**10*3 if self.split else 3, [N_dec_ipt, channels[0]//8][ipt_cha_opt], inter_channels=ic)
2091
+ self.ipt_blk4 = DBlock(2**8*3 if self.split else 3, [N_dec_ipt, channels[0]//8][ipt_cha_opt], inter_channels=ic)
2092
+ self.ipt_blk3 = DBlock(2**6*3 if self.split else 3, [N_dec_ipt, channels[1]//8][ipt_cha_opt], inter_channels=ic)
2093
+ self.ipt_blk2 = DBlock(2**4*3 if self.split else 3, [N_dec_ipt, channels[2]//8][ipt_cha_opt], inter_channels=ic)
2094
+ self.ipt_blk1 = DBlock(2**0*3 if self.split else 3, [N_dec_ipt, channels[3]//8][ipt_cha_opt], inter_channels=ic)
2095
+ else:
2096
+ self.split = None
2097
+
2098
+ self.decoder_block4 = DecoderBlock(channels[0]+([N_dec_ipt, channels[0]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[1])
2099
+ self.decoder_block3 = DecoderBlock(channels[1]+([N_dec_ipt, channels[0]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[2])
2100
+ self.decoder_block2 = DecoderBlock(channels[2]+([N_dec_ipt, channels[1]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[3])
2101
+ self.decoder_block1 = DecoderBlock(channels[3]+([N_dec_ipt, channels[2]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[3]//2)
2102
+ self.conv_out1 = nn.Sequential(nn.Conv2d(channels[3]//2+([N_dec_ipt, channels[3]//8][ipt_cha_opt] if self.config.dec_ipt else 0), 1, 1, 1, 0))
2103
+
2104
+ self.lateral_block4 = LateralBlock(channels[1], channels[1])
2105
+ self.lateral_block3 = LateralBlock(channels[2], channels[2])
2106
+ self.lateral_block2 = LateralBlock(channels[3], channels[3])
2107
+
2108
+ if self.config.ms_supervision:
2109
+ self.conv_ms_spvn_4 = nn.Conv2d(channels[1], 1, 1, 1, 0)
2110
+ self.conv_ms_spvn_3 = nn.Conv2d(channels[2], 1, 1, 1, 0)
2111
+ self.conv_ms_spvn_2 = nn.Conv2d(channels[3], 1, 1, 1, 0)
2112
+
2113
+ if self.config.out_ref:
2114
+ _N = 16
2115
+ self.gdt_convs_4 = nn.Sequential(nn.Conv2d(channels[1], _N, 3, 1, 1), nn.BatchNorm2d(_N) if self.config.batch_size > 1 else nn.Identity(), nn.ReLU(inplace=True))
2116
+ self.gdt_convs_3 = nn.Sequential(nn.Conv2d(channels[2], _N, 3, 1, 1), nn.BatchNorm2d(_N) if self.config.batch_size > 1 else nn.Identity(), nn.ReLU(inplace=True))
2117
+ self.gdt_convs_2 = nn.Sequential(nn.Conv2d(channels[3], _N, 3, 1, 1), nn.BatchNorm2d(_N) if self.config.batch_size > 1 else nn.Identity(), nn.ReLU(inplace=True))
2118
+
2119
+ self.gdt_convs_pred_4 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
2120
+ self.gdt_convs_pred_3 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
2121
+ self.gdt_convs_pred_2 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
2122
+
2123
+ self.gdt_convs_attn_4 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
2124
+ self.gdt_convs_attn_3 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
2125
+ self.gdt_convs_attn_2 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
2126
+
2127
+ def get_patches_batch(self, x, p):
2128
+ _size_h, _size_w = p.shape[2:]
2129
+ patches_batch = []
2130
+ for idx in range(x.shape[0]):
2131
+ columns_x = torch.split(x[idx], split_size_or_sections=_size_w, dim=-1)
2132
+ patches_x = []
2133
+ for column_x in columns_x:
2134
+ patches_x += [p.unsqueeze(0) for p in torch.split(column_x, split_size_or_sections=_size_h, dim=-2)]
2135
+ patch_sample = torch.cat(patches_x, dim=1)
2136
+ patches_batch.append(patch_sample)
2137
+ return torch.cat(patches_batch, dim=0)
2138
+
2139
+ def forward(self, features):
2140
+ if self.training and self.config.out_ref:
2141
+ outs_gdt_pred = []
2142
+ outs_gdt_label = []
2143
+ x, x1, x2, x3, x4, gdt_gt = features
2144
+ else:
2145
+ x, x1, x2, x3, x4 = features
2146
+ outs = []
2147
+
2148
+ if self.config.dec_ipt:
2149
+ patches_batch = self.get_patches_batch(x, x4) if self.split else x
2150
+ x4 = torch.cat((x4, self.ipt_blk5(F.interpolate(patches_batch, size=x4.shape[2:], mode='bilinear', align_corners=True))), 1)
2151
+ p4 = self.decoder_block4(x4)
2152
+ m4 = self.conv_ms_spvn_4(p4) if self.config.ms_supervision else None
2153
+ if self.config.out_ref:
2154
+ p4_gdt = self.gdt_convs_4(p4)
2155
+ if self.training:
2156
+ # >> GT:
2157
+ m4_dia = m4
2158
+ gdt_label_main_4 = gdt_gt * F.interpolate(m4_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
2159
+ outs_gdt_label.append(gdt_label_main_4)
2160
+ # >> Pred:
2161
+ gdt_pred_4 = self.gdt_convs_pred_4(p4_gdt)
2162
+ outs_gdt_pred.append(gdt_pred_4)
2163
+ gdt_attn_4 = self.gdt_convs_attn_4(p4_gdt).sigmoid()
2164
+ # >> Finally:
2165
+ p4 = p4 * gdt_attn_4
2166
+ _p4 = F.interpolate(p4, size=x3.shape[2:], mode='bilinear', align_corners=True)
2167
+ _p3 = _p4 + self.lateral_block4(x3)
2168
+
2169
+ if self.config.dec_ipt:
2170
+ patches_batch = self.get_patches_batch(x, _p3) if self.split else x
2171
+ _p3 = torch.cat((_p3, self.ipt_blk4(F.interpolate(patches_batch, size=x3.shape[2:], mode='bilinear', align_corners=True))), 1)
2172
+ p3 = self.decoder_block3(_p3)
2173
+ m3 = self.conv_ms_spvn_3(p3) if self.config.ms_supervision else None
2174
+ if self.config.out_ref:
2175
+ p3_gdt = self.gdt_convs_3(p3)
2176
+ if self.training:
2177
+ # >> GT:
2178
+ # m3 --dilation--> m3_dia
2179
+ # G_3^gt * m3_dia --> G_3^m, which is the label of gradient
2180
+ m3_dia = m3
2181
+ gdt_label_main_3 = gdt_gt * F.interpolate(m3_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
2182
+ outs_gdt_label.append(gdt_label_main_3)
2183
+ # >> Pred:
2184
+ # p3 --conv--BN--> F_3^G, where F_3^G predicts the \hat{G_3} with xx
2185
+ # F_3^G --sigmoid--> A_3^G
2186
+ gdt_pred_3 = self.gdt_convs_pred_3(p3_gdt)
2187
+ outs_gdt_pred.append(gdt_pred_3)
2188
+ gdt_attn_3 = self.gdt_convs_attn_3(p3_gdt).sigmoid()
2189
+ # >> Finally:
2190
+ # p3 = p3 * A_3^G
2191
+ p3 = p3 * gdt_attn_3
2192
+ _p3 = F.interpolate(p3, size=x2.shape[2:], mode='bilinear', align_corners=True)
2193
+ _p2 = _p3 + self.lateral_block3(x2)
2194
+
2195
+ if self.config.dec_ipt:
2196
+ patches_batch = self.get_patches_batch(x, _p2) if self.split else x
2197
+ _p2 = torch.cat((_p2, self.ipt_blk3(F.interpolate(patches_batch, size=x2.shape[2:], mode='bilinear', align_corners=True))), 1)
2198
+ p2 = self.decoder_block2(_p2)
2199
+ m2 = self.conv_ms_spvn_2(p2) if self.config.ms_supervision else None
2200
+ if self.config.out_ref:
2201
+ p2_gdt = self.gdt_convs_2(p2)
2202
+ if self.training:
2203
+ # >> GT:
2204
+ m2_dia = m2
2205
+ gdt_label_main_2 = gdt_gt * F.interpolate(m2_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
2206
+ outs_gdt_label.append(gdt_label_main_2)
2207
+ # >> Pred:
2208
+ gdt_pred_2 = self.gdt_convs_pred_2(p2_gdt)
2209
+ outs_gdt_pred.append(gdt_pred_2)
2210
+ gdt_attn_2 = self.gdt_convs_attn_2(p2_gdt).sigmoid()
2211
+ # >> Finally:
2212
+ p2 = p2 * gdt_attn_2
2213
+ _p2 = F.interpolate(p2, size=x1.shape[2:], mode='bilinear', align_corners=True)
2214
+ _p1 = _p2 + self.lateral_block2(x1)
2215
+
2216
+ if self.config.dec_ipt:
2217
+ patches_batch = self.get_patches_batch(x, _p1) if self.split else x
2218
+ _p1 = torch.cat((_p1, self.ipt_blk2(F.interpolate(patches_batch, size=x1.shape[2:], mode='bilinear', align_corners=True))), 1)
2219
+ _p1 = self.decoder_block1(_p1)
2220
+ _p1 = F.interpolate(_p1, size=x.shape[2:], mode='bilinear', align_corners=True)
2221
+
2222
+ if self.config.dec_ipt:
2223
+ patches_batch = self.get_patches_batch(x, _p1) if self.split else x
2224
+ _p1 = torch.cat((_p1, self.ipt_blk1(F.interpolate(patches_batch, size=x.shape[2:], mode='bilinear', align_corners=True))), 1)
2225
+ p1_out = self.conv_out1(_p1)
2226
+
2227
+ if self.config.ms_supervision:
2228
+ outs.append(m4)
2229
+ outs.append(m3)
2230
+ outs.append(m2)
2231
+ outs.append(p1_out)
2232
+ return outs if not (self.config.out_ref and self.training) else ([outs_gdt_pred, outs_gdt_label], outs)
2233
+
2234
+
2235
+ class SimpleConvs(nn.Module):
2236
+ def __init__(
2237
+ self, in_channels: int, out_channels: int, inter_channels=64
2238
+ ) -> None:
2239
+ super().__init__()
2240
+ self.conv1 = nn.Conv2d(in_channels, inter_channels, 3, 1, 1)
2241
+ self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, 1)
2242
+
2243
+ def forward(self, x):
2244
+ return self.conv_out(self.conv1(x))
config.json ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "ZhengPeng7/BiRefNet-w_codes",
3
+ "architectures": [
4
+ "BiRefNet"
5
+ ],
6
+ "auto_map": {
7
+ "AutoConfig": "BiRefNet_config.BiRefNetConfig",
8
+ "AutoModelForImageSegmentation": "birefnet.BiRefNet"
9
+ },
10
+ "custom_pipelines": {
11
+ "image-segmentation": {
12
+ "pt": [
13
+ "AutoModelForImageSegmentation"
14
+ ],
15
+ "tf": [],
16
+ "type": "image"
17
+ }
18
+ },
19
+ "bb_pretrained": false
20
+ }
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6c321deba6f15147f325566d0470e75e7258bf481f5e0b6c667659530d40019e
3
+ size 884878856