Upload folder using huggingface_hub

extras/birefnet/.gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

extras/birefnet/.gitignore

@@ -0,0 +1,142 @@
+# Custom
+e_*
+.vscode
+ckpt
+preds
+evaluation/eval-*
+nohup.out*
+tmp*
+*.pth
+core-*-python-*
+.DS_Store
+__MACOSX/
+
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+pip-wheel-metadata/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+.python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/

extras/birefnet/BiRefNet_config.py

@@ -0,0 +1,11 @@
+from transformers import PretrainedConfig
+
+class BiRefNetConfig(PretrainedConfig):
+    model_type = "SegformerForSemanticSegmentation"
+    def __init__(
+        self,
+        bb_pretrained=False,
+        **kwargs
+    ):
+        self.bb_pretrained = bb_pretrained
+        super().__init__(**kwargs)

extras/birefnet/README.md

@@ -0,0 +1,226 @@
+---
+library_name: birefnet
+tags:
+- background-removal
+- mask-generation
+- Dichotomous Image Segmentation
+- Camouflaged Object Detection
+- Salient Object Detection
+- pytorch_model_hub_mixin
+- model_hub_mixin
+- transformers
+repo_url: https://github.com/ZhengPeng7/BiRefNet
+pipeline_tag: image-segmentation
+license: mit
+---
+<h1 align="center">Bilateral Reference for High-Resolution Dichotomous Image Segmentation</h1>
+
+<div align='center'>
+    <a href='https://scholar.google.com/citations?user=TZRzWOsAAAAJ' target='_blank'><strong>Peng Zheng</strong></a><sup> 1,4,5,6</sup>,&thinsp;
+    <a href='https://scholar.google.com/citations?user=0uPb8MMAAAAJ' target='_blank'><strong>Dehong Gao</strong></a><sup> 2</sup>,&thinsp;
+    <a href='https://scholar.google.com/citations?user=kakwJ5QAAAAJ' target='_blank'><strong>Deng-Ping Fan</strong></a><sup> 1*</sup>,&thinsp;
+    <a href='https://scholar.google.com/citations?user=9cMQrVsAAAAJ' target='_blank'><strong>Li Liu</strong></a><sup> 3</sup>,&thinsp;
+    <a href='https://scholar.google.com/citations?user=qQP6WXIAAAAJ' target='_blank'><strong>Jorma Laaksonen</strong></a><sup> 4</sup>,&thinsp;
+    <a href='https://scholar.google.com/citations?user=pw_0Z_UAAAAJ' target='_blank'><strong>Wanli Ouyang</strong></a><sup> 5</sup>,&thinsp;
+    <a href='https://scholar.google.com/citations?user=stFCYOAAAAAJ' target='_blank'><strong>Nicu Sebe</strong></a><sup> 6</sup>
+</div>
+
+<div align='center'>
+    <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; 
+</div>
+
+<div align="center" style="display: flex; justify-content: center; flex-wrap: wrap;">
+  <a href='https://www.sciopen.com/article/pdf/10.26599/AIR.2024.9150038.pdf'><img src='https://img.shields.io/badge/Journal-Paper-red'></a>&ensp; 
+  <a href='https://arxiv.org/pdf/2401.03407'><img src='https://img.shields.io/badge/arXiv-BiRefNet-red'></a>&ensp; 
+  <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; 
+  <a href='https://www.birefnet.top'><img src='https://img.shields.io/badge/Page-BiRefNet-red'></a>&ensp; 
+  <a href='https://drive.google.com/drive/folders/1s2Xe0cjq-2ctnJBR24563yMSCOu4CcxM'><img src='https://img.shields.io/badge/Drive-Stuff-green'></a>&ensp; 
+  <a href='LICENSE'><img src='https://img.shields.io/badge/License-MIT-yellow'></a>&ensp; 
+  <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; 
+  <a href='https://huggingface.co/ZhengPeng7/BiRefNet'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20HF%20Models-BiRefNet-blue'></a>&ensp; 
+  <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; 
+  <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; 
+</div>
+
+
+|            *DIS-Sample_1*        |             *DIS-Sample_2*        |
+| :------------------------------: | :-------------------------------: |
+| <img src="https://drive.google.com/thumbnail?id=1ItXaA26iYnE8XQ_GgNLy71MOWePoS2-g&sz=w400" /> |  <img src="https://drive.google.com/thumbnail?id=1Z-esCujQF_uEa_YJjkibc3NUrW4aR_d4&sz=w400" /> |
+
+This repo is the official implementation of "[**Bilateral Reference for High-Resolution Dichotomous Image Segmentation**](https://arxiv.org/pdf/2401.03407.pdf)" (___CAAI AIR 2024___). 
+
+Visit our GitHub repo: [https://github.com/ZhengPeng7/BiRefNet](https://github.com/ZhengPeng7/BiRefNet) for more details -- **codes**, **docs**, and **model zoo**!
+
+## How to use
+
+### 0. Install Packages:
+```
+pip install -qr https://raw.githubusercontent.com/ZhengPeng7/BiRefNet/main/requirements.txt
+```
+
+### 1. Load BiRefNet:
+
+#### Use codes + weights from HuggingFace
+> Only use the weights on HuggingFace -- Pro: No need to download BiRefNet codes manually; Con: Codes on HuggingFace might not be latest version (I'll try to keep them always latest).
+
+```python
+# Load BiRefNet with weights
+from transformers import AutoModelForImageSegmentation
+birefnet = AutoModelForImageSegmentation.from_pretrained('ZhengPeng7/BiRefNet', trust_remote_code=True)
+```
+
+#### Use codes from GitHub + weights from HuggingFace
+> Only use the weights on HuggingFace -- Pro: codes are always latest; Con: Need to clone the BiRefNet repo from my GitHub.
+
+```shell
+# Download codes
+git clone https://github.com/ZhengPeng7/BiRefNet.git
+cd BiRefNet
+```
+
+```python
+# Use codes locally
+from models.birefnet import BiRefNet
+
+# Load weights from Hugging Face Models
+birefnet = BiRefNet.from_pretrained('ZhengPeng7/BiRefNet')
+```
+
+#### Use codes from GitHub + weights from local space
+> Only use the weights and codes both locally.
+
+```python
+# Use codes and weights locally
+import torch
+from utils import check_state_dict
+
+birefnet = BiRefNet(bb_pretrained=False)
+state_dict = torch.load(PATH_TO_WEIGHT, map_location='cpu')
+state_dict = check_state_dict(state_dict)
+birefnet.load_state_dict(state_dict)
+```
+
+#### Use the loaded BiRefNet for inference
+```python
+# Imports
+from PIL import Image
+import matplotlib.pyplot as plt
+import torch
+from torchvision import transforms
+from models.birefnet import BiRefNet
+
+birefnet = ... # -- BiRefNet should be loaded with codes above, either way.
+torch.set_float32_matmul_precision(['high', 'highest'][0])
+birefnet.to('cuda')
+birefnet.eval()
+birefnet.half()
+
+def extract_object(birefnet, imagepath):
+    # Data settings
+    image_size = (1024, 1024)
+    transform_image = transforms.Compose([
+        transforms.Resize(image_size),
+        transforms.ToTensor(),
+        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
+    ])
+
+    image = Image.open(imagepath)
+    input_images = transform_image(image).unsqueeze(0).to('cuda').half()
+
+    # Prediction
+    with torch.no_grad():
+        preds = birefnet(input_images)[-1].sigmoid().cpu()
+    pred = preds[0].squeeze()
+    pred_pil = transforms.ToPILImage()(pred)
+    mask = pred_pil.resize(image.size)
+    image.putalpha(mask)
+    return image, mask
+
+# Visualization
+plt.axis("off")
+plt.imshow(extract_object(birefnet, imagepath='PATH-TO-YOUR_IMAGE.jpg')[0])
+plt.show()
+
+```
+
+### 2. Use inference endpoint locally:
+> You may need to click the *deploy* and set up the endpoint by yourself, which would make some costs.
+```
+import requests
+import base64
+from io import BytesIO
+from PIL import Image
+
+
+YOUR_HF_TOKEN = 'xxx'
+API_URL = "xxx"
+headers = {
+	"Authorization": "Bearer {}".format(YOUR_HF_TOKEN)
+}
+
+def base64_to_bytes(base64_string):
+    # Remove the data URI prefix if present
+    if "data:image" in base64_string:
+        base64_string = base64_string.split(",")[1]
+
+    # Decode the Base64 string into bytes
+    image_bytes = base64.b64decode(base64_string)
+    return image_bytes
+
+def bytes_to_base64(image_bytes):
+    # Create a BytesIO object to handle the image data
+    image_stream = BytesIO(image_bytes)
+
+    # Open the image using Pillow (PIL)
+    image = Image.open(image_stream)
+    return image
+
+def query(payload):
+	response = requests.post(API_URL, headers=headers, json=payload)
+	return response.json()
+
+output = query({
+	"inputs": "https://hips.hearstapps.com/hmg-prod/images/gettyimages-1229892983-square.jpg",
+	"parameters": {}
+})
+
+output_image = bytes_to_base64(base64_to_bytes(output))
+output_image
+```
+
+
+> This BiRefNet for standard dichotomous image segmentation (DIS) is trained on **DIS-TR** and validated on **DIS-TEs and DIS-VD**.
+
+## This repo holds the official model weights of "[<ins>Bilateral Reference for High-Resolution Dichotomous Image Segmentation</ins>](https://arxiv.org/pdf/2401.03407)" (_CAAI AIR 2024_).
+
+This repo contains the weights of BiRefNet proposed in our paper, which has achieved the SOTA performance on three tasks (DIS, HRSOD, and COD).
+
+Go to my GitHub page for BiRefNet codes and the latest updates: https://github.com/ZhengPeng7/BiRefNet :)
+
+
+#### Try our online demos for inference:
+
++ Online **Image Inference** on Colab: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/14Dqg7oeBkFEtchaHLNpig2BcdkZEogba?usp=drive_link)
++ **Online Inference with GUI on Hugging Face** with adjustable resolutions: [![Hugging Face Spaces](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue)](https://huggingface.co/spaces/ZhengPeng7/BiRefNet_demo)  
++ **Inference and evaluation** of your given weights: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1MaEiBfJ4xIaZZn0DqKrhydHB8X97hNXl#scrollTo=DJ4meUYjia6S)
+<img src="https://drive.google.com/thumbnail?id=12XmDhKtO1o2fEvBu4OE4ULVB2BK0ecWi&sz=w1080" />
+
+## Acknowledgement:
+
++ Many thanks to @Freepik for their generous support on GPU resources for training higher resolution BiRefNet models and more of my explorations.
++ Many thanks to @fal for their generous support on GPU resources for training better general BiRefNet models.
++ Many thanks to @not-lain for his help on the better deployment of our BiRefNet model on HuggingFace.
+
+
+## Citation
+
+```
+@article{zheng2024birefnet,
+  title={Bilateral Reference for High-Resolution Dichotomous Image Segmentation},
+  author={Zheng, Peng and Gao, Dehong and Fan, Deng-Ping and Liu, Li and Laaksonen, Jorma and Ouyang, Wanli and Sebe, Nicu},
+  journal={CAAI Artificial Intelligence Research},
+  volume = {3},
+  pages = {9150038},
+  year={2024}
+}
+```

extras/birefnet/birefnet.py

@@ -0,0 +1,2248 @@
+### config.py
+
+import os
+import math
+from transformers import PretrainedConfig
+
+
+class Config(PretrainedConfig):
+    def __init__(self) -> None:
+        # Compatible with the latest version of transformers
+        super().__init__()
+
+        # PATH settings
+        self.sys_home_dir = os.path.expanduser('~')     # Make up your file system as: SYS_HOME_DIR/codes/dis/BiRefNet, SYS_HOME_DIR/datasets/dis/xx, SYS_HOME_DIR/weights/xx
+
+        # TASK settings
+        self.task = ['DIS5K', 'COD', 'HRSOD', 'DIS5K+HRSOD+HRS10K', 'P3M-10k'][0]
+        self.training_set = {
+            'DIS5K': ['DIS-TR', 'DIS-TR+DIS-TE1+DIS-TE2+DIS-TE3+DIS-TE4'][0],
+            'COD': 'TR-COD10K+TR-CAMO',
+            '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],
+            '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.
+            'P3M-10k': 'TR-P3M-10k',
+        }[self.task]
+        self.prompt4loc = ['dense', 'sparse'][0]
+
+        # Faster-Training settings
+        self.load_all = True
+        self.compile = True     # 1. Trigger CPU memory leak in some extend, which is an inherent problem of PyTorch.
+                                #   Machines with > 70GB CPU memory can run the whole training on DIS5K with default setting.
+                                # 2. Higher PyTorch version may fix it: https://github.com/pytorch/pytorch/issues/119607.
+                                # 3. But compile in Pytorch > 2.0.1 seems to bring no acceleration for training.
+        self.precisionHigh = True
+
+        # MODEL settings
+        self.ms_supervision = True
+        self.out_ref = self.ms_supervision and True
+        self.dec_ipt = True
+        self.dec_ipt_split = True
+        self.cxt_num = [0, 3][1]    # multi-scale skip connections from encoder
+        self.mul_scl_ipt = ['', 'add', 'cat'][2]
+        self.dec_att = ['', 'ASPP', 'ASPPDeformable'][2]
+        self.squeeze_block = ['', 'BasicDecBlk_x1', 'ResBlk_x4', 'ASPP_x3', 'ASPPDeformable_x3'][1]
+        self.dec_blk = ['BasicDecBlk', 'ResBlk', 'HierarAttDecBlk'][0]
+
+        # TRAINING settings
+        self.batch_size = 4
+        self.IoU_finetune_last_epochs = [
+            0,
+            {
+                'DIS5K': -50,
+                'COD': -20,
+                'HRSOD': -20,
+                'DIS5K+HRSOD+HRS10K': -20,
+                'P3M-10k': -20,
+            }[self.task]
+        ][1]    # choose 0 to skip
+        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
+        self.size = 1024
+        self.num_workers = max(4, self.batch_size)          # will be decrease to min(it, batch_size) at the initialization of the data_loader
+
+        # Backbone settings
+        self.bb = [
+            'vgg16', 'vgg16bn', 'resnet50',         # 0, 1, 2
+            'swin_v1_t', 'swin_v1_s',               # 3, 4
+            'swin_v1_b', 'swin_v1_l',               # 5-bs9, 6-bs4
+            'pvt_v2_b0', 'pvt_v2_b1',               # 7, 8
+            'pvt_v2_b2', 'pvt_v2_b5',               # 9-bs10, 10-bs5
+        ][6]
+        self.lateral_channels_in_collection = {
+            'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
+            'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
+            'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
+            'swin_v1_t': [768, 384, 192, 96], 'swin_v1_s': [768, 384, 192, 96],
+            'pvt_v2_b0': [256, 160, 64, 32], 'pvt_v2_b1': [512, 320, 128, 64],
+        }[self.bb]
+        if self.mul_scl_ipt == 'cat':
+            self.lateral_channels_in_collection = [channel * 2 for channel in self.lateral_channels_in_collection]
+        self.cxt = self.lateral_channels_in_collection[1:][::-1][-self.cxt_num:] if self.cxt_num else []
+
+        # MODEL settings - inactive
+        self.lat_blk = ['BasicLatBlk'][0]
+        self.dec_channels_inter = ['fixed', 'adap'][0]
+        self.refine = ['', 'itself', 'RefUNet', 'Refiner', 'RefinerPVTInChannels4'][0]
+        self.progressive_ref = self.refine and True
+        self.ender = self.progressive_ref and False
+        self.scale = self.progressive_ref and 2
+        self.auxiliary_classification = False       # Only for DIS5K, where class labels are saved in `dataset.py`.
+        self.refine_iteration = 1
+        self.freeze_bb = False
+        self.model = [
+            'BiRefNet',
+        ][0]
+        if self.dec_blk == 'HierarAttDecBlk':
+            self.batch_size = 2 ** [0, 1, 2, 3, 4][2]
+
+        # TRAINING settings - inactive
+        self.preproc_methods = ['flip', 'enhance', 'rotate', 'pepper', 'crop'][:4]
+        self.optimizer = ['Adam', 'AdamW'][1]
+        self.lr_decay_epochs = [1e5]    # Set to negative N to decay the lr in the last N-th epoch.
+        self.lr_decay_rate = 0.5
+        # Loss
+        self.lambdas_pix_last = {
+            # not 0 means opening this loss
+            # original rate -- 1 : 30 : 1.5 : 0.2, bce x 30
+            'bce': 30 * 1,          # high performance
+            'iou': 0.5 * 1,         # 0 / 255
+            'iou_patch': 0.5 * 0,   # 0 / 255, win_size = (64, 64)
+            'mse': 150 * 0,         # can smooth the saliency map
+            'triplet': 3 * 0,
+            'reg': 100 * 0,
+            'ssim': 10 * 1,          # help contours,
+            'cnt': 5 * 0,          # help contours
+            '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.
+        }
+        self.lambdas_cls = {
+            'ce': 5.0
+        }
+        # Adv
+        self.lambda_adv_g = 10. * 0        # turn to 0 to avoid adv training
+        self.lambda_adv_d = 3. * (self.lambda_adv_g > 0)
+
+        # PATH settings - inactive
+        self.data_root_dir = os.path.join(self.sys_home_dir, 'datasets/dis')
+        self.weights_root_dir = os.path.join(self.sys_home_dir, 'weights')
+        self.weights = {
+            'pvt_v2_b2': os.path.join(self.weights_root_dir, 'pvt_v2_b2.pth'),
+            'pvt_v2_b5': os.path.join(self.weights_root_dir, ['pvt_v2_b5.pth', 'pvt_v2_b5_22k.pth'][0]),
+            '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]),
+            '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]),
+            'swin_v1_t': os.path.join(self.weights_root_dir, ['swin_tiny_patch4_window7_224_22kto1k_finetune.pth'][0]),
+            'swin_v1_s': os.path.join(self.weights_root_dir, ['swin_small_patch4_window7_224_22kto1k_finetune.pth'][0]),
+            'pvt_v2_b0': os.path.join(self.weights_root_dir, ['pvt_v2_b0.pth'][0]),
+            'pvt_v2_b1': os.path.join(self.weights_root_dir, ['pvt_v2_b1.pth'][0]),
+        }
+
+        # Callbacks - inactive
+        self.verbose_eval = True
+        self.only_S_MAE = False
+        self.use_fp16 = False   # Bugs. It may cause nan in training.
+        self.SDPA_enabled = False    # Bugs. Slower and errors occur in multi-GPUs
+
+        # others
+        self.device = [0, 'cpu'][0]     # .to(0) == .to('cuda:0')
+
+        self.batch_size_valid = 1
+        self.rand_seed = 7
+        # 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]
+        # with open(run_sh_file[0], 'r') as f:
+        #     lines = f.readlines()
+        #     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])
+        #     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])
+        # self.val_step = [0, self.save_step][0]
+
+    def print_task(self) -> None:
+        # Return task for choosing settings in shell scripts.
+        print(self.task)
+
+
+
+### models/backbones/pvt_v2.py
+
+import torch
+import torch.nn as nn
+from functools import partial
+
+from timm.layers import DropPath, to_2tuple, trunc_normal_
+
+
+import math
+
+# from config import Config
+
+# config = Config()
+
+class Mlp(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.dwconv = DWConv(hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x, H, W):
+        x = self.fc1(x)
+        x = self.dwconv(x, H, W)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., sr_ratio=1):
+        super().__init__()
+        assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
+
+        self.dim = dim
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.q = nn.Linear(dim, dim, bias=qkv_bias)
+        self.kv = nn.Linear(dim, dim * 2, bias=qkv_bias)
+        self.attn_drop_prob = attn_drop
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        self.sr_ratio = sr_ratio
+        if sr_ratio > 1:
+            self.sr = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)
+            self.norm = nn.LayerNorm(dim)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x, H, W):
+        B, N, C = x.shape
+        q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
+
+        if self.sr_ratio > 1:
+            x_ = x.permute(0, 2, 1).reshape(B, C, H, W)
+            x_ = self.sr(x_).reshape(B, C, -1).permute(0, 2, 1)
+            x_ = self.norm(x_)
+            kv = self.kv(x_).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        else:
+            kv = self.kv(x).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        k, v = kv[0], kv[1]
+
+        if config.SDPA_enabled:
+            x = torch.nn.functional.scaled_dot_product_attention(
+                q, k, v,
+                attn_mask=None, dropout_p=self.attn_drop_prob, is_causal=False
+            ).transpose(1, 2).reshape(B, N, C)
+        else:
+            attn = (q @ k.transpose(-2, -1)) * self.scale
+            attn = attn.softmax(dim=-1)
+            attn = self.attn_drop(attn)
+
+            x = (attn @ v).transpose(1, 2).reshape(B, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+
+        return x
+
+
+class Block(nn.Module):
+
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, sr_ratio=1):
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim,
+            num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            attn_drop=attn_drop, proj_drop=drop, sr_ratio=sr_ratio)
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x, H, W):
+        x = x + self.drop_path(self.attn(self.norm1(x), H, W))
+        x = x + self.drop_path(self.mlp(self.norm2(x), H, W))
+
+        return x
+
+
+class OverlapPatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+    """
+
+    def __init__(self, img_size=224, patch_size=7, stride=4, in_channels=3, embed_dim=768):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.H, self.W = img_size[0] // patch_size[0], img_size[1] // patch_size[1]
+        self.num_patches = self.H * self.W
+        self.proj = nn.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=stride,
+                              padding=(patch_size[0] // 2, patch_size[1] // 2))
+        self.norm = nn.LayerNorm(embed_dim)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x):
+        x = self.proj(x)
+        _, _, H, W = x.shape
+        x = x.flatten(2).transpose(1, 2)
+        x = self.norm(x)
+
+        return x, H, W
+
+
+class PyramidVisionTransformerImpr(nn.Module):
+    def __init__(self, img_size=224, patch_size=16, in_channels=3, num_classes=1000, embed_dims=[64, 128, 256, 512],
+                 num_heads=[1, 2, 4, 8], mlp_ratios=[4, 4, 4, 4], qkv_bias=False, qk_scale=None, drop_rate=0.,
+                 attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm,
+                 depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1]):
+        super().__init__()
+        self.num_classes = num_classes
+        self.depths = depths
+
+        # patch_embed
+        self.patch_embed1 = OverlapPatchEmbed(img_size=img_size, patch_size=7, stride=4, in_channels=in_channels,
+                                              embed_dim=embed_dims[0])
+        self.patch_embed2 = OverlapPatchEmbed(img_size=img_size // 4, patch_size=3, stride=2, in_channels=embed_dims[0],
+                                              embed_dim=embed_dims[1])
+        self.patch_embed3 = OverlapPatchEmbed(img_size=img_size // 8, patch_size=3, stride=2, in_channels=embed_dims[1],
+                                              embed_dim=embed_dims[2])
+        self.patch_embed4 = OverlapPatchEmbed(img_size=img_size // 16, patch_size=3, stride=2, in_channels=embed_dims[2],
+                                              embed_dim=embed_dims[3])
+
+        # transformer encoder
+        dpr = np.linspace(0, drop_path_rate, sum(depths)).tolist() # stochastic depth decay rule
+        cur = 0
+        self.block1 = nn.ModuleList([Block(
+            dim=embed_dims[0], num_heads=num_heads[0], mlp_ratio=mlp_ratios[0], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[0])
+            for i in range(depths[0])])
+        self.norm1 = norm_layer(embed_dims[0])
+
+        cur += depths[0]
+        self.block2 = nn.ModuleList([Block(
+            dim=embed_dims[1], num_heads=num_heads[1], mlp_ratio=mlp_ratios[1], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[1])
+            for i in range(depths[1])])
+        self.norm2 = norm_layer(embed_dims[1])
+
+        cur += depths[1]
+        self.block3 = nn.ModuleList([Block(
+            dim=embed_dims[2], num_heads=num_heads[2], mlp_ratio=mlp_ratios[2], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[2])
+            for i in range(depths[2])])
+        self.norm3 = norm_layer(embed_dims[2])
+
+        cur += depths[2]
+        self.block4 = nn.ModuleList([Block(
+            dim=embed_dims[3], num_heads=num_heads[3], mlp_ratio=mlp_ratios[3], qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
+            sr_ratio=sr_ratios[3])
+            for i in range(depths[3])])
+        self.norm4 = norm_layer(embed_dims[3])
+
+        # classification head
+        # self.head = nn.Linear(embed_dims[3], num_classes) if num_classes > 0 else nn.Identity()
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def init_weights(self, pretrained=None):
+        if isinstance(pretrained, str):
+            logger = 1
+            #load_checkpoint(self, pretrained, map_location='cpu', strict=False, logger=logger)
+
+    def reset_drop_path(self, drop_path_rate):
+        dpr = np.linspace(0, drop_path_rate, sum(self.depths)).tolist()
+        cur = 0
+        for i in range(self.depths[0]):
+            self.block1[i].drop_path.drop_prob = dpr[cur + i]
+
+        cur += self.depths[0]
+        for i in range(self.depths[1]):
+            self.block2[i].drop_path.drop_prob = dpr[cur + i]
+
+        cur += self.depths[1]
+        for i in range(self.depths[2]):
+            self.block3[i].drop_path.drop_prob = dpr[cur + i]
+
+        cur += self.depths[2]
+        for i in range(self.depths[3]):
+            self.block4[i].drop_path.drop_prob = dpr[cur + i]
+
+    def freeze_patch_emb(self):
+        self.patch_embed1.requires_grad = False
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'pos_embed1', 'pos_embed2', 'pos_embed3', 'pos_embed4', 'cls_token'}  # has pos_embed may be better
+
+    def get_classifier(self):
+        return self.head
+
+    def reset_classifier(self, num_classes, global_pool=''):
+        self.num_classes = num_classes
+        self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
+
+    def forward_features(self, x):
+        B = x.shape[0]
+        outs = []
+
+        # stage 1
+        x, H, W = self.patch_embed1(x)
+        for i, blk in enumerate(self.block1):
+            x = blk(x, H, W)
+        x = self.norm1(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        # stage 2
+        x, H, W = self.patch_embed2(x)
+        for i, blk in enumerate(self.block2):
+            x = blk(x, H, W)
+        x = self.norm2(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        # stage 3
+        x, H, W = self.patch_embed3(x)
+        for i, blk in enumerate(self.block3):
+            x = blk(x, H, W)
+        x = self.norm3(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        # stage 4
+        x, H, W = self.patch_embed4(x)
+        for i, blk in enumerate(self.block4):
+            x = blk(x, H, W)
+        x = self.norm4(x)
+        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+        outs.append(x)
+
+        return outs
+
+        # return x.mean(dim=1)
+
+    def forward(self, x):
+        x = self.forward_features(x)
+        # x = self.head(x)
+
+        return x
+
+
+class DWConv(nn.Module):
+    def __init__(self, dim=768):
+        super(DWConv, self).__init__()
+        self.dwconv = nn.Conv2d(dim, dim, 3, 1, 1, bias=True, groups=dim)
+
+    def forward(self, x, H, W):
+        B, N, C = x.shape
+        x = x.transpose(1, 2).view(B, C, H, W).contiguous()
+        x = self.dwconv(x)
+        x = x.flatten(2).transpose(1, 2)
+
+        return x
+
+
+def _conv_filter(state_dict, patch_size=16):
+    """ convert patch embedding weight from manual patchify + linear proj to conv"""
+    out_dict = {}
+    for k, v in state_dict.items():
+        if 'patch_embed.proj.weight' in k:
+            v = v.reshape((v.shape[0], 3, patch_size, patch_size))
+        out_dict[k] = v
+
+    return out_dict
+
+
+class pvt_v2_b0(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b0, self).__init__(
+            patch_size=4, embed_dims=[32, 64, 160, 256], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+
+
+class pvt_v2_b1(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b1, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+class pvt_v2_b2(PyramidVisionTransformerImpr):
+    def __init__(self, in_channels=3, **kwargs):
+        super(pvt_v2_b2, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1, in_channels=in_channels)
+
+class pvt_v2_b3(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b3, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 18, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+class pvt_v2_b4(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b4, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 8, 27, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+
+class pvt_v2_b5(PyramidVisionTransformerImpr):
+    def __init__(self, **kwargs):
+        super(pvt_v2_b5, self).__init__(
+            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],
+            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 6, 40, 3], sr_ratios=[8, 4, 2, 1],
+            drop_rate=0.0, drop_path_rate=0.1)
+
+
+
+### models/backbones/swin_v1.py
+
+# --------------------------------------------------------
+# Swin Transformer
+# Copyright (c) 2021 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Ze Liu, Yutong Lin, Yixuan Wei
+# --------------------------------------------------------
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+import numpy as np
+from timm.layers import DropPath, to_2tuple, trunc_normal_
+
+# from config import Config
+
+
+# config = Config()
+
+
+class Mlp(nn.Module):
+    """ Multilayer perceptron."""
+
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+def window_partition(x, window_size):
+    """
+    Args:
+        x: (B, H, W, C)
+        window_size (int): window size
+
+    Returns:
+        windows: (num_windows*B, window_size, window_size, C)
+    """
+    B, H, W, C = x.shape
+    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows
+
+
+def window_reverse(windows, window_size, H, W):
+    """
+    Args:
+        windows: (num_windows*B, window_size, window_size, C)
+        window_size (int): Window size
+        H (int): Height of image
+        W (int): Width of image
+
+    Returns:
+        x: (B, H, W, C)
+    """
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+
+class WindowAttention(nn.Module):
+    """ Window based multi-head self attention (W-MSA) module with relative position bias.
+    It supports both of shifted and non-shifted window.
+
+    Args:
+        dim (int): Number of input channels.
+        window_size (tuple[int]): The height and width of the window.
+        num_heads (int): Number of attention heads.
+        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
+        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
+        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
+    """
+
+    def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        # define a parameter table of relative position bias
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(self.window_size[0])
+        coords_w = torch.arange(self.window_size[1])
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w], indexing='ij'))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop_prob = attn_drop
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        trunc_normal_(self.relative_position_bias_table, std=.02)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask=None):
+        """ Forward function.
+
+        Args:
+            x: input features with shape of (num_windows*B, N, C)
+            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
+        """
+        B_, N, C = x.shape
+        qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+
+        if config.SDPA_enabled:
+            x = torch.nn.functional.scaled_dot_product_attention(
+                q, k, v,
+                attn_mask=None, dropout_p=self.attn_drop_prob, is_causal=False
+            ).transpose(1, 2).reshape(B_, N, C)
+        else:
+            attn = (q @ k.transpose(-2, -1))
+
+            relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
+                self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1
+            )   # Wh*Ww, Wh*Ww, nH
+            relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+            attn = attn + relative_position_bias.unsqueeze(0)
+
+            if mask is not None:
+                nW = mask.shape[0]
+                attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
+                attn = attn.view(-1, self.num_heads, N, N)
+                attn = self.softmax(attn)
+            else:
+                attn = self.softmax(attn)
+
+            attn = self.attn_drop(attn)
+
+            x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+class SwinTransformerBlock(nn.Module):
+    """ Swin Transformer Block.
+
+    Args:
+        dim (int): Number of input channels.
+        num_heads (int): Number of attention heads.
+        window_size (int): Window size.
+        shift_size (int): Shift size for SW-MSA.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float, optional): Stochastic depth rate. Default: 0.0
+        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, dim, num_heads, window_size=7, shift_size=0,
+                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
+                 act_layer=nn.GELU, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
+
+        self.norm1 = norm_layer(dim)
+        self.attn = WindowAttention(
+            dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
+            qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        self.H = None
+        self.W = None
+
+    def forward(self, x, mask_matrix):
+        """ Forward function.
+
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+            mask_matrix: Attention mask for cyclic shift.
+        """
+        B, L, C = x.shape
+        H, W = self.H, self.W
+        assert L == H * W, "input feature has wrong size"
+
+        shortcut = x
+        x = self.norm1(x)
+        x = x.view(B, H, W, C)
+
+        # pad feature maps to multiples of window size
+        pad_l = pad_t = 0
+        pad_r = (self.window_size - W % self.window_size) % self.window_size
+        pad_b = (self.window_size - H % self.window_size) % self.window_size
+        x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
+        _, Hp, Wp, _ = x.shape
+
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
+            attn_mask = mask_matrix
+        else:
+            shifted_x = x
+            attn_mask = None
+
+        # partition windows
+        x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
+        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C
+
+        # W-MSA/SW-MSA
+        attn_windows = self.attn(x_windows, mask=attn_mask)  # nW*B, window_size*window_size, C
+
+        # merge windows
+        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
+        shifted_x = window_reverse(attn_windows, self.window_size, Hp, Wp)  # B H' W' C
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
+        else:
+            x = shifted_x
+
+        if pad_r > 0 or pad_b > 0:
+            x = x[:, :H, :W, :].contiguous()
+
+        x = x.view(B, H * W, C)
+
+        # FFN
+        x = shortcut + self.drop_path(x)
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+
+        return x
+
+
+class PatchMerging(nn.Module):
+    """ Patch Merging Layer
+
+    Args:
+        dim (int): Number of input channels.
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+    def __init__(self, dim, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.dim = dim
+        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
+        self.norm = norm_layer(4 * dim)
+
+    def forward(self, x, H, W):
+        """ Forward function.
+
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+        """
+        B, L, C = x.shape
+        assert L == H * W, "input feature has wrong size"
+
+        x = x.view(B, H, W, C)
+
+        # padding
+        pad_input = (H % 2 == 1) or (W % 2 == 1)
+        if pad_input:
+            x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2))
+
+        x0 = x[:, 0::2, 0::2, :]  # B H/2 W/2 C
+        x1 = x[:, 1::2, 0::2, :]  # B H/2 W/2 C
+        x2 = x[:, 0::2, 1::2, :]  # B H/2 W/2 C
+        x3 = x[:, 1::2, 1::2, :]  # B H/2 W/2 C
+        x = torch.cat([x0, x1, x2, x3], -1)  # B H/2 W/2 4*C
+        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C
+
+        x = self.norm(x)
+        x = self.reduction(x)
+
+        return x
+
+
+class BasicLayer(nn.Module):
+    """ A basic Swin Transformer layer for one stage.
+
+    Args:
+        dim (int): Number of feature channels
+        depth (int): Depths of this stage.
+        num_heads (int): Number of attention head.
+        window_size (int): Local window size. Default: 7.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(self,
+                 dim,
+                 depth,
+                 num_heads,
+                 window_size=7,
+                 mlp_ratio=4.,
+                 qkv_bias=True,
+                 qk_scale=None,
+                 drop=0.,
+                 attn_drop=0.,
+                 drop_path=0.,
+                 norm_layer=nn.LayerNorm,
+                 downsample=None,
+                 use_checkpoint=False):
+        super().__init__()
+        self.window_size = window_size
+        self.shift_size = window_size // 2
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+
+        # build blocks
+        self.blocks = nn.ModuleList([
+            SwinTransformerBlock(
+                dim=dim,
+                num_heads=num_heads,
+                window_size=window_size,
+                shift_size=0 if (i % 2 == 0) else window_size // 2,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                drop=drop,
+                attn_drop=attn_drop,
+                drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
+                norm_layer=norm_layer)
+            for i in range(depth)])
+
+        # patch merging layer
+        if downsample is not None:
+            self.downsample = downsample(dim=dim, norm_layer=norm_layer)
+        else:
+            self.downsample = None
+
+    def forward(self, x, H, W):
+        """ Forward function.
+
+        Args:
+            x: Input feature, tensor size (B, H*W, C).
+            H, W: Spatial resolution of the input feature.
+        """
+
+        # calculate attention mask for SW-MSA
+        # Turn int to torch.tensor for the compatiability with torch.compile in PyTorch 2.5.
+        Hp = torch.ceil(torch.tensor(H) / self.window_size).to(torch.int64) * self.window_size
+        Wp = torch.ceil(torch.tensor(W) / self.window_size).to(torch.int64) * self.window_size
+        img_mask = torch.zeros((1, Hp, Wp, 1), device=x.device)  # 1 Hp Wp 1
+        h_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        w_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        cnt = 0
+        for h in h_slices:
+            for w in w_slices:
+                img_mask[:, h, w, :] = cnt
+                cnt += 1
+
+        mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
+        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0)).to(x.dtype)
+
+        for blk in self.blocks:
+            blk.H, blk.W = H, W
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x, attn_mask)
+            else:
+                x = blk(x, attn_mask)
+        if self.downsample is not None:
+            x_down = self.downsample(x, H, W)
+            Wh, Ww = (H + 1) // 2, (W + 1) // 2
+            return x, H, W, x_down, Wh, Ww
+        else:
+            return x, H, W, x, H, W
+
+
+class PatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+
+    Args:
+        patch_size (int): Patch token size. Default: 4.
+        in_channels (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        norm_layer (nn.Module, optional): Normalization layer. Default: None
+    """
+
+    def __init__(self, patch_size=4, in_channels=3, embed_dim=96, norm_layer=None):
+        super().__init__()
+        patch_size = to_2tuple(patch_size)
+        self.patch_size = patch_size
+
+        self.in_channels = in_channels
+        self.embed_dim = embed_dim
+
+        self.proj = nn.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=patch_size)
+        if norm_layer is not None:
+            self.norm = norm_layer(embed_dim)
+        else:
+            self.norm = None
+
+    def forward(self, x):
+        """Forward function."""
+        # padding
+        _, _, H, W = x.size()
+        if W % self.patch_size[1] != 0:
+            x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1]))
+        if H % self.patch_size[0] != 0:
+            x = F.pad(x, (0, 0, 0, self.patch_size[0] - H % self.patch_size[0]))
+
+        x = self.proj(x)  # B C Wh Ww
+        if self.norm is not None:
+            Wh, Ww = x.size(2), x.size(3)
+            x = x.flatten(2).transpose(1, 2)
+            x = self.norm(x)
+            x = x.transpose(1, 2).view(-1, self.embed_dim, Wh, Ww)
+
+        return x
+
+
+class SwinTransformer(nn.Module):
+    """ Swin Transformer backbone.
+        A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows`  -
+          https://arxiv.org/pdf/2103.14030
+
+    Args:
+        pretrain_img_size (int): Input image size for training the pretrained model,
+            used in absolute postion embedding. Default 224.
+        patch_size (int | tuple(int)): Patch size. Default: 4.
+        in_channels (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        depths (tuple[int]): Depths of each Swin Transformer stage.
+        num_heads (tuple[int]): Number of attention head of each stage.
+        window_size (int): Window size. Default: 7.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float): Override default qk scale of head_dim ** -0.5 if set.
+        drop_rate (float): Dropout rate.
+        attn_drop_rate (float): Attention dropout rate. Default: 0.
+        drop_path_rate (float): Stochastic depth rate. Default: 0.2.
+        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
+        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False.
+        patch_norm (bool): If True, add normalization after patch embedding. Default: True.
+        out_indices (Sequence[int]): Output from which stages.
+        frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
+            -1 means not freezing any parameters.
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(self,
+                 pretrain_img_size=224,
+                 patch_size=4,
+                 in_channels=3,
+                 embed_dim=96,
+                 depths=[2, 2, 6, 2],
+                 num_heads=[3, 6, 12, 24],
+                 window_size=7,
+                 mlp_ratio=4.,
+                 qkv_bias=True,
+                 qk_scale=None,
+                 drop_rate=0.,
+                 attn_drop_rate=0.,
+                 drop_path_rate=0.2,
+                 norm_layer=nn.LayerNorm,
+                 ape=False,
+                 patch_norm=True,
+                 out_indices=(0, 1, 2, 3),
+                 frozen_stages=-1,
+                 use_checkpoint=False):
+        super().__init__()
+
+        self.pretrain_img_size = pretrain_img_size
+        self.num_layers = len(depths)
+        self.embed_dim = embed_dim
+        self.ape = ape
+        self.patch_norm = patch_norm
+        self.out_indices = out_indices
+        self.frozen_stages = frozen_stages
+
+        # split image into non-overlapping patches
+        self.patch_embed = PatchEmbed(
+            patch_size=patch_size, in_channels=in_channels, embed_dim=embed_dim,
+            norm_layer=norm_layer if self.patch_norm else None)
+
+        # absolute position embedding
+        if self.ape:
+            pretrain_img_size = to_2tuple(pretrain_img_size)
+            patch_size = to_2tuple(patch_size)
+            patches_resolution = [pretrain_img_size[0] // patch_size[0], pretrain_img_size[1] // patch_size[1]]
+
+            self.absolute_pos_embed = nn.Parameter(torch.zeros(1, embed_dim, patches_resolution[0], patches_resolution[1]))
+            trunc_normal_(self.absolute_pos_embed, std=.02)
+
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        # stochastic depth
+        dpr = np.linspace(0, drop_path_rate, sum(depths)).tolist() # stochastic depth decay rule
+
+        # build layers
+        self.layers = nn.ModuleList()
+        for i_layer in range(self.num_layers):
+            layer = BasicLayer(
+                dim=int(embed_dim * 2 ** i_layer),
+                depth=depths[i_layer],
+                num_heads=num_heads[i_layer],
+                window_size=window_size,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                drop=drop_rate,
+                attn_drop=attn_drop_rate,
+                drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],
+                norm_layer=norm_layer,
+                downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
+                use_checkpoint=use_checkpoint)
+            self.layers.append(layer)
+
+        num_features = [int(embed_dim * 2 ** i) for i in range(self.num_layers)]
+        self.num_features = num_features
+
+        # add a norm layer for each output
+        for i_layer in out_indices:
+            layer = norm_layer(num_features[i_layer])
+            layer_name = f'norm{i_layer}'
+            self.add_module(layer_name, layer)
+
+        self._freeze_stages()
+
+    def _freeze_stages(self):
+        if self.frozen_stages >= 0:
+            self.patch_embed.eval()
+            for param in self.patch_embed.parameters():
+                param.requires_grad = False
+
+        if self.frozen_stages >= 1 and self.ape:
+            self.absolute_pos_embed.requires_grad = False
+
+        if self.frozen_stages >= 2:
+            self.pos_drop.eval()
+            for i in range(0, self.frozen_stages - 1):
+                m = self.layers[i]
+                m.eval()
+                for param in m.parameters():
+                    param.requires_grad = False
+
+
+    def forward(self, x):
+        """Forward function."""
+        x = self.patch_embed(x)
+
+        Wh, Ww = x.size(2), x.size(3)
+        if self.ape:
+            # interpolate the position embedding to the corresponding size
+            absolute_pos_embed = F.interpolate(self.absolute_pos_embed, size=(Wh, Ww), mode='bicubic')
+            x = (x + absolute_pos_embed) # B Wh*Ww C
+
+        outs = []#x.contiguous()]
+        x = x.flatten(2).transpose(1, 2)
+        x = self.pos_drop(x)
+        for i in range(self.num_layers):
+            layer = self.layers[i]
+            x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww)
+
+            if i in self.out_indices:
+                norm_layer = getattr(self, f'norm{i}')
+                x_out = norm_layer(x_out)
+
+                out = x_out.view(-1, H, W, self.num_features[i]).permute(0, 3, 1, 2).contiguous()
+                outs.append(out)
+
+        return tuple(outs)
+
+    def train(self, mode=True):
+        """Convert the model into training mode while keep layers freezed."""
+        super(SwinTransformer, self).train(mode)
+        self._freeze_stages()
+
+def swin_v1_t():
+    model = SwinTransformer(embed_dim=96, depths=[2, 2, 6, 2], num_heads=[3, 6, 12, 24], window_size=7)
+    return model
+
+def swin_v1_s():
+    model = SwinTransformer(embed_dim=96, depths=[2, 2, 18, 2], num_heads=[3, 6, 12, 24], window_size=7)
+    return model
+
+def swin_v1_b():
+    model = SwinTransformer(embed_dim=128, depths=[2, 2, 18, 2], num_heads=[4, 8, 16, 32], window_size=12)
+    return model
+
+def swin_v1_l():
+    model = SwinTransformer(embed_dim=192, depths=[2, 2, 18, 2], num_heads=[6, 12, 24, 48], window_size=12)
+    return model
+
+
+
+### models/modules/deform_conv.py
+
+import torch
+import torch.nn as nn
+from torchvision.ops import deform_conv2d
+
+
+class DeformableConv2d(nn.Module):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size=3,
+                 stride=1,
+                 padding=1,
+                 bias=False):
+
+        super(DeformableConv2d, self).__init__()
+
+        assert type(kernel_size) == tuple or type(kernel_size) == int
+
+        kernel_size = kernel_size if type(kernel_size) == tuple else (kernel_size, kernel_size)
+        self.stride = stride if type(stride) == tuple else (stride, stride)
+        self.padding = padding
+
+        self.offset_conv = nn.Conv2d(in_channels,
+                                     2 * kernel_size[0] * kernel_size[1],
+                                     kernel_size=kernel_size,
+                                     stride=stride,
+                                     padding=self.padding,
+                                     bias=True)
+
+        nn.init.constant_(self.offset_conv.weight, 0.)
+        nn.init.constant_(self.offset_conv.bias, 0.)
+
+        self.modulator_conv = nn.Conv2d(in_channels,
+                                     1 * kernel_size[0] * kernel_size[1],
+                                     kernel_size=kernel_size,
+                                     stride=stride,
+                                     padding=self.padding,
+                                     bias=True)
+
+        nn.init.constant_(self.modulator_conv.weight, 0.)
+        nn.init.constant_(self.modulator_conv.bias, 0.)
+
+        self.regular_conv = nn.Conv2d(in_channels,
+                                      out_channels=out_channels,
+                                      kernel_size=kernel_size,
+                                      stride=stride,
+                                      padding=self.padding,
+                                      bias=bias)
+
+    def forward(self, x):
+        #h, w = x.shape[2:]
+        #max_offset = max(h, w)/4.
+
+        offset = self.offset_conv(x)#.clamp(-max_offset, max_offset)
+        modulator = 2. * torch.sigmoid(self.modulator_conv(x))
+
+        x = deform_conv2d(
+            input=x,
+            offset=offset,
+            weight=self.regular_conv.weight,
+            bias=self.regular_conv.bias,
+            padding=self.padding,
+            mask=modulator,
+            stride=self.stride,
+        )
+        return x
+
+
+
+
+### utils.py
+
+import torch.nn as nn
+
+
+def build_act_layer(act_layer):
+    if act_layer == 'ReLU':
+        return nn.ReLU(inplace=True)
+    elif act_layer == 'SiLU':
+        return nn.SiLU(inplace=True)
+    elif act_layer == 'GELU':
+        return nn.GELU()
+
+    raise NotImplementedError(f'build_act_layer does not support {act_layer}')
+
+
+def build_norm_layer(dim,
+                     norm_layer,
+                     in_format='channels_last',
+                     out_format='channels_last',
+                     eps=1e-6):
+    layers = []
+    if norm_layer == 'BN':
+        if in_format == 'channels_last':
+            layers.append(to_channels_first())
+        layers.append(nn.BatchNorm2d(dim))
+        if out_format == 'channels_last':
+            layers.append(to_channels_last())
+    elif norm_layer == 'LN':
+        if in_format == 'channels_first':
+            layers.append(to_channels_last())
+        layers.append(nn.LayerNorm(dim, eps=eps))
+        if out_format == 'channels_first':
+            layers.append(to_channels_first())
+    else:
+        raise NotImplementedError(
+            f'build_norm_layer does not support {norm_layer}')
+    return nn.Sequential(*layers)
+
+
+class to_channels_first(nn.Module):
+
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x):
+        return x.permute(0, 3, 1, 2)
+
+
+class to_channels_last(nn.Module):
+
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x):
+        return x.permute(0, 2, 3, 1)
+
+
+
+### dataset.py
+
+_class_labels_TR_sorted = (
+    'Airplane, Ant, Antenna, Archery, Axe, BabyCarriage, Bag, BalanceBeam, Balcony, Balloon, Basket, BasketballHoop, Beatle, Bed, Bee, Bench, Bicycle, '
+    'BicycleFrame, BicycleStand, Boat, Bonsai, BoomLift, Bridge, BunkBed, Butterfly, Button, Cable, CableLift, Cage, Camcorder, Cannon, Canoe, Car, '
+    'CarParkDropArm, Carriage, Cart, Caterpillar, CeilingLamp, Centipede, Chair, Clip, Clock, Clothes, CoatHanger, Comb, ConcretePumpTruck, Crack, Crane, '
+    'Cup, DentalChair, Desk, DeskChair, Diagram, DishRack, DoorHandle, Dragonfish, Dragonfly, Drum, Earphone, Easel, ElectricIron, Excavator, Eyeglasses, '
+    'Fan, Fence, Fencing, FerrisWheel, FireExtinguisher, Fishing, Flag, FloorLamp, Forklift, GasStation, Gate, Gear, Goal, Golf, GymEquipment, Hammock, '
+    'Handcart, Handcraft, Handrail, HangGlider, Harp, Harvester, Headset, Helicopter, Helmet, Hook, HorizontalBar, Hydrovalve, IroningTable, Jewelry, Key, '
+    'KidsPlayground, Kitchenware, Kite, Knife, Ladder, LaundryRack, Lightning, Lobster, Locust, Machine, MachineGun, MagazineRack, Mantis, Medal, MemorialArchway, '
+    'Microphone, Missile, MobileHolder, Monitor, Mosquito, Motorcycle, MovingTrolley, Mower, MusicPlayer, MusicStand, ObservationTower, Octopus, OilWell, '
+    'OlympicLogo, OperatingTable, OutdoorFitnessEquipment, Parachute, Pavilion, Piano, Pipe, PlowHarrow, PoleVault, Punchbag, Rack, Racket, Rifle, Ring, Robot, '
+    'RockClimbing, Rope, Sailboat, Satellite, Scaffold, Scale, Scissor, Scooter, Sculpture, Seadragon, Seahorse, Seal, SewingMachine, Ship, Shoe, ShoppingCart, '
+    'ShoppingTrolley, Shower, Shrimp, Signboard, Skateboarding, Skeleton, Skiing, Spade, SpeedBoat, Spider, Spoon, Stair, Stand, Stationary, SteeringWheel, '
+    'Stethoscope, Stool, Stove, StreetLamp, SweetStand, Swing, Sword, TV, Table, TableChair, TableLamp, TableTennis, Tank, Tapeline, Teapot, Telescope, Tent, '
+    'TobaccoPipe, Toy, Tractor, TrafficLight, TrafficSign, Trampoline, TransmissionTower, Tree, Tricycle, TrimmerCover, Tripod, Trombone, Truck, Trumpet, Tuba, '
+    'UAV, Umbrella, UnevenBars, UtilityPole, VacuumCleaner, Violin, Wakesurfing, Watch, WaterTower, WateringPot, Well, WellLid, Wheel, Wheelchair, WindTurbine, Windmill, WineGlass, WireWhisk, Yacht'
+)
+class_labels_TR_sorted = _class_labels_TR_sorted.split(', ')
+
+
+### models/backbones/build_backbones.py
+
+import torch
+import torch.nn as nn
+from collections import OrderedDict
+from torchvision.models import vgg16, vgg16_bn, VGG16_Weights, VGG16_BN_Weights, resnet50, ResNet50_Weights
+# from models.pvt_v2 import pvt_v2_b0, pvt_v2_b1, pvt_v2_b2, pvt_v2_b5
+# from models.swin_v1 import swin_v1_t, swin_v1_s, swin_v1_b, swin_v1_l
+# from config import Config
+
+
+config = Config()
+
+def build_backbone(bb_name, pretrained=True, params_settings=''):
+    if bb_name == 'vgg16':
+        bb_net = list(vgg16(pretrained=VGG16_Weights.DEFAULT if pretrained else None).children())[0]
+        bb = nn.Sequential(OrderedDict({'conv1': bb_net[:4], 'conv2': bb_net[4:9], 'conv3': bb_net[9:16], 'conv4': bb_net[16:23]}))
+    elif bb_name == 'vgg16bn':
+        bb_net = list(vgg16_bn(pretrained=VGG16_BN_Weights.DEFAULT if pretrained else None).children())[0]
+        bb = nn.Sequential(OrderedDict({'conv1': bb_net[:6], 'conv2': bb_net[6:13], 'conv3': bb_net[13:23], 'conv4': bb_net[23:33]}))
+    elif bb_name == 'resnet50':
+        bb_net = list(resnet50(pretrained=ResNet50_Weights.DEFAULT if pretrained else None).children())
+        bb = nn.Sequential(OrderedDict({'conv1': nn.Sequential(*bb_net[0:3]), 'conv2': bb_net[4], 'conv3': bb_net[5], 'conv4': bb_net[6]}))
+    else:
+        bb = eval('{}({})'.format(bb_name, params_settings))
+        if pretrained:
+            bb = load_weights(bb, bb_name)
+    return bb
+
+def load_weights(model, model_name):
+    save_model = torch.load(config.weights[model_name], map_location='cpu')
+    model_dict = model.state_dict()
+    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()}
+    # to ignore the weights with mismatched size when I modify the backbone itself.
+    if not state_dict:
+        save_model_keys = list(save_model.keys())
+        sub_item = save_model_keys[0] if len(save_model_keys) == 1 else None
+        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()}
+        if not state_dict or not sub_item:
+            print('Weights are not successully loaded. Check the state dict of weights file.')
+            return None
+        else:
+            print('Found correct weights in the "{}" item of loaded state_dict.'.format(sub_item))
+    model_dict.update(state_dict)
+    model.load_state_dict(model_dict)
+    return model
+
+
+
+### models/modules/decoder_blocks.py
+
+import torch
+import torch.nn as nn
+# from models.aspp import ASPP, ASPPDeformable
+# from config import Config
+
+
+# config = Config()
+
+
+class BasicDecBlk(nn.Module):
+    def __init__(self, in_channels=64, out_channels=64, inter_channels=64):
+        super(BasicDecBlk, self).__init__()
+        inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
+        self.conv_in = nn.Conv2d(in_channels, inter_channels, 3, 1, padding=1)
+        self.relu_in = nn.ReLU(inplace=True)
+        if config.dec_att == 'ASPP':
+            self.dec_att = ASPP(in_channels=inter_channels)
+        elif config.dec_att == 'ASPPDeformable':
+            self.dec_att = ASPPDeformable(in_channels=inter_channels)
+        self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, padding=1)
+        self.bn_in = nn.BatchNorm2d(inter_channels) if config.batch_size > 1 else nn.Identity()
+        self.bn_out = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
+
+    def forward(self, x):
+        x = self.conv_in(x)
+        x = self.bn_in(x)
+        x = self.relu_in(x)
+        if hasattr(self, 'dec_att'):
+            x = self.dec_att(x)
+        x = self.conv_out(x)
+        x = self.bn_out(x)
+        return x
+
+
+class ResBlk(nn.Module):
+    def __init__(self, in_channels=64, out_channels=None, inter_channels=64):
+        super(ResBlk, self).__init__()
+        if out_channels is None:
+            out_channels = in_channels
+        inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
+
+        self.conv_in = nn.Conv2d(in_channels, inter_channels, 3, 1, padding=1)
+        self.bn_in = nn.BatchNorm2d(inter_channels) if config.batch_size > 1 else nn.Identity()
+        self.relu_in = nn.ReLU(inplace=True)
+
+        if config.dec_att == 'ASPP':
+            self.dec_att = ASPP(in_channels=inter_channels)
+        elif config.dec_att == 'ASPPDeformable':
+            self.dec_att = ASPPDeformable(in_channels=inter_channels)
+
+        self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, padding=1)
+        self.bn_out = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
+
+        self.conv_resi = nn.Conv2d(in_channels, out_channels, 1, 1, 0)
+
+    def forward(self, x):
+        _x = self.conv_resi(x)
+        x = self.conv_in(x)
+        x = self.bn_in(x)
+        x = self.relu_in(x)
+        if hasattr(self, 'dec_att'):
+            x = self.dec_att(x)
+        x = self.conv_out(x)
+        x = self.bn_out(x)
+        return x + _x
+
+
+
+### models/modules/lateral_blocks.py
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from functools import partial
+
+# from config import Config
+
+
+# config = Config()
+
+
+class BasicLatBlk(nn.Module):
+    def __init__(self, in_channels=64, out_channels=64, inter_channels=64):
+        super(BasicLatBlk, self).__init__()
+        inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
+        self.conv = nn.Conv2d(in_channels, out_channels, 1, 1, 0)
+
+    def forward(self, x):
+        x = self.conv(x)
+        return x
+
+
+
+### models/modules/aspp.py
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+# from models.deform_conv import DeformableConv2d
+# from config import Config
+
+
+# config = Config()
+
+
+class _ASPPModule(nn.Module):
+    def __init__(self, in_channels, planes, kernel_size, padding, dilation):
+        super(_ASPPModule, self).__init__()
+        self.atrous_conv = nn.Conv2d(in_channels, planes, kernel_size=kernel_size,
+                                            stride=1, padding=padding, dilation=dilation, bias=False)
+        self.bn = nn.BatchNorm2d(planes) if config.batch_size > 1 else nn.Identity()
+        self.relu = nn.ReLU(inplace=True)
+
+    def forward(self, x):
+        x = self.atrous_conv(x)
+        x = self.bn(x)
+
+        return self.relu(x)
+
+
+class ASPP(nn.Module):
+    def __init__(self, in_channels=64, out_channels=None, output_stride=16):
+        super(ASPP, self).__init__()
+        self.down_scale = 1
+        if out_channels is None:
+            out_channels = in_channels
+        self.in_channelster = 256 // self.down_scale
+        if output_stride == 16:
+            dilations = [1, 6, 12, 18]
+        elif output_stride == 8:
+            dilations = [1, 12, 24, 36]
+        else:
+            raise NotImplementedError
+
+        self.aspp1 = _ASPPModule(in_channels, self.in_channelster, 1, padding=0, dilation=dilations[0])
+        self.aspp2 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[1], dilation=dilations[1])
+        self.aspp3 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[2], dilation=dilations[2])
+        self.aspp4 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[3], dilation=dilations[3])
+
+        self.global_avg_pool = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)),
+                                             nn.Conv2d(in_channels, self.in_channelster, 1, stride=1, bias=False),
+                                             nn.BatchNorm2d(self.in_channelster) if config.batch_size > 1 else nn.Identity(),
+                                             nn.ReLU(inplace=True))
+        self.conv1 = nn.Conv2d(self.in_channelster * 5, out_channels, 1, bias=False)
+        self.bn1 = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
+        self.relu = nn.ReLU(inplace=True)
+        self.dropout = nn.Dropout(0.5)
+
+    def forward(self, x):
+        x1 = self.aspp1(x)
+        x2 = self.aspp2(x)
+        x3 = self.aspp3(x)
+        x4 = self.aspp4(x)
+        x5 = self.global_avg_pool(x)
+        x5 = F.interpolate(x5, size=x1.size()[2:], mode='bilinear', align_corners=True)
+        x = torch.cat((x1, x2, x3, x4, x5), dim=1)
+
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+
+        return self.dropout(x)
+
+
+##################### Deformable
+class _ASPPModuleDeformable(nn.Module):
+    def __init__(self, in_channels, planes, kernel_size, padding):
+        super(_ASPPModuleDeformable, self).__init__()
+        self.atrous_conv = DeformableConv2d(in_channels, planes, kernel_size=kernel_size,
+                                            stride=1, padding=padding, bias=False)
+        self.bn = nn.BatchNorm2d(planes) if config.batch_size > 1 else nn.Identity()
+        self.relu = nn.ReLU(inplace=True)
+
+    def forward(self, x):
+        x = self.atrous_conv(x)
+        x = self.bn(x)
+
+        return self.relu(x)
+
+
+class ASPPDeformable(nn.Module):
+    def __init__(self, in_channels, out_channels=None, parallel_block_sizes=[1, 3, 7]):
+        super(ASPPDeformable, self).__init__()
+        self.down_scale = 1
+        if out_channels is None:
+            out_channels = in_channels
+        self.in_channelster = 256 // self.down_scale
+
+        self.aspp1 = _ASPPModuleDeformable(in_channels, self.in_channelster, 1, padding=0)
+        self.aspp_deforms = nn.ModuleList([
+            _ASPPModuleDeformable(in_channels, self.in_channelster, conv_size, padding=int(conv_size//2)) for conv_size in parallel_block_sizes
+        ])
+
+        self.global_avg_pool = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)),
+                                             nn.Conv2d(in_channels, self.in_channelster, 1, stride=1, bias=False),
+                                             nn.BatchNorm2d(self.in_channelster) if config.batch_size > 1 else nn.Identity(),
+                                             nn.ReLU(inplace=True))
+        self.conv1 = nn.Conv2d(self.in_channelster * (2 + len(self.aspp_deforms)), out_channels, 1, bias=False)
+        self.bn1 = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
+        self.relu = nn.ReLU(inplace=True)
+        self.dropout = nn.Dropout(0.5)
+
+    def forward(self, x):
+        x1 = self.aspp1(x)
+        x_aspp_deforms = [aspp_deform(x) for aspp_deform in self.aspp_deforms]
+        x5 = self.global_avg_pool(x)
+        x5 = F.interpolate(x5, size=x1.size()[2:], mode='bilinear', align_corners=True)
+        x = torch.cat((x1, *x_aspp_deforms, x5), dim=1)
+
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+
+        return self.dropout(x)
+
+
+
+### models/refinement/refiner.py
+
+import torch
+import torch.nn as nn
+from collections import OrderedDict
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torchvision.models import vgg16, vgg16_bn
+from torchvision.models import resnet50
+
+# from config import Config
+# from dataset import class_labels_TR_sorted
+# from models.build_backbone import build_backbone
+# from models.decoder_blocks import BasicDecBlk
+# from models.lateral_blocks import BasicLatBlk
+# from models.ing import *
+# from models.stem_layer import StemLayer
+
+
+class RefinerPVTInChannels4(nn.Module):
+    def __init__(self, in_channels=3+1):
+        super(RefinerPVTInChannels4, self).__init__()
+        self.config = Config()
+        self.epoch = 1
+        self.bb = build_backbone(self.config.bb, params_settings='in_channels=4')
+
+        lateral_channels_in_collection = {
+            'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
+            'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
+            'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
+        }
+        channels = lateral_channels_in_collection[self.config.bb]
+        self.squeeze_module = BasicDecBlk(channels[0], channels[0])
+
+        self.decoder = Decoder(channels)
+
+        if 0:
+            for key, value in self.named_parameters():
+                if 'bb.' in key:
+                    value.requires_grad = False
+
+    def forward(self, x):
+        if isinstance(x, list):
+            x = torch.cat(x, dim=1)
+        ########## Encoder ##########
+        if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
+            x1 = self.bb.conv1(x)
+            x2 = self.bb.conv2(x1)
+            x3 = self.bb.conv3(x2)
+            x4 = self.bb.conv4(x3)
+        else:
+            x1, x2, x3, x4 = self.bb(x)
+
+        x4 = self.squeeze_module(x4)
+
+        ########## Decoder ##########
+
+        features = [x, x1, x2, x3, x4]
+        scaled_preds = self.decoder(features)
+
+        return scaled_preds
+
+
+class Refiner(nn.Module):
+    def __init__(self, in_channels=3+1):
+        super(Refiner, self).__init__()
+        self.config = Config()
+        self.epoch = 1
+        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')
+        self.bb = build_backbone(self.config.bb)
+
+        lateral_channels_in_collection = {
+            'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
+            'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
+            'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
+        }
+        channels = lateral_channels_in_collection[self.config.bb]
+        self.squeeze_module = BasicDecBlk(channels[0], channels[0])
+
+        self.decoder = Decoder(channels)
+
+        if 0:
+            for key, value in self.named_parameters():
+                if 'bb.' in key:
+                    value.requires_grad = False
+
+    def forward(self, x):
+        if isinstance(x, list):
+            x = torch.cat(x, dim=1)
+        x = self.stem_layer(x)
+        ########## Encoder ##########
+        if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
+            x1 = self.bb.conv1(x)
+            x2 = self.bb.conv2(x1)
+            x3 = self.bb.conv3(x2)
+            x4 = self.bb.conv4(x3)
+        else:
+            x1, x2, x3, x4 = self.bb(x)
+
+        x4 = self.squeeze_module(x4)
+
+        ########## Decoder ##########
+
+        features = [x, x1, x2, x3, x4]
+        scaled_preds = self.decoder(features)
+
+        return scaled_preds
+
+
+class Decoder(nn.Module):
+    def __init__(self, channels):
+        super(Decoder, self).__init__()
+        self.config = Config()
+        DecoderBlock = eval('BasicDecBlk')
+        LateralBlock = eval('BasicLatBlk')
+
+        self.decoder_block4 = DecoderBlock(channels[0], channels[1])
+        self.decoder_block3 = DecoderBlock(channels[1], channels[2])
+        self.decoder_block2 = DecoderBlock(channels[2], channels[3])
+        self.decoder_block1 = DecoderBlock(channels[3], channels[3]//2)
+
+        self.lateral_block4 = LateralBlock(channels[1], channels[1])
+        self.lateral_block3 = LateralBlock(channels[2], channels[2])
+        self.lateral_block2 = LateralBlock(channels[3], channels[3])
+
+        if self.config.ms_supervision:
+            self.conv_ms_spvn_4 = nn.Conv2d(channels[1], 1, 1, 1, 0)
+            self.conv_ms_spvn_3 = nn.Conv2d(channels[2], 1, 1, 1, 0)
+            self.conv_ms_spvn_2 = nn.Conv2d(channels[3], 1, 1, 1, 0)
+        self.conv_out1 = nn.Sequential(nn.Conv2d(channels[3]//2, 1, 1, 1, 0))
+
+    def forward(self, features):
+        x, x1, x2, x3, x4 = features
+        outs = []
+        p4 = self.decoder_block4(x4)
+        _p4 = F.interpolate(p4, size=x3.shape[2:], mode='bilinear', align_corners=True)
+        _p3 = _p4 + self.lateral_block4(x3)
+
+        p3 = self.decoder_block3(_p3)
+        _p3 = F.interpolate(p3, size=x2.shape[2:], mode='bilinear', align_corners=True)
+        _p2 = _p3 + self.lateral_block3(x2)
+
+        p2 = self.decoder_block2(_p2)
+        _p2 = F.interpolate(p2, size=x1.shape[2:], mode='bilinear', align_corners=True)
+        _p1 = _p2 + self.lateral_block2(x1)
+
+        _p1 = self.decoder_block1(_p1)
+        _p1 = F.interpolate(_p1, size=x.shape[2:], mode='bilinear', align_corners=True)
+        p1_out = self.conv_out1(_p1)
+
+        if self.config.ms_supervision:
+            outs.append(self.conv_ms_spvn_4(p4))
+            outs.append(self.conv_ms_spvn_3(p3))
+            outs.append(self.conv_ms_spvn_2(p2))
+        outs.append(p1_out)
+        return outs
+
+
+class RefUNet(nn.Module):
+    # Refinement
+    def __init__(self, in_channels=3+1):
+        super(RefUNet, self).__init__()
+        self.encoder_1 = nn.Sequential(
+            nn.Conv2d(in_channels, 64, 3, 1, 1),
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.encoder_2 = nn.Sequential(
+            nn.MaxPool2d(2, 2, ceil_mode=True),
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.encoder_3 = nn.Sequential(
+            nn.MaxPool2d(2, 2, ceil_mode=True),
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.encoder_4 = nn.Sequential(
+            nn.MaxPool2d(2, 2, ceil_mode=True),
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.pool4 = nn.MaxPool2d(2, 2, ceil_mode=True)
+        #####
+        self.decoder_5 = nn.Sequential(
+            nn.Conv2d(64, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+        #####
+        self.decoder_4 = nn.Sequential(
+            nn.Conv2d(128, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.decoder_3 = nn.Sequential(
+            nn.Conv2d(128, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.decoder_2 = nn.Sequential(
+            nn.Conv2d(128, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.decoder_1 = nn.Sequential(
+            nn.Conv2d(128, 64, 3, 1, 1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+
+        self.conv_d0 = nn.Conv2d(64, 1, 3, 1, 1)
+
+        self.upscore2 = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)
+
+    def forward(self, x):
+        outs = []
+        if isinstance(x, list):
+            x = torch.cat(x, dim=1)
+        hx = x
+
+        hx1 = self.encoder_1(hx)
+        hx2 = self.encoder_2(hx1)
+        hx3 = self.encoder_3(hx2)
+        hx4 = self.encoder_4(hx3)
+
+        hx = self.decoder_5(self.pool4(hx4))
+        hx = torch.cat((self.upscore2(hx), hx4), 1)
+
+        d4 = self.decoder_4(hx)
+        hx = torch.cat((self.upscore2(d4), hx3), 1)
+
+        d3 = self.decoder_3(hx)
+        hx = torch.cat((self.upscore2(d3), hx2), 1)
+
+        d2 = self.decoder_2(hx)
+        hx = torch.cat((self.upscore2(d2), hx1), 1)
+
+        d1 = self.decoder_1(hx)
+
+        x = self.conv_d0(d1)
+        outs.append(x)
+        return outs
+
+
+
+### models/stem_layer.py
+
+import torch.nn as nn
+# from utils import build_act_layer, build_norm_layer
+
+
+class StemLayer(nn.Module):
+    r""" Stem layer of InternImage
+    Args:
+        in_channels (int): number of input channels
+        out_channels (int): number of output channels
+        act_layer (str): activation layer
+        norm_layer (str): normalization layer
+    """
+
+    def __init__(self,
+                 in_channels=3+1,
+                 inter_channels=48,
+                 out_channels=96,
+                 act_layer='GELU',
+                 norm_layer='BN'):
+        super().__init__()
+        self.conv1 = nn.Conv2d(in_channels,
+                               inter_channels,
+                               kernel_size=3,
+                               stride=1,
+                               padding=1)
+        self.norm1 = build_norm_layer(
+            inter_channels, norm_layer, 'channels_first', 'channels_first'
+        )
+        self.act = build_act_layer(act_layer)
+        self.conv2 = nn.Conv2d(inter_channels,
+                               out_channels,
+                               kernel_size=3,
+                               stride=1,
+                               padding=1)
+        self.norm2 = build_norm_layer(
+            out_channels, norm_layer, 'channels_first', 'channels_first'
+        )
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.norm1(x)
+        x = self.act(x)
+        x = self.conv2(x)
+        x = self.norm2(x)
+        return x
+
+
+### models/birefnet.py
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from kornia.filters import laplacian
+from transformers import PreTrainedModel
+from einops import rearrange
+
+# from config import Config
+# from dataset import class_labels_TR_sorted
+# from models.build_backbone import build_backbone
+# from models.decoder_blocks import BasicDecBlk, ResBlk, HierarAttDecBlk
+# from models.lateral_blocks import BasicLatBlk
+# from models.aspp import ASPP, ASPPDeformable
+# from models.ing import *
+# from models.refiner import Refiner, RefinerPVTInChannels4, RefUNet
+# from models.stem_layer import StemLayer
+from .BiRefNet_config import BiRefNetConfig
+
+
+def image2patches(image, grid_h=2, grid_w=2, patch_ref=None, transformation='b c (hg h) (wg w) -> (b hg wg) c h w'):
+    if patch_ref is not None:
+        grid_h, grid_w = image.shape[-2] // patch_ref.shape[-2], image.shape[-1] // patch_ref.shape[-1]
+    patches = rearrange(image, transformation, hg=grid_h, wg=grid_w)
+    return patches
+
+def patches2image(patches, grid_h=2, grid_w=2, patch_ref=None, transformation='(b hg wg) c h w -> b c (hg h) (wg w)'):
+    if patch_ref is not None:
+        grid_h, grid_w = patch_ref.shape[-2] // patches[0].shape[-2], patch_ref.shape[-1] // patches[0].shape[-1]
+    image = rearrange(patches, transformation, hg=grid_h, wg=grid_w)
+    return image
+
+class BiRefNet(
+    PreTrainedModel
+):
+    config_class = BiRefNetConfig
+    def __init__(self, bb_pretrained=True, config=BiRefNetConfig()):
+        super(BiRefNet, self).__init__(config)
+        bb_pretrained = config.bb_pretrained
+        self.config = Config()
+        self.epoch = 1
+        self.bb = build_backbone(self.config.bb, pretrained=bb_pretrained)
+
+        channels = self.config.lateral_channels_in_collection
+
+        if self.config.auxiliary_classification:
+            self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
+            self.cls_head = nn.Sequential(
+                nn.Linear(channels[0], len(class_labels_TR_sorted))
+            )
+
+        if self.config.squeeze_block:
+            self.squeeze_module = nn.Sequential(*[
+                eval(self.config.squeeze_block.split('_x')[0])(channels[0]+sum(self.config.cxt), channels[0])
+                for _ in range(eval(self.config.squeeze_block.split('_x')[1]))
+            ])
+
+        self.decoder = Decoder(channels)
+
+        if self.config.ender:
+            self.dec_end = nn.Sequential(
+                nn.Conv2d(1, 16, 3, 1, 1),
+                nn.Conv2d(16, 1, 3, 1, 1),
+                nn.ReLU(inplace=True),
+            )
+
+        # refine patch-level segmentation
+        if self.config.refine:
+            if self.config.refine == 'itself':
+                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')
+            else:
+                self.refiner = eval('{}({})'.format(self.config.refine, 'in_channels=3+1'))
+
+        if self.config.freeze_bb:
+            # Freeze the backbone...
+            print(self.named_parameters())
+            for key, value in self.named_parameters():
+                if 'bb.' in key and 'refiner.' not in key:
+                    value.requires_grad = False
+
+        self.post_init()
+
+    def forward_enc(self, x):
+        if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
+            x1 = self.bb.conv1(x); x2 = self.bb.conv2(x1); x3 = self.bb.conv3(x2); x4 = self.bb.conv4(x3)
+        else:
+            x1, x2, x3, x4 = self.bb(x)
+            if self.config.mul_scl_ipt == 'cat':
+                B, C, H, W = x.shape
+                x1_, x2_, x3_, x4_ = self.bb(F.interpolate(x, size=(H//2, W//2), mode='bilinear', align_corners=True))
+                x1 = torch.cat([x1, F.interpolate(x1_, size=x1.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+                x2 = torch.cat([x2, F.interpolate(x2_, size=x2.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+                x3 = torch.cat([x3, F.interpolate(x3_, size=x3.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+                x4 = torch.cat([x4, F.interpolate(x4_, size=x4.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+            elif self.config.mul_scl_ipt == 'add':
+                B, C, H, W = x.shape
+                x1_, x2_, x3_, x4_ = self.bb(F.interpolate(x, size=(H//2, W//2), mode='bilinear', align_corners=True))
+                x1 = x1 + F.interpolate(x1_, size=x1.shape[2:], mode='bilinear', align_corners=True)
+                x2 = x2 + F.interpolate(x2_, size=x2.shape[2:], mode='bilinear', align_corners=True)
+                x3 = x3 + F.interpolate(x3_, size=x3.shape[2:], mode='bilinear', align_corners=True)
+                x4 = x4 + F.interpolate(x4_, size=x4.shape[2:], mode='bilinear', align_corners=True)
+        class_preds = self.cls_head(self.avgpool(x4).view(x4.shape[0], -1)) if self.training and self.config.auxiliary_classification else None
+        if self.config.cxt:
+            x4 = torch.cat(
+                (
+                    *[
+                        F.interpolate(x1, size=x4.shape[2:], mode='bilinear', align_corners=True),
+                        F.interpolate(x2, size=x4.shape[2:], mode='bilinear', align_corners=True),
+                        F.interpolate(x3, size=x4.shape[2:], mode='bilinear', align_corners=True),
+                    ][-len(self.config.cxt):],
+                    x4
+                ),
+                dim=1
+            )
+        return (x1, x2, x3, x4), class_preds
+
+    def forward_ori(self, x):
+        ########## Encoder ##########
+        (x1, x2, x3, x4), class_preds = self.forward_enc(x)
+        if self.config.squeeze_block:
+            x4 = self.squeeze_module(x4)
+        ########## Decoder ##########
+        features = [x, x1, x2, x3, x4]
+        if self.training and self.config.out_ref:
+            features.append(laplacian(torch.mean(x, dim=1).unsqueeze(1), kernel_size=5))
+        scaled_preds = self.decoder(features)
+        return scaled_preds, class_preds
+
+    def forward(self, x):
+        scaled_preds, class_preds = self.forward_ori(x)
+        class_preds_lst = [class_preds]
+        return [scaled_preds, class_preds_lst] if self.training else scaled_preds
+
+
+class Decoder(nn.Module):
+    def __init__(self, channels):
+        super(Decoder, self).__init__()
+        self.config = Config()
+        DecoderBlock = eval(self.config.dec_blk)
+        LateralBlock = eval(self.config.lat_blk)
+
+        if self.config.dec_ipt:
+            self.split = self.config.dec_ipt_split
+            N_dec_ipt = 64
+            DBlock = SimpleConvs
+            ic = 64
+            ipt_cha_opt = 1
+            self.ipt_blk5 = DBlock(2**10*3 if self.split else 3, [N_dec_ipt, channels[0]//8][ipt_cha_opt], inter_channels=ic)
+            self.ipt_blk4 = DBlock(2**8*3 if self.split else 3, [N_dec_ipt, channels[0]//8][ipt_cha_opt], inter_channels=ic)
+            self.ipt_blk3 = DBlock(2**6*3 if self.split else 3, [N_dec_ipt, channels[1]//8][ipt_cha_opt], inter_channels=ic)
+            self.ipt_blk2 = DBlock(2**4*3 if self.split else 3, [N_dec_ipt, channels[2]//8][ipt_cha_opt], inter_channels=ic)
+            self.ipt_blk1 = DBlock(2**0*3 if self.split else 3, [N_dec_ipt, channels[3]//8][ipt_cha_opt], inter_channels=ic)
+        else:
+            self.split = None
+
+        self.decoder_block4 = DecoderBlock(channels[0]+([N_dec_ipt, channels[0]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[1])
+        self.decoder_block3 = DecoderBlock(channels[1]+([N_dec_ipt, channels[0]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[2])
+        self.decoder_block2 = DecoderBlock(channels[2]+([N_dec_ipt, channels[1]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[3])
+        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)
+        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))
+
+        self.lateral_block4 = LateralBlock(channels[1], channels[1])
+        self.lateral_block3 = LateralBlock(channels[2], channels[2])
+        self.lateral_block2 = LateralBlock(channels[3], channels[3])
+
+        if self.config.ms_supervision:
+            self.conv_ms_spvn_4 = nn.Conv2d(channels[1], 1, 1, 1, 0)
+            self.conv_ms_spvn_3 = nn.Conv2d(channels[2], 1, 1, 1, 0)
+            self.conv_ms_spvn_2 = nn.Conv2d(channels[3], 1, 1, 1, 0)
+
+            if self.config.out_ref:
+                _N = 16
+                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))
+                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))
+                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))
+
+                self.gdt_convs_pred_4 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                self.gdt_convs_pred_3 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                self.gdt_convs_pred_2 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+
+                self.gdt_convs_attn_4 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                self.gdt_convs_attn_3 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+                self.gdt_convs_attn_2 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
+
+    def forward(self, features):
+        if self.training and self.config.out_ref:
+            outs_gdt_pred = []
+            outs_gdt_label = []
+            x, x1, x2, x3, x4, gdt_gt = features
+        else:
+            x, x1, x2, x3, x4 = features
+        outs = []
+
+        if self.config.dec_ipt:
+            patches_batch = image2patches(x, patch_ref=x4, transformation='b c (hg h) (wg w) -> b (c hg wg) h w') if self.split else x
+            x4 = torch.cat((x4, self.ipt_blk5(F.interpolate(patches_batch, size=x4.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p4 = self.decoder_block4(x4)
+        m4 = self.conv_ms_spvn_4(p4) if self.config.ms_supervision and self.training else None
+        if self.config.out_ref:
+            p4_gdt = self.gdt_convs_4(p4)
+            if self.training:
+                # >> GT:
+                m4_dia = m4
+                gdt_label_main_4 = gdt_gt * F.interpolate(m4_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
+                outs_gdt_label.append(gdt_label_main_4)
+                # >> Pred:
+                gdt_pred_4 = self.gdt_convs_pred_4(p4_gdt)
+                outs_gdt_pred.append(gdt_pred_4)
+            gdt_attn_4 = self.gdt_convs_attn_4(p4_gdt).sigmoid()
+            # >> Finally:
+            p4 = p4 * gdt_attn_4
+        _p4 = F.interpolate(p4, size=x3.shape[2:], mode='bilinear', align_corners=True)
+        _p3 = _p4 + self.lateral_block4(x3)
+
+        if self.config.dec_ipt:
+            patches_batch = image2patches(x, patch_ref=_p3, transformation='b c (hg h) (wg w) -> b (c hg wg) h w') if self.split else x
+            _p3 = torch.cat((_p3, self.ipt_blk4(F.interpolate(patches_batch, size=x3.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p3 = self.decoder_block3(_p3)
+        m3 = self.conv_ms_spvn_3(p3) if self.config.ms_supervision and self.training else None
+        if self.config.out_ref:
+            p3_gdt = self.gdt_convs_3(p3)
+            if self.training:
+                # >> GT:
+                # m3 --dilation--> m3_dia
+                # G_3^gt * m3_dia --> G_3^m, which is the label of gradient
+                m3_dia = m3
+                gdt_label_main_3 = gdt_gt * F.interpolate(m3_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
+                outs_gdt_label.append(gdt_label_main_3)
+                # >> Pred:
+                # p3 --conv--BN--> F_3^G, where F_3^G predicts the \hat{G_3} with xx
+                # F_3^G --sigmoid--> A_3^G
+                gdt_pred_3 = self.gdt_convs_pred_3(p3_gdt)
+                outs_gdt_pred.append(gdt_pred_3)
+            gdt_attn_3 = self.gdt_convs_attn_3(p3_gdt).sigmoid()
+            # >> Finally:
+            # p3 = p3 * A_3^G
+            p3 = p3 * gdt_attn_3
+        _p3 = F.interpolate(p3, size=x2.shape[2:], mode='bilinear', align_corners=True)
+        _p2 = _p3 + self.lateral_block3(x2)
+
+        if self.config.dec_ipt:
+            patches_batch = image2patches(x, patch_ref=_p2, transformation='b c (hg h) (wg w) -> b (c hg wg) h w') if self.split else x
+            _p2 = torch.cat((_p2, self.ipt_blk3(F.interpolate(patches_batch, size=x2.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p2 = self.decoder_block2(_p2)
+        m2 = self.conv_ms_spvn_2(p2) if self.config.ms_supervision and self.training else None
+        if self.config.out_ref:
+            p2_gdt = self.gdt_convs_2(p2)
+            if self.training:
+                # >> GT:
+                m2_dia = m2
+                gdt_label_main_2 = gdt_gt * F.interpolate(m2_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
+                outs_gdt_label.append(gdt_label_main_2)
+                # >> Pred:
+                gdt_pred_2 = self.gdt_convs_pred_2(p2_gdt)
+                outs_gdt_pred.append(gdt_pred_2)
+            gdt_attn_2 = self.gdt_convs_attn_2(p2_gdt).sigmoid()
+            # >> Finally:
+            p2 = p2 * gdt_attn_2
+        _p2 = F.interpolate(p2, size=x1.shape[2:], mode='bilinear', align_corners=True)
+        _p1 = _p2 + self.lateral_block2(x1)
+
+        if self.config.dec_ipt:
+            patches_batch = image2patches(x, patch_ref=_p1, transformation='b c (hg h) (wg w) -> b (c hg wg) h w') if self.split else x
+            _p1 = torch.cat((_p1, self.ipt_blk2(F.interpolate(patches_batch, size=x1.shape[2:], mode='bilinear', align_corners=True))), 1)
+        _p1 = self.decoder_block1(_p1)
+        _p1 = F.interpolate(_p1, size=x.shape[2:], mode='bilinear', align_corners=True)
+
+        if self.config.dec_ipt:
+            patches_batch = image2patches(x, patch_ref=_p1, transformation='b c (hg h) (wg w) -> b (c hg wg) h w') if self.split else x
+            _p1 = torch.cat((_p1, self.ipt_blk1(F.interpolate(patches_batch, size=x.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p1_out = self.conv_out1(_p1)
+
+        if self.config.ms_supervision and self.training:
+            outs.append(m4)
+            outs.append(m3)
+            outs.append(m2)
+        outs.append(p1_out)
+        return outs if not (self.config.out_ref and self.training) else ([outs_gdt_pred, outs_gdt_label], outs)
+
+
+class SimpleConvs(nn.Module):
+    def __init__(
+        self, in_channels: int, out_channels: int, inter_channels=64
+    ) -> None:
+        super().__init__()
+        self.conv1 = nn.Conv2d(in_channels, inter_channels, 3, 1, 1)
+        self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, 1)
+
+    def forward(self, x):
+        return self.conv_out(self.conv1(x))

extras/birefnet/config.json

@@ -0,0 +1,20 @@
+{
+  "_name_or_path": "ZhengPeng7/BiRefNet",
+  "architectures": [
+    "BiRefNet"
+  ],
+  "auto_map": {
+    "AutoConfig": "BiRefNet_config.BiRefNetConfig",
+    "AutoModelForImageSegmentation": "birefnet.BiRefNet"
+  },
+  "custom_pipelines": {
+    "image-segmentation": {
+      "pt": [
+        "AutoModelForImageSegmentation"
+      ],
+      "tf": [],
+      "type": "image"
+    }
+  },
+  "bb_pretrained": false
+}

extras/birefnet/handler.py

@@ -0,0 +1,139 @@
+# These HF deployment codes refer to https://huggingface.co/not-lain/BiRefNet/raw/main/handler.py.
+from typing import Dict, List, Any, Tuple
+import os
+import requests
+from io import BytesIO
+import cv2
+import numpy as np
+from PIL import Image
+import torch
+from torchvision import transforms
+from transformers import AutoModelForImageSegmentation
+
+torch.set_float32_matmul_precision(["high", "highest"][0])
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+### image_proc.py
+def refine_foreground(image, mask, r=90):
+    if mask.size != image.size:
+        mask = mask.resize(image.size)
+    image = np.array(image) / 255.0
+    mask = np.array(mask) / 255.0
+    estimated_foreground = FB_blur_fusion_foreground_estimator_2(image, mask, r=r)
+    image_masked = Image.fromarray((estimated_foreground * 255.0).astype(np.uint8))
+    return image_masked
+
+
+def FB_blur_fusion_foreground_estimator_2(image, alpha, r=90):
+    # Thanks to the source: https://github.com/Photoroom/fast-foreground-estimation
+    alpha = alpha[:, :, None]
+    F, blur_B = FB_blur_fusion_foreground_estimator(image, image, image, alpha, r)
+    return FB_blur_fusion_foreground_estimator(image, F, blur_B, alpha, r=6)[0]
+
+
+def FB_blur_fusion_foreground_estimator(image, F, B, alpha, r=90):
+    if isinstance(image, Image.Image):
+        image = np.array(image) / 255.0
+    blurred_alpha = cv2.blur(alpha, (r, r))[:, :, None]
+
+    blurred_FA = cv2.blur(F * alpha, (r, r))
+    blurred_F = blurred_FA / (blurred_alpha + 1e-5)
+
+    blurred_B1A = cv2.blur(B * (1 - alpha), (r, r))
+    blurred_B = blurred_B1A / ((1 - blurred_alpha) + 1e-5)
+    F = blurred_F + alpha * \
+        (image - alpha * blurred_F - (1 - alpha) * blurred_B)
+    F = np.clip(F, 0, 1)
+    return F, blurred_B
+
+
+class ImagePreprocessor():
+    def __init__(self, resolution: Tuple[int, int] = (1024, 1024)) -> None:
+        self.transform_image = transforms.Compose([
+            transforms.Resize(resolution),
+            transforms.ToTensor(),
+            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
+        ])
+
+    def proc(self, image: Image.Image) -> torch.Tensor:
+        image = self.transform_image(image)
+        return image
+
+usage_to_weights_file = {
+    'General': 'BiRefNet',
+    'General-HR': 'BiRefNet_HR',
+    'General-Lite': 'BiRefNet_lite',
+    'General-Lite-2K': 'BiRefNet_lite-2K',
+    'General-reso_512': 'BiRefNet-reso_512',
+    'Matting': 'BiRefNet-matting',
+    'Matting-HR': 'BiRefNet_HR-Matting',
+    'Portrait': 'BiRefNet-portrait',
+    'DIS': 'BiRefNet-DIS5K',
+    'HRSOD': 'BiRefNet-HRSOD',
+    'COD': 'BiRefNet-COD',
+    'DIS-TR_TEs': 'BiRefNet-DIS5K-TR_TEs',
+    'General-legacy': 'BiRefNet-legacy'
+}
+
+# Choose the version of BiRefNet here.
+usage = 'General'
+
+# Set resolution
+if usage in ['General-Lite-2K']:
+    resolution = (2560, 1440)
+elif usage in ['General-reso_512']:
+    resolution = (512, 512)
+elif usage in ['General-HR', 'Matting-HR']:
+    resolution = (2048, 2048)
+else:
+    resolution = (1024, 1024) 
+
+half_precision = True
+
+class EndpointHandler():
+    def __init__(self, path=''):
+        self.birefnet = AutoModelForImageSegmentation.from_pretrained(
+            '/'.join(('zhengpeng7', usage_to_weights_file[usage])), trust_remote_code=True
+        )
+        self.birefnet.to(device)
+        self.birefnet.eval()
+        if half_precision:
+            self.birefnet.half()
+
+    def __call__(self, data: Dict[str, Any]):
+        """
+        data args:
+            inputs (:obj: `str`)
+            date (:obj: `str`)
+        Return:
+            A :obj:`list` | `dict`: will be serialized and returned
+        """
+        print('data["inputs"] = ', data["inputs"])
+        image_src = data["inputs"]
+        if isinstance(image_src, str):
+            if os.path.isfile(image_src):
+                image_ori = Image.open(image_src)
+            else:
+                response = requests.get(image_src)
+                image_data = BytesIO(response.content)
+                image_ori = Image.open(image_data)
+        else:
+            image_ori = Image.fromarray(image_src)
+
+        image = image_ori.convert('RGB')
+        # Preprocess the image
+        image_preprocessor = ImagePreprocessor(resolution=tuple(resolution))
+        image_proc = image_preprocessor.proc(image)
+        image_proc = image_proc.unsqueeze(0)
+
+        # Prediction
+        with torch.no_grad():
+            preds = self.birefnet(image_proc.to(device).half() if half_precision else image_proc.to(device))[-1].sigmoid().cpu()
+        pred = preds[0].squeeze()
+
+        # Show Results
+        pred_pil = transforms.ToPILImage()(pred)
+        image_masked = refine_foreground(image, pred_pil)
+        image_masked.putalpha(pred_pil.resize(image.size))
+        return image_masked

extras/birefnet/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9ab37426bf4de0567af6b5d21b16151357149139362e6e8992021b8ce356a154
+size 444473596

extras/birefnet/requirements.txt

@@ -0,0 +1,16 @@
+torch==2.5.1
+torchvision
+numpy<2
+opencv-python
+timm
+scipy
+scikit-image
+kornia
+einops
+
+tqdm
+prettytable
+
+transformers
+huggingface-hub>0.25
+accelerate

extras/dinov3-vitl16/.gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

extras/dinov3-vitl16/LICENSE.md

@@ -0,0 +1,66 @@
+# DINOv3 License
+
+*Last Updated: August 19, 2025*
+
+**“Agreement”** means the terms and conditions for use, reproduction, distribution and modification of the DINO Materials set forth herein.
+
+**“DINO Materials”** means, collectively, Documentation and the models, software and algorithms, including machine-learning model code, trained model weights, inference-enabling code, training-enabling code, fine-tuning enabling code, and other elements of the foregoing distributed by Meta and made available under this Agreement.
+
+**“Documentation”** means the specifications, manuals and documentation accompanying
+DINO Materials distributed by Meta.
+
+**“Licensee”** or **“you”** means you, or your employer or any other person or entity (if you are entering into this Agreement on such person or entity’s behalf), of the age required under applicable laws, rules or regulations to provide legal consent and that has legal authority to bind your employer or such other person or entity if you are entering in this Agreement on their behalf.
+
+**“Meta”** or **“we”** means Meta Platforms Ireland Limited (if you are located in or, if you are an entity, your principal place of business is in the EEA or Switzerland) or Meta Platforms, Inc. (if you are located outside of the EEA or Switzerland).
+
+**“Sanctions”** means any economic or trade sanctions or restrictions administered or enforced by the United States (including the Office of Foreign Assets Control of the U.S. Department of the Treasury (“OFAC”), the U.S. Department of State and the U.S. Department of Commerce), the United Nations, the European Union, or the United Kingdom.
+
+**“Trade Controls”** means any of the following: Sanctions and applicable export and import controls.
+
+By clicking “I Accept” below or by using or distributing any portion or element of the DINO Materials, you agree to be bound by this Agreement.
+
+## 1. License Rights and Redistribution.
+
+a. <ins>Grant of Rights</ins>. You are granted a non-exclusive, worldwide, non-transferable and royalty-free limited license under Meta’s intellectual property or other rights owned by Meta embodied in the DINO Materials to use, reproduce, distribute, copy, create derivative works of, and make modifications to the DINO Materials.
+
+b. <ins>Redistribution and Use</ins>.
+
+i. Distribution of DINO Materials, and any derivative works thereof, are subject to the terms of this Agreement. If you distribute or make the DINO Materials, or any derivative works thereof, available to a third party, you may only do so under the terms of this Agreement and you shall provide a copy of this Agreement with any such DINO Materials.
+
+ii.  If you submit for publication the results of research you perform on, using, or otherwise in connection with DINO Materials, you must acknowledge the use of DINO Materials in your publication.
+
+iii. Your use of the DINO Materials must comply with applicable laws and regulations, including Trade Control Laws and applicable privacy and data protection laws.
+
+iv. Your use of the DINO Materials will not involve or encourage others to reverse engineer, decompile or discover the underlying components of the DINO Materials.
+
+v. You are not the target of Trade Controls and your use of DINO Materials must comply with Trade Controls. You agree not to use, or permit others to use, DINO Materials for any activities subject to the International Traffic in Arms Regulations (ITAR) or end uses prohibited by Trade Controls, including those related to military or warfare purposes, nuclear industries or applications, espionage, or the development or use of guns or illegal weapons.
+
+## 2. User Support.
+
+Your use of the DINO Materials is done at your own discretion; Meta does not process any information nor provide any service in relation to such use.  Meta is under no obligation to provide any support services for the DINO Materials. Any support provided is “as is”, “with all faults”, and without warranty of any kind.
+
+## 3. Disclaimer of Warranty.
+
+UNLESS REQUIRED BY APPLICABLE LAW, THE DINO MATERIALS AND ANY OUTPUT AND RESULTS THEREFROM ARE PROVIDED ON AN “AS IS” BASIS, WITHOUT WARRANTIES OF ANY KIND, AND META DISCLAIMS ALL WARRANTIES OF ANY KIND, BOTH EXPRESS AND IMPLIED, INCLUDING, WITHOUT LIMITATION, ANY WARRANTIES OF TITLE, NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. YOU ARE SOLELY RESPONSIBLE FOR DETERMINING THE APPROPRIATENESS OF USING OR REDISTRIBUTING THE DINO MATERIALS AND ASSUME ANY RISKS ASSOCIATED WITH YOUR USE OF THE DINO MATERIALS AND ANY OUTPUT AND RESULTS.
+
+## 4. Limitation of Liability.
+
+IN NO EVENT WILL META OR ITS AFFILIATES BE LIABLE UNDER ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, TORT, NEGLIGENCE, PRODUCTS LIABILITY, OR OTHERWISE, ARISING OUT OF THIS AGREEMENT, FOR ANY LOST PROFITS OR ANY DIRECT OR INDIRECT, SPECIAL, CONSEQUENTIAL, INCIDENTAL, EXEMPLARY OR PUNITIVE DAMAGES, EVEN IF META OR ITS AFFILIATES HAVE BEEN ADVISED OF THE POSSIBILITY OF ANY OF THE FOREGOING.
+
+## 5. Intellectual Property.
+
+a. Subject to Meta’s ownership of DINO Materials and derivatives made by or for Meta, with respect to any derivative works and modifications of the DINO Materials that are made by you, as between you and Meta, you are and will be the owner of such derivative works and modifications.
+
+b. If you institute litigation or other proceedings against Meta or any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the DINO Materials, outputs or results, or any portion of any of the foregoing, constitutes infringement of intellectual property or other rights owned or licensable by you, then any licenses granted to you under this Agreement shall terminate as of the date such litigation or claim is filed or instituted. You will indemnify and hold harmless Meta from and against any claim by any third party arising out of or related to your use or distribution of the DINO Materials.
+
+## 6. Term and Termination.
+
+The term of this Agreement will commence upon your acceptance of this Agreement or access to the DINO Materials and will continue in full force and effect until terminated in accordance with the terms and conditions herein. Meta may terminate this Agreement if you are in breach of any term or condition of this Agreement. Upon termination of this Agreement, you shall delete and cease use of the DINO Materials. Sections 3, 4 and 7 shall survive the termination of this Agreement.
+
+## 7. Governing Law and Jurisdiction.
+
+This Agreement will be governed and construed under the laws of the State of California without regard to choice of law principles, and the UN Convention on Contracts for the International Sale of Goods does not apply to this Agreement. The courts of California shall have exclusive jurisdiction of any dispute arising out of this Agreement.
+
+## 8. Modifications and Amendments.
+
+Meta may modify this Agreement from time to time; provided that they are similar in spirit to the current version of the Agreement, but may differ in detail to address new problems or concerns. All such changes will be effective immediately. Your continued use of the DINO Materials after any modification to this Agreement constitutes your agreement to such modification. Except as provided in this Agreement, no modification or addition to any provision of this Agreement will be binding unless it is in writing and signed by an authorized representative of both you and Meta.

extras/dinov3-vitl16/README.md

@@ -0,0 +1,477 @@
+---
+extra_gated_fields:
+  First Name: text
+  Last Name: text
+  Date of birth: date_picker
+  Country: country
+  Affiliation: text
+  Job title:
+    type: select
+    options:
+      - Student
+      - Research Graduate
+      - AI researcher
+      - AI developer/engineer
+      - Reporter
+      - Other
+  geo: ip_location
+  By clicking Submit below I accept the terms of the license and acknowledge that the information I provide will be collected stored processed and shared in accordance with the Meta Privacy Policy: checkbox
+extra_gated_description: >-
+  The information you provide will be collected, stored, processed and shared in
+  accordance with the [Meta Privacy
+  Policy](https://www.facebook.com/privacy/policy/).
+extra_gated_button_content: Submit
+language:
+- en
+tags:
+- dino
+- dinov3
+- arxiv:2508.10104
+license: other
+license_name: dinov3-license
+license_link: https://ai.meta.com/resources/models-and-libraries/dinov3-license
+base_model: facebook/dinov3-vit7b16-pretrain-lvd1689m
+pipeline_tag: image-feature-extraction
+library_name: transformers
+---
+
+# Model Card for DINOv3
+
+DINOv3 is a family of versatile vision foundation models that outperforms the specialized state of the art across a broad range of settings, without fine-tuning. DINOv3 produces high-quality dense features that achieve outstanding performance on various vision tasks, significantly surpassing previous self- and weakly-supervised foundation models.
+
+## Model Details
+
+These are Vision Transformer and ConvNeXt models trained following the method described in the DINOv3 paper. 12 models are provided:
+
+- 10 models pretrained on web data (LVD-1689M dataset)
+  - 1 ViT-7B trained from scratch,
+  - 5 ViT-S/S+/B/L/H+ models distilled from the ViT-7B,
+  - 4 ConvNeXt-{T/S/B/L} models distilled from the ViT-7B,
+- 2 models pretrained on satellite data (SAT-493M dataset)
+  - 1 ViT-7B trained from scratch
+  - 1 ViT-L distilled from the ViT-7B
+
+
+Each Transformer-based model takes an image as input and returns a class token, patch tokens (and register tokens). These models follow a ViT architecture, with a patch size of 16. For a 224x224 image, this results in 1 class token + 4 register tokens + 196 patch tokens = 201 tokens (for DINOv2 with registers this resulted in 1 + 4 + 256 = 261 tokens).
+
+The models can accept larger images provided the image shapes are multiples of the patch size (16). If this condition is not verified, the model will crop to the closest smaller multiple of the patch size.
+
+### Model Description
+
+- **Developed by:** Meta AI
+- **Model type:** Vision Transformer, ConvNeXt
+- **License:** [DINOv3 License](https://ai.meta.com/resources/models-and-libraries/dinov3-license/)
+
+### Model Sources
+
+- **Repository:** [https://github.com/facebookresearch/dinov3](https://github.com/facebookresearch/dinov3)
+- **Paper:** [https://arxiv.org/abs/2508.10104](https://arxiv.org/abs/2508.10104)
+
+## Uses
+
+The models are vision backbones providing multi-purpose features for downstream tasks.
+
+### Direct Use
+
+The models can be used without fine-tuning, with downstream classifiers as simple as linear layers, to obtain competitive results:
+
+- on image classification, using k-NN classifiers on the class token
+- on image classification, with logistic regression classifiers applied on the class token
+- on image classification, with a linear layer applied on the class token and the average of the patch tokens
+- on image retrieval using nearest neighbors
+- on geometric and semantic 3D keypoint correspondances
+- on depth estimation, semantic segmentation, using linear layers
+- on unsupervised object discovery
+- on video segmentation tracking
+- on video classification, using a small 4-layer attentive probe
+
+### Downstream Use
+
+While fine-tuning the models can yield some gains, it is recommended to keep this option as a last resort: the frozen features are expected to provide good performance out-of-the-box.
+
+## Bias, Risks, and Limitations
+
+Compared to DINOv2 and SEERv2, DINOv3 delivers somewhat consistent performance across income categories on geographical fairness and diversity, although with a notable performance drop in the low-income bucket compared to the highest-income bucket.
+
+DINOv3 also achieves relatively good scores across different regions, improving over its predecessor DINOv2. However, a relative difference is still observed between Europe and Africa.
+
+### Recommendations
+
+Fine-tuning is expected to increase the biases in the features produced by the model as they will be tuned to the fine-tuning labels.
+
+## How to Get Started with the Model
+
+The example below demonstrates how to obtain an image embedding with [Pipeline] or the [AutoModel] class.
+
+```python
+from transformers import pipeline
+from transformers.image_utils import load_image
+
+url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"
+image = load_image(url)
+
+feature_extractor = pipeline(
+    model="facebook/dinov3-vitl16-pretrain-lvd1689m",
+    task="image-feature-extraction", 
+)
+features = feature_extractor(image)
+```
+
+```python
+import torch
+from transformers import AutoImageProcessor, AutoModel
+from transformers.image_utils import load_image
+
+url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+image = load_image(url)
+
+pretrained_model_name = "facebook/dinov3-vitl16-pretrain-lvd1689m"
+processor = AutoImageProcessor.from_pretrained(pretrained_model_name)
+model = AutoModel.from_pretrained(
+    pretrained_model_name, 
+    device_map="auto", 
+)
+
+inputs = processor(images=image, return_tensors="pt").to(model.device)
+with torch.inference_mode():
+    outputs = model(**inputs)
+
+pooled_output = outputs.pooler_output
+print("Pooled output shape:", pooled_output.shape)
+```
+
+## Training Details
+
+### Training Data
+
+- Web dataset (LVD-1689M): a curated dataset of 1,689 millions of images extracted from a large data
+pool of 17 billions web images collected from public posts on Instagram
+
+- Satellite dataset (SAT-493M): a dataset of 493 millions of 512x512 images sampled randomly from Maxar RGB ortho-rectified imagery at 0.6 meter resolution
+
+### Training Procedure
+
+**Training objective:**
+
+- DINO self-distillation loss with multi-crop
+- iBOT masked-image modeling loss
+- KoLeo regularization on [CLS] tokens
+- Gram anchoring
+
+- **Training regime:** PyTorch FSDP2 (with bf16 and fp8 matrix multiplications)
+
+**Distillation:**
+
+- Distillation follows the standard DINOv3 pretraining procedure, except the teacher is a frozen pretrained ViT-7B.
+
+## Evaluation
+
+**Results**
+
+The reader is referred to the associated paper for details on the evaluation protocols
+
+*Results for ViT backbones pretrained (or distilled) on web (LVD-1689M)*
+
+<table>
+  <tr>
+    <th></th>
+    <!-- <th></th> -->
+    <th colspan="4">Global Tasks</th>
+    <th colspan="5">Dense Tasks</th>
+  </tr>
+  <tr>
+    <th>Model</th>
+    <!-- <th>Dataset</th> -->
+    <th>IN-ReaL</th>
+    <th>IN-R</th>
+    <th>Obj.Net</th>
+    <th>Ox.-H</th>
+    <th>ADE20k</th>
+    <th>NYU↓</th>
+    <th>DAVIS</th>
+    <th>NAVI</th>
+    <th>SPair</th>
+  </tr>
+  <tr>
+    <td>DINOv3 ViT-S/16</td>
+    <!-- <td>LVD-1689M</td> -->
+    <td align="right">87.0</td>
+    <td align="right">60.4</td>
+    <td align="right">50.9</td>
+    <td align="right">49.5</td>
+    <td align="right">47.0</td>
+    <td align="right">0.403</td>
+    <td align="right">72.7</td>
+    <td align="right">56.3</td>
+    <td align="right">50.4</td>
+  </tr>
+  <tr>
+    <td>DINOv3 ViT-S+/16</td>
+    <!-- <td>LVD-1689M</td> -->
+    <td align="right">88.0</td>
+    <td align="right">68.8</td>
+    <td align="right">54.6</td>
+    <td align="right">50.0</td>
+    <td align="right">48.8</td>
+    <td align="right">0.399</td>
+    <td align="right">75.5</td>
+    <td align="right">57.1</td>
+    <td align="right">55.2</td>
+  </tr>
+  <tr>
+    <td>DINOv3 ViT-B/16</td>
+    <!-- <td>LVD-1689M</td> -->
+    <td align="right">89.3</td>
+    <td align="right">76.7</td>
+    <td align="right">64.1</td>
+    <td align="right">58.5</td>
+    <td align="right">51.8</td>
+    <td align="right">0.373</td>
+    <td align="right">77.2</td>
+    <td align="right">58.8</td>
+    <td align="right">57.2</td>
+  </tr>
+  <tr>
+    <td>DINOv3 ViT-L/16</td>
+    <!-- <td>LVD-1689M</td> -->
+    <td align="right">90.2</td>
+    <td align="right">88.1</td>
+    <td align="right">74.8</td>
+    <td align="right">63.1</td>
+    <td align="right">54.9</td>
+    <td align="right">0.352</td>
+    <td align="right">79.9</td>
+    <td align="right">62.3</td>
+    <td align="right">61.3</td>
+  </tr>
+  <tr>
+    <td>DINOv3 ViT-H+/16</td>
+    <!-- <td>LVD-1689M</td> -->
+    <td align="right">90.3</td>
+    <td align="right">90.0</td>
+    <td align="right">78.6</td>
+    <td align="right">64.5</td>
+    <td align="right">54.8</td>
+    <td align="right">0.352</td>
+    <td align="right">79.3</td>
+    <td align="right">63.3</td>
+    <td align="right">56.3</td>
+  </tr>
+  <tr>
+    <td>DINOv3 ViT-7B/16</td>
+    <!-- <td>LVD-1689M</td> -->
+    <td align="right">90.4</td>
+    <td align="right">91.1</td>
+    <td align="right">91.1</td>
+    <td align="right">72.8</td>
+    <td align="right">55.9</td>
+    <td align="right">0.309</td>
+    <td align="right">79.7</td>
+    <td align="right">64.4</td>
+    <td align="right">58.7</td>
+  </tr>
+</table>
+
+*Results for ConvNeXt backbones distilled on web (LVD-1689M)*
+
+<table>
+  <tr>
+    <th></th>
+    <th colspan="6">Global Tasks</th>
+    <th colspan="2">Dense Tasks</th>
+  </tr>
+  <tr>
+    <th>Model</th>
+    <th colspan="2">IN-ReaL</th>
+    <th colspan="2">IN-R</th>
+    <th colspan="2">Obj.Net</th>
+    <th>ADE20k</th>
+    <th>NYU↓</th>
+  </tr>
+  <tr>
+    <td></th>
+    <td>@256px</td>
+    <td>@512px</td>
+    <td>@256px</td>
+    <td>@512px</td>
+    <td>@256px</td>
+    <td>@512px</td>
+    <td colspan="2"></td>
+  </tr>
+  <tr>
+    <td>DINOv3 ConvNeXt Tiny</td>
+    <td align="right">86.6</td>
+    <td align="right">87.7</td>
+    <td align="right">73.7</td>
+    <td align="right">74.1</td>
+    <td align="right">52.6</td>
+    <td align="right">58.7</td>
+    <td align="right">42.7</td>
+    <td align="right">0.448</td>
+  </tr>
+  <tr>
+    <td>DINOv3 ConvNeXt Small</td>
+    <td align="right">87.9</td>
+    <td align="right">88.7</td>
+    <td align="right">73.7</td>
+    <td align="right">74.1</td>
+    <td align="right">52.6</td>
+    <td align="right">58.7</td>
+    <td align="right">44.8</td>
+    <td align="right">0.432</td>
+  </tr>
+  <tr>
+    <td>DINOv3 ConvNeXt Base</td>
+    <td align="right">88.5</td>
+    <td align="right">89.2</td>
+    <td align="right">77.2</td>
+    <td align="right">78.2</td>
+    <td align="right">56.2</td>
+    <td align="right">61.3</td>
+    <td align="right">46.3</td>
+    <td align="right">0.420</td>
+  </tr>
+  <tr>
+    <td>DINOv3 ConvNeXt Large</td>
+    <td align="right">88.9</td>
+    <td align="right">89.4</td>
+    <td align="right">81.3</td>
+    <td align="right">82.4</td>
+    <td align="right">59.3</td>
+    <td align="right">65.2</td>
+    <td align="right">47.8</td>
+    <td align="right">0.403</td>
+  </tr>
+</table>
+
+*Results for ViT backbones pretrained (or distilled) on satellite (SAT-493M)*
+
+<table>
+  <tr>
+    <th></th>
+    <th colspan="7">(GEO-Bench) Classification</th>
+  </tr>
+  <tr>
+    <th>Model</ht>
+    <th>m-BEnet</th>
+    <th>m-brick-kiln
+    <th>m-eurosat</th>
+    <th>m-forestnet</th>
+    <th>m-pv4ger</th>
+    <th>m-so2sat</th>
+    <th>mean</th>
+  </tr>
+  <tr>
+    <td>DINOv3 ViT-L/16</td>
+    <td>73.0</td>
+    <td>96.5</td>
+    <td>94.1</td>
+    <td>60.6</td>
+    <td>96.0</td>
+    <td>57.4</td>
+    <td>79.6</td>
+  </tr>
+  <tr>
+    <td>DINOv3 ViT-7B/16</td>
+    <td>74.0</td>
+    <td>97.2</td>
+    <td>94.8</td>
+    <td>62.3</td>
+    <td>96.1</td>
+    <td>62.1</td>
+    <td>81.1</td>
+  </tr>
+  <tr>
+    <th></th>
+    <th colspan="7">(GEO-Bench) Segmentation</th>
+  </tr>
+  <tr>
+    <th>Model</th>
+    <th>m-cashew</th>
+    <th>m-chesapeake</th>
+    <th>m-NeonTree</th>
+    <th>m-nz-cattle</th>
+    <th>m-pv4ger-seg</th>
+    <th>m-SA-crop</th>
+    <th>mean</th>
+  </tr>
+  <tr>
+    <td>DINOv3 ViT-L/16</td>
+    <td>94.2</td>
+    <td>75.6</td>
+    <td>61.8</td>
+    <td>83.7</td>
+    <td>95.2</td>
+    <td>36.8</td>
+    <td>74.5</td>
+  </tr>
+  <tr>
+    <td>DINOv3 ViT-7B/16</td>
+    <td>94.1</td>
+    <td>76.6</td>
+    <td>62.6</td>
+    <td>83.4</td>
+    <td>95.5</td>
+    <td>37.6</td>
+    <td>75.0</td>
+  </tr>
+</table>
+
+
+## Environmental Impact
+
+- **Hardware Type:** Nvidia H100
+- **Hours used:** 61,440 hours for ViT-7B model training
+- **Cloud Provider:** Private infrastructure
+- **Compute Region:** USA
+- **Carbon Emitted:** 18t CO2eq
+
+## Technical Specifications
+
+### Model Architecture and Objective
+
+Vision Transformer models:
+
+- ViT-S (21M parameters): patch size 16, embedding dimension 384, 4 register tokens, 6 heads, MLP FFN, RoPE
+- ViT-S+ (29M parameters): patch size 16, embedding dimension 384, 4 register tokens, 6 heads, SwiGLU FFN, RoPE
+- ViT-B (86M parameters): patch size 16, embedding dimension 768, 4 register tokens, 12 heads, MLP FFN, RoPE
+- ViT-L (300M parameters): patch size 16, embedding dimension 1024, 4 register tokens, 16 heads, MLP FFN, RoPE
+- ViT-H+ (840M parameters): patch size 16, embedding dimension 1280, 4 register tokens, 20 heads, SwiGLU FFN, RoPE
+- ViT-7B (6716M parameters): patch size 16, embedding dimension 4096, 4 register tokens, 32 heads, SwiGLU FFN, RoPE
+
+ConvNeXt models:
+
+- ConvNeXt Tiny (29M parameters)
+- ConvNeXt Small (50M parameters)
+- ConvNeXt Base (89M parameters)
+- ConvNeXt Large (198M parameters)
+
+### Compute Infrastructure
+
+#### Hardware
+
+Nvidia H100 GPUs
+
+#### Software
+
+PyTorch 2.7
+
+## More Information
+
+See the [blog post](https://ai.meta.com/blog/dinov3-self-supervised-vision-model/) and the associated [website](https://ai.meta.com/dinov3/).
+
+## Citation
+
+**BibTeX**
+
+```
+@misc{simeoni2025dinov3,
+  title={{DINOv3}},
+  author={Sim{\'e}oni, Oriane and Vo, Huy V. and Seitzer, Maximilian and Baldassarre, Federico and Oquab, Maxime and Jose, Cijo and Khalidov, Vasil and Szafraniec, Marc and Yi, Seungeun and Ramamonjisoa, Micha{\"e}l and Massa, Francisco and Haziza, Daniel and Wehrstedt, Luca and Wang, Jianyuan and Darcet, Timoth{\'e}e and Moutakanni, Th{\'e}o and Sentana, Leonel and Roberts, Claire and Vedaldi, Andrea and Tolan, Jamie and Brandt, John and Couprie, Camille and Mairal, Julien and J{\'e}gou, Herv{\'e} and Labatut, Patrick and Bojanowski, Piotr},
+  year={2025},
+  eprint={2508.10104},
+  archivePrefix={arXiv},
+  primaryClass={cs.CV},
+  url={https://arxiv.org/abs/2508.10104},
+}
+```

extras/dinov3-vitl16/config.json

@@ -0,0 +1,32 @@
+{
+  "architectures": [
+    "DINOv3ViTModel"
+  ],
+  "attention_dropout": 0.0,
+  "drop_path_rate": 0.0,
+  "hidden_act": "gelu",
+  "hidden_size": 1024,
+  "image_size": 224,
+  "initializer_range": 0.02,
+  "intermediate_size": 4096,
+  "key_bias": false,
+  "layer_norm_eps": 1e-05,
+  "layerscale_value": 1.0,
+  "mlp_bias": true,
+  "model_type": "dinov3_vit",
+  "num_attention_heads": 16,
+  "num_channels": 3,
+  "num_hidden_layers": 24,
+  "num_register_tokens": 4,
+  "patch_size": 16,
+  "pos_embed_jitter": null,
+  "pos_embed_rescale": 2.0,
+  "pos_embed_shift": null,
+  "proj_bias": true,
+  "query_bias": true,
+  "rope_theta": 100.0,
+  "torch_dtype": "float32",
+  "transformers_version": "4.56.0.dev0",
+  "use_gated_mlp": false,
+  "value_bias": true
+}

extras/dinov3-vitl16/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dcb2e45127cccbf1601e5f42fef165eea275c8e5213197e8dcf3f48822718179
+size 1212559808

extras/dinov3-vitl16/preprocessor_config.json

@@ -0,0 +1,31 @@
+{
+  "crop_size": null,
+  "data_format": "channels_first",
+  "default_to_square": true,
+  "device": null,
+  "disable_grouping": null,
+  "do_center_crop": null,
+  "do_convert_rgb": null,
+  "do_normalize": true,
+  "do_rescale": true,
+  "do_resize": true,
+  "image_mean": [
+    0.485,
+    0.456,
+    0.406
+  ],
+  "image_processor_type": "DINOv3ViTImageProcessorFast",
+  "image_std": [
+    0.229,
+    0.224,
+    0.225
+  ],
+  "input_data_format": null,
+  "resample": 2,
+  "rescale_factor": 0.00392156862745098,
+  "return_tensors": null,
+  "size": {
+    "height": 224,
+    "width": 224
+  }
+}

extras/moge-2-vitl/.gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

extras/moge-2-vitl/README.md

@@ -0,0 +1,3 @@
+---
+license: mit
+---

extras/moge-2-vitl/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3eefd4abb2102f38f12b2d1992e5ff15e4923e5431c67dd494afe157e0111cd5
+size 1305030700

extras/rmbg-2.0/README.md

@@ -0,0 +1,238 @@
+---
+license: other
+license_name: bria-rmbg-2.0
+license_link: https://creativecommons.org/licenses/by-nc/4.0/deed.en
+pipeline_tag: image-segmentation
+tags:
+- remove background
+- background
+- background-removal
+- Pytorch
+- vision
+- legal liability
+- transformers
+- transformers.js
+extra_gated_description: >-
+  Bria AI Model weights are open source for non commercial use only, per the
+  provided [license](https://creativecommons.org/licenses/by-nc/4.0/deed.en).
+extra_gated_heading: Fill in this form to immediatly access the model for non commercial use
+extra_gated_fields:
+  Name: text
+  Email: text
+  Company/Org name: text
+  Company Website URL: text
+  Discord user: text
+  I agree to BRIA’s Privacy policy, Terms & conditions, and acknowledge Non commercial use to be Personal use / Academy / Non profit (direct or indirect): checkbox
+---
+
+# BRIA Background Removal v2.0 Model Card
+<p align="center"><img src="https://bria-public.s3.us-east-1.amazonaws.com/Bria-logo.svg" alt="BRIA Logo" width="200" /></p>
+
+<!-- RMBG Card wrapper -->
+<div class="rmbg-card" style="position: relative; border-radius: 12px; overflow: hidden;">
+
+  <!-- FIBO Promo Banner (Top) -->
+  <a
+    href="https://huggingface.co/briaai/FIBO"
+    target="_blank"
+    rel="noopener"
+    aria-label="Explore FIBO on Hugging Face"
+    style="
+      position: absolute;
+      top: 0;
+      left: 0;
+      width: 100%;
+      display: flex;
+      align-items: center;
+      justify-content: center;
+      gap: 10px;
+      background: linear-gradient(90deg, #fff6b7 0%, #fde047 100%);
+      color: #1f2937;
+      text-decoration: none;
+      font-family: Inter, system-ui, -apple-system, Segoe UI, Roboto, Arial, sans-serif;
+      font-weight: 600;
+      font-size: 13px;
+      padding: 10px 0;
+      border-bottom: 1px solid rgba(0,0,0,0.08);
+      box-shadow: 0 2px 8px rgba(0,0,0,0.08);
+      z-index: 10;
+    ">
+    <img
+      src="https://huggingface.co/front/assets/huggingface_logo-noborder.svg"
+      alt="Hugging Face"
+      width="18"
+      height="18"
+      style="display:block"
+    />
+    <span>✨ Discover <strong>FIBO</strong> on Hugging Face</span>
+  </a>
+
+  <!-- ... your RMBG content below ... -->
+<p align="center">
+         💜 <a href="https://go.bria.ai/46gzn20"><b>Bria AI</b></a>&nbsp&nbsp | &nbsp&nbsp🤗 <a href="https://huggingface.co/briaai/">Hugging Face</a> &nbsp&nbsp | &nbsp&nbsp 📑 <a href="https://blog.bria.ai/">Blog</a> &nbsp&nbsp 
+<br>
+🖥️ <a href="https://huggingface.co/spaces/briaai/BRIA-RMBG-2.0">Demo</a>&nbsp&nbsp| &nbsp&nbsp <a href="https://github.com/Bria-AI/RMBG-2.0">Github</a>&nbsp&nbsp
+</p>
+
+RMBG v2.0 is our new state-of-the-art background removal model significantly improves RMBG v1.4. The model is designed to effectively separate foreground from background in a range of
+categories and image types. This model has been trained on a carefully selected dataset, which includes:
+general stock images, e-commerce, gaming, and advertising content, making it suitable for commercial use cases powering enterprise content creation at scale. 
+The accuracy, efficiency, and versatility currently rival leading source-available models. 
+It is ideal where content safety, legally licensed datasets, and bias mitigation are paramount. 
+
+**[→ Try the API Sandbox (no signup required)](https://catalog.bria.ai/image-editing/remove-background/sandbox)**
+
+Developed by BRIA AI, RMBG v2.0 is available as a source-available model for non-commercial use.
+
+### Get Access
+
+Bria RMBG2.0 is availabe everywhere you build, either as source-code and weights, ComfyUI nodes or API endpoints.
+
+
+-
+
+
+#####
+### Model Description
+
+**For production / commercial deployment**, use the Bria API — same RMBG-2.0 quality, fully licensed, zero infrastructure:
+
+| Use                   | Self-Hosted (HF Weights) | Bria API    |
+| :-------------------- |:-------------------------| :---------- |
+| Quality               | ✅ RMBG-2.0              | ✅ RMBG-2.0 |
+| Commercial License    | ❌ Requires agreement    | ✅ Included |
+| GPU Infrastructure    | ❌ You manage            | ✅ Managed  |
+| Legally Licensed Data | ✅ Yes                   | ✅ Yes      |
+| Setup Time            | Hours                    | Minutes     |
+
+**[→ Try the API Sandbox](https://catalog.bria.ai/image-editing/remove-background/sandbox)** — test it live, no signup required.
+
+
+
+## Model Details
+- **Developed by:** [BRIA AI](https://bria.ai/)
+- **Model type:** Background Removal 
+- **License:** [Creative Commons Attribution–Non-Commercial (CC BY-NC 4.0)](https://creativecommons.org/licenses/by-nc/4.0/deed.en)
+  - The model is released under a CC BY-NC 4.0 license for non-commercial use.
+  - Commercial use is subject to a commercial agreement with BRIA. Available [here](https://share-eu1.hsforms.com/2sj9FVZTGSFmFRibDLhr_ZAf4e04)
+
+ **API Endpoint**: [Sandbox](https://catalog.bria.ai/image-editing/remove-background/sandbox)
+- **ComfyUI**: [Use it in workflows](https://github.com/Bria-AI/ComfyUI-BRIA-API)
+- **GitHub**: [github.com/Bria-AI/RMBG-2.0](https://github.com/Bria-AI/RMBG-2.0)
+
+ **Purchase:** To purchase a Self-Hosted (HF Weights) commercial license [Click Here](https://share-eu1.hsforms.com/2sj9FVZTGSFmFRibDLhr_ZAf4e04).
+
+For more information, please visit our [website](https://catalog.bria.ai/image-editing/remove-background/sandbox).
+
+Join our [Discord community](https://discord.gg/Nxe9YW9zHS) for more information, tutorials, tools, and to connect with other users!
+
+[CLICK HERE FOR A DEMO](https://huggingface.co/spaces/briaai/BRIA-RMBG-2.0)
+
+[fal.ai](https://fal.ai/models/fal-ai/bria/background/remove), [Replicate](https://replicate.com/bria/remove-background)
+
+![examples](t4.png)
+
+- **Model Description:** BRIA RMBG-2.0 is a dichotomous image segmentation model trained exclusively on a professional-grade dataset. The model output includes a single-channel 8-bit grayscale alpha matte, where each pixel value indicates the opacity level of the corresponding pixel in the original image. This non-binary output approach offers developers the flexibility to define custom thresholds for foreground-background separation, catering to varied use cases requirements and enhancing integration into complex pipelines.
+- **BRIA:** Resources for more information: [BRIA AI](https://bria.ai/)
+
+
+
+## Training data
+Bria-RMBG model was trained with over 15,000 high-quality, high-resolution, manually labeled (pixel-wise accuracy), fully licensed images.
+Our benchmark included balanced gender, balanced ethnicity, and people with different types of disabilities.
+For clarity, we provide our data distribution according to different categories, demonstrating our model’s versatility.
+
+### Distribution of images:
+
+| Category | Distribution |
+| -----------------------------------| -----------------------------------:|
+| Objects only | 45.11% |
+| People with objects/animals | 25.24% |
+| People only | 17.35% |
+| people/objects/animals with text | 8.52% |
+| Text only | 2.52% |
+| Animals only | 1.89% |
+
+| Category | Distribution |
+| -----------------------------------| -----------------------------------------:|
+| Photorealistic | 87.70% |
+| Non-Photorealistic | 12.30% |
+
+
+| Category | Distribution |
+| -----------------------------------| -----------------------------------:|
+| Non Solid Background | 52.05% |
+| Solid Background | 47.95% 
+
+
+| Category | Distribution |
+| -----------------------------------| -----------------------------------:|
+| Single main foreground object | 51.42% |
+| Multiple objects in the foreground | 48.58% |
+
+
+## Qualitative Evaluation
+Open source models comparison
+![diagram](diagram1.png)
+![examples](collage5.png)
+
+### Architecture
+RMBG-2.0 is developed on the [BiRefNet](https://github.com/ZhengPeng7/BiRefNet) architecture enhanced with our proprietary dataset and training scheme. This training data significantly improves the model’s accuracy and effectiveness for background-removal task.<br>
+If you use this model in your research, please cite:
+
+```
+@article{BiRefNet,
+  title={Bilateral Reference for High-Resolution Dichotomous Image Segmentation},
+  author={Zheng, Peng and Gao, Dehong and Fan, Deng-Ping and Liu, Li and Laaksonen, Jorma and Ouyang, Wanli and Sebe, Nicu},
+  journal={CAAI Artificial Intelligence Research},
+  year={2024}
+}
+```
+
+#### Requirements
+```bash
+torch
+torchvision
+pillow
+kornia
+transformers
+```
+
+### Usage
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+
+```python
+from PIL import Image
+import torch
+from torchvision import transforms
+from transformers import AutoModelForImageSegmentation
+
+device = 'cuda' if torch.cuda.is_available() else 'cpu'
+model = AutoModelForImageSegmentation.from_pretrained('briaai/RMBG-2.0', trust_remote_code=True).eval().to(device)
+
+# Data settings
+image_size = (1024, 1024)
+transform_image = transforms.Compose([
+    transforms.Resize(image_size),
+    transforms.ToTensor(),
+    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
+])
+
+image = Image.open(input_image_path)
+input_images = transform_image(image).unsqueeze(0).to(device)
+
+# Prediction
+with torch.no_grad():
+    preds = model(input_images)[-1].sigmoid().cpu()
+pred = preds[0].squeeze()
+pred_pil = transforms.ToPILImage()(pred)
+mask = pred_pil.resize(image.size)
+image.putalpha(mask)
+
+image.save("no_bg_image.png")
+```
+
+
+</div>