diff --git a/README.MD b/README.MD index ea4b83f..22c49f4 100644 --- a/README.MD +++ b/README.MD @@ -87,6 +87,7 @@ When this error has occurred, please check the network environment. ## Update **If the dependency package error after updating, please reinstall the relevant dependency packages. for details, please refer to [here](https://github.com/chflame163/ComfyUI_LayerStyle/issues/5).
+* Commit [BiRefNetUltra](#BiRefNetUltra) node, it using the BiRefNet model to remove background has better recognition ability and ultra-high edge details. * Commit [ImageAutoCropV2](#ImageAutoCropV2) node, it can choose not to remove the background, support mask input, and scale by long or short side size. * Commit [ImageHub](#ImageHub) node, supports up to 9 sets of Image and Mask switching output, and supports random output. * Commit [TextJoin](#TextJoin) node. @@ -627,7 +628,7 @@ Node option: * [note](#notes) ### ColorPicker -Modify web extensions from [mtb nodes](https://github.com/melMass/comfy_mtb). Select colors on the color palette and output RGB values. +Modify web extensions from [mtb nodes](https://github.com/melMass/comfy_mtb). Select colors on the color palette and output RGB values, thanks to the original author. ![image](image/color_picker.png) Node options: @@ -752,7 +753,7 @@ The following changes have been made based on GradientImage: ### SimpleTextImage ![image](image/simple_text_image_example.png) -Generate simple typesetting images and masks from text. This node references some of the functionalities and code of [ZHO-ZHO-ZHO/ComfyUI-Text_Image-Composite](https://github.com/ZHO-ZHO-ZHO/ComfyUI-Text_Image-Composite). +Generate simple typesetting images and masks from text. This node references some of the functionalities and code of [ZHO-ZHO-ZHO/ComfyUI-Text_Image-Composite](https://github.com/ZHO-ZHO-ZHO/ComfyUI-Text_Image-Composite), thanks to the original author. ![image](image/simple_text_image_node.png) Node options: @@ -809,7 +810,7 @@ If the folder set in ini is invalid, the font folder that comes with the plugin ### LaMa ![image](image/lama_example.png) -Erase objects from the image based on the mask. this node is repackage of [IOPaint](https://www.iopaint.com), powered by state-of-the-art AI models. +Erase objects from the image based on the mask. this node is repackage of [IOPaint](https://www.iopaint.com), powered by state-of-the-art AI models, thanks to the original author. It is have [LaMa](https://github.com/advimman/lama), [LDM](https://github.com/CompVis/latent-diffusion), [ZITS](https://github.com/DQiaole/ZITS_inpainting),[MAT](https://github.com/fenglinglwb/MAT), [FcF](https://github.com/SHI-Labs/FcF-Inpainting), [Manga](https://github.com/msxie92/MangaInpainting) models and the SPREAD method to erase. Please refer to the original link for the introduction of each model. Please download the model files from [lama models(Baidu Netdisk)](https://pan.baidu.com/s/1LllR9TJHP1G9uEwWT3Mvkg?pwd=tvzv) or [lama models(Google Drive)](https://drive.google.com/drive/folders/1Aq0a4sybb3SRxi7j1e1_ZbBRjaWDdP9e?usp=sharing) to ```ComfyUI/models/lama``` folder. @@ -945,7 +946,7 @@ Output: * y: The y-coordinate of the top left corner position. ### SegmentAnythingUltra -Improvements to [ComfyUI Segment Anything](https://github.com/storyicon/comfyui_segment_anything), combined with the Alpha Matte node of [ComfyUI-Image-Filters](https://github.com/spacepxl/ComfyUI-Image-Filters) in spacepxl, result in more detailed edges for masks. +Improvements to [ComfyUI Segment Anything](https://github.com/storyicon/comfyui_segment_anything), combined with the Alpha Matte node of [ComfyUI-Image-Filters](https://github.com/spacepxl/ComfyUI-Image-Filters) in spacepxl, result in more detailed edges for masks, thanks to the original author. *Please refer to the installation of ComfyUI Segment Anything to install the model. If ComfyUI Segment Anything has been correctly installed, you can skip this step. * From [here](https://huggingface.co/bert-base-uncased/tree/main) download the config.json，model.safetensors，tokenizer_config.json，tokenizer.json and vocab.txt 5 files to ```ComfyUI/models/bert-base-uncased``` folder. @@ -984,7 +985,7 @@ On the basis of SegmentAnythingUltra, the following changes have been made: ### RemBgUltra Remove background. compared to the similar background removal nodes, this node has ultra-high edge details. -This node combines the Alpha Matte node of Spacepxl's [ComfyUI-Image-Filters](https://github.com/spacepxl/ComfyUI-Image-Filters) and the functionality of ZHO-ZHO-ZHO's [ComfyUI-BRIA_AI-RMBG](https://github.com/ZHO-ZHO-ZHO/ComfyUI-BRIA_AI-RMBG). +This node combines the Alpha Matte node of Spacepxl's [ComfyUI-Image-Filters](https://github.com/spacepxl/ComfyUI-Image-Filters) and the functionality of ZHO-ZHO-ZHO's [ComfyUI-BRIA_AI-RMBG](https://github.com/ZHO-ZHO-ZHO/ComfyUI-BRIA_AI-RMBG), thanks to the original author. *Download the [BRIA Background Removal v1.4](https://huggingface.co/briaai/RMBG-1.4) model file (```model.pth```) to the ```/ComfyUI/models/rmbg/RMBG-1.4``` folder. This model was developed by BRIA AI and can be used as an open-source model for non-commercial purposes. **After downloading the model, recommended to add a line 'rmbg: models/rmbg/RMBG-1.4' in ComfyUI/extra_model_paths.yaml. (This step is not mandatory) @@ -1007,9 +1008,29 @@ On the basis of RemBgUltra, the following changes have been made: * detail_erode: Mask the erosion range inward from the edge. the larger the value, the larger the range of inward repair. * detail_dilate: The edge of the mask expands outward. the larger the value, the wider the range of outward repair. + + +### BiRefNetUltra +Using the BiRefNet model to remove background has better recognition ability and ultra-high edge details. +The code for the model part of this node comes from Viper's [ComfyUI-BiRefNet](https://github.com/viperyl/ComfyUI-BiRefNet)，thanks to the original author. + +*From [https://huggingface.co/ViperYX/BiRefNet](https://huggingface.co/ViperYX/BiRefNet/tree/main) download the ```BiRefNet-ep480.pth```,```pvt_v2_b2.pth```,```pvt_v2_b5.pth```,```swin_base_patch4_window12_384_22kto1k.pth```, ```swin_large_patch4_window12_384_22kto1k.pth``` 5 files to ```ComfyUI/models/BiRefNet``` folder. + +![image](image/birefnet_ultra_example.png) + +Node options: +![image](image/birefnet_ultra_node.png) +* detail_method: Edge processing methods. provides three methods: VITMatte, PyMatting and GuidedFilter. +* detail_erode: Mask the erosion range inward from the edge. the larger the value, the larger the range of inward repair. +* detail_dilate: The edge of the mask expands outward. the larger the value, the wider the range of outward repair. +* black_point: Edge black sampling threshold. +* white_point: Edge white sampling threshold. +* process_detail: Set to false here will skip edge processing to save runtime. + + ### PersonMaskUltra Generate masks for portrait's face, hair, body skin, clothing, or accessories. Compared to the previous A Person Mask Generator node, this node has ultra-high edge details. -The model code for this node comes from [a-person-mask-generator](https://github.com/djbielejeski/a-person-mask-generator)， edge processing code from [ComfyUI-Image-Filters](https://github.com/spacepxl/ComfyUI-Image-Filters). +The model code for this node comes from [a-person-mask-generator](https://github.com/djbielejeski/a-person-mask-generator)， edge processing code from [ComfyUI-Image-Filters](https://github.com/spacepxl/ComfyUI-Image-Filters)，thanks to the original author. ![image](image/person_mask_ultra_example.png) @@ -1074,7 +1095,7 @@ Node options: ### MaskEdgeUltraDetail Process rough masks to ultra fine edges. -This node combines the Alpha Matte and the Guided Filter Alpha nodes functions of Spacepxl's [ComfyUI-Image-Filters](https://github.com/spacepxl/ComfyUI-Image-Filters). +This node combines the Alpha Matte and the Guided Filter Alpha nodes functions of Spacepxl's [ComfyUI-Image-Filters](https://github.com/spacepxl/ComfyUI-Image-Filters), thanks to the original author. ![image](image/mask_edge_ultra_detail_example.png) Node options: @@ -1239,7 +1260,7 @@ Node options: ### Film Simulate the grain, dark edge, and blurred edge of the film, support input depth map to simulate defocus. -This node is reorganize and encapsulate of [digitaljohn/comfyui-propost](https://github.com/digitaljohn/comfyui-propost). +This node is reorganize and encapsulate of [digitaljohn/comfyui-propost](https://github.com/digitaljohn/comfyui-propost), thanks to the original author. ![image](image/film_example.png) Node options: @@ -1309,4 +1330,4 @@ Node options: 5 The layer_image and layer_mask must be of the same size. # statement -LayerStyle nodes follows the MIT license, Some of its functional code comes from other open-source projects. If used for commercial purposes, please refer to the original project license to authorization agreement. +LayerStyle nodes follows the MIT license, Some of its functional code comes from other open-source projects. Thanks to the original author. If used for commercial purposes, please refer to the original project license to authorization agreement. diff --git a/README_CN.MD b/README_CN.MD index a3717f7..a5307f1 100644 --- a/README_CN.MD +++ b/README_CN.MD @@ -88,6 +88,7 @@ git clone https://github.com/chflame163/ComfyUI_LayerStyle.git ## 更新说明 **如果本插件更新后出现依赖包错误，请重新安装相关依赖包。详情见[这里](https://github.com/chflame163/ComfyUI_LayerStyle/issues/5)。
+* 添加 [BiRefNetUltra](#BiRefNetUltra) 节点, 使用BiRefNet模型去除背景，有更好的识别能力，同时具有超高的边缘细节。 * 添加 [ImageAutoCropV2](#ImageAutoCropV2) 节点，可选择不去除背景，支持mask输入，支持按长边或短边尺寸缩放。 * 添加 [ImageHub](#ImageHub) 节点，支持最多9组Image和Mask切换，支持随机输出。 * 添加 [TextJoin](#TextJoin) 节点。 @@ -625,7 +626,7 @@ ImageScaleByAspectRatio的V2升级版 ### ColorPicker 在色板上选取颜色并输出。 -改自[mtb nodes](https://github.com/melMass/comfy_mtb)的web extensions。 +改自[mtb nodes](https://github.com/melMass/comfy_mtb)的web extensions，感谢原作者。 ![image](image/color_picker.png) 节点选项说明: @@ -751,7 +752,7 @@ GradientImage的V2升级版。 ### SimpleTextImage ![image](image/simple_text_image_example.png) -从文字生成简单排版的图片以及遮罩。这个节点参考了[ZHO-ZHO-ZHO/ComfyUI-Text_Image-Composite](https://github.com/ZHO-ZHO-ZHO/ComfyUI-Text_Image-Composite)的部分功能和代码。 +从文字生成简单排版的图片以及遮罩。这个节点参考了[ZHO-ZHO-ZHO/ComfyUI-Text_Image-Composite](https://github.com/ZHO-ZHO-ZHO/ComfyUI-Text_Image-Composite)的部分功能和代码，感谢原作者。 ![image](image/simple_text_image_node.png) 节点选项说明: @@ -805,7 +806,7 @@ GradientImage的V2升级版。 ### LaMa ![image](image/lama_example.png) -根据图像遮罩擦除物体。本节点是对[IOPaint](https://www.iopaint.com)的封装，由 SOTA AI 模型提供支持。 +根据图像遮罩擦除物体。本节点是对[IOPaint](https://www.iopaint.com)的封装，由 SOTA AI 模型提供支持， 感谢原作者。 提供[LaMa](https://github.com/advimman/lama), [LDM](https://github.com/CompVis/latent-diffusion), [ZITS](https://github.com/DQiaole/ZITS_inpainting),[MAT](https://github.com/fenglinglwb/MAT), [FcF](https://github.com/SHI-Labs/FcF-Inpainting), [Manga](https://github.com/msxie92/MangaInpainting) 模型以及 SPREAD 擦除方法。请查看链接了解各个模型的介绍。 请下载模型文件 [lama models(百度网盘)](https://pan.baidu.com/s/1LllR9TJHP1G9uEwWT3Mvkg?pwd=tvzv) 或者 [lama models(Google Drive)](https://drive.google.com/drive/folders/1Aq0a4sybb3SRxi7j1e1_ZbBRjaWDdP9e?usp=sharing), 将文件放到```ComfyUI/models/lama``` @@ -943,7 +944,7 @@ cropped_mask: 裁切后的遮罩。 * y: 左上角位置y坐标输出。 ### SegmentAnythingUltra -对[ComfyUI Segment Anything](https://github.com/storyicon/comfyui_segment_anything)的改进，结合了spacepxl的[ComfyUI-Image-Filters](https://github.com/spacepxl/ComfyUI-Image-Filters)的Alpha Matte节点，使遮罩有更具细节的边缘。 +对[ComfyUI Segment Anything](https://github.com/storyicon/comfyui_segment_anything)的改进，结合了spacepxl的[ComfyUI-Image-Filters](https://github.com/spacepxl/ComfyUI-Image-Filters)的Alpha Matte节点，使遮罩有更具细节的边缘，感谢原作者。 *请参照ComfyUI Segment Anything的安装方法安装模型。如果已经正确安装了ComfyUI Segment Anything，可跳过此步骤。 * 从 [这里](https://huggingface.co/bert-base-uncased/tree/main) 下载 config.json，model.safetensors，tokenizer_config.json，tokenizer.json 和 vocab.txt 5个文件到 ```ComfyUI/models/bert-base-uncased```文件夹。 @@ -980,14 +981,13 @@ SegmentAnythingUltra的V2升级版，增加了VITMatte边缘处理方法。(注 * detail_dilate: 遮罩边缘向外扩张范围。数值越大，向外修复的范围越大。 + ### RemBgUltra 去除背景。与类似的背景移除节点相比，这个节点具有超高的边缘细节。 -本节点结合了spacepxl的[ComfyUI-Image-Filters](https://github.com/spacepxl/ComfyUI-Image-Filters)的Alpha Matte节点，以及ZHO-ZHO-ZHO的[ComfyUI-BRIA_AI-RMBG](https://github.com/ZHO-ZHO-ZHO/ComfyUI-BRIA_AI-RMBG)的功能。 +本节点结合了spacepxl的[ComfyUI-Image-Filters](https://github.com/spacepxl/ComfyUI-Image-Filters)的Alpha Matte节点，以及ZHO-ZHO-ZHO的[ComfyUI-BRIA_AI-RMBG](https://github.com/ZHO-ZHO-ZHO/ComfyUI-BRIA_AI-RMBG)的功能，感谢原作者。 *将[BRIA Background Removal v1.4](https://huggingface.co/briaai/RMBG-1.4)模型文件(model.pth)下载至```ComfyUI/models/rmbg/RMBG-1.4```文件夹。该模型由 BRIA AI 开发，可作为非商业用途的开源模型。 -**下载模型后，建议在ComfyUI/extra_model_paths.yaml中添加一行'rmbg: models/rmbg/RMBG-1.4'。(此步骤非必须） - ![image](image/rembg_ultra_example.png) @@ -1007,10 +1007,27 @@ RemBgUltra的V2升级版，增加了VITMatte边缘处理方法。(注意:超过2 * detail_erode: 遮罩边缘向内侵蚀范围。数值越大，向内修复的范围越大。 * detail_dilate: 遮罩边缘向外扩张范围。数值越大，向外修复的范围越大。 +### BiRefNetUltra +使用BiRefNet模型去除背景，有更好的识别能力，同时具有超高的边缘细节。 +本节点模型部分的代码来自vipery的[ComfyUI-BiRefNet](https://github.com/viperyl/ComfyUI-BiRefNet)，感谢原作者。 + +*从[https://huggingface.co/ViperYX/BiRefNet](https://huggingface.co/ViperYX/BiRefNet/tree/main)下载```BiRefNet-ep480.pth```,```pvt_v2_b2.pth```,```pvt_v2_b5.pth```,```swin_base_patch4_window12_384_22kto1k.pth```, ```swin_large_patch4_window12_384_22kto1k.pth```5个文件至```ComfyUI/models/BiRefNet```文件夹。 + +![image](image/birefnet_ultra_example.png) + +节点选项说明: +![image](image/birefnet_ultra_node.png) +* detail_method: 边缘处理方法。提供了VITMatte, PyMatting, GuidedFilter三种方法。 +* detail_erode: 遮罩边缘向内侵蚀范围。数值越大，向内修复的范围越大。 +* detail_dilate: 遮罩边缘向外扩张范围。数值越大，向外修复的范围越大。 +* black_point: 边缘黑色采样阈值。 +* white_point: 边缘黑色采样阈值。 +* process_detail: 此处设为False将跳过边缘处理以节省运行时间。 + ### PersonMaskUltra 为人物生成脸、头发、身体皮肤、衣服或配饰的遮罩。与之前的A Person Mask Generator节点相比，这个节点具有超高的边缘细节。 -本节点的模型代码来自[a-person-mask-generator](https://github.com/djbielejeski/a-person-mask-generator)，边缘处理代码来自spacepxl的[ComfyUI-Image-Filters](https://github.com/spacepxl/ComfyUI-Image-Filters)。 +本节点的模型代码来自[a-person-mask-generator](https://github.com/djbielejeski/a-person-mask-generator)，边缘处理代码来自spacepxl的[ComfyUI-Image-Filters](https://github.com/spacepxl/ComfyUI-Image-Filters)，感谢原作者。 ![image](image/person_mask_ultra_example.png) @@ -1076,7 +1093,7 @@ PersonMaskUltra的V2升级版，增加了VITMatte边缘处理方法。(注意: ### MaskEdgeUltraDetail 处理较粗糙的遮罩使其获得超精细边缘。 -本节点结合了spacepxl的[ComfyUI-Image-Filters](https://github.com/spacepxl/ComfyUI-Image-Filters)的Alpha Matte和Guided Filter Alpha两者功能。 +本节点结合了spacepxl的[ComfyUI-Image-Filters](https://github.com/spacepxl/ComfyUI-Image-Filters)的Alpha Matte和Guided Filter Alpha两者功能，感谢原作者。 ![image](image/mask_edge_ultra_detail_example.png) 节点选项说明: @@ -1242,7 +1259,7 @@ mask反转 ### Film 模拟胶片的颗粒、暗边和边缘模糊，支持输入深度图模拟虚焦。 -这个节点是[digitaljohn/comfyui-propost](https://github.com/digitaljohn/comfyui-propost)的重新封装。 +这个节点是[digitaljohn/comfyui-propost](https://github.com/digitaljohn/comfyui-propost)的重新封装，感谢原作者。 ![image](image/film_example.png) 节点选项说明: @@ -1314,4 +1331,4 @@ mask反转 ## 声明 -LayerStyle节点遵照MIT开源协议，有部分功能代码和模型来自其他开源项目。如果作为商业用途，请查阅原项目授权协议使用。 +LayerStyle节点遵照MIT开源协议，有部分功能代码和模型来自其他开源项目，感谢原作者。如果作为商业用途，请查阅原项目授权协议使用。 diff --git a/image/birefnet_ultra_example.png b/image/birefnet_ultra_example.png new file mode 100644 index 0000000..e56b3c1 Binary files /dev/null and b/image/birefnet_ultra_example.png differ diff --git a/image/birefnet_ultra_node.png b/image/birefnet_ultra_node.png new file mode 100644 index 0000000..5d52281 Binary files /dev/null and b/image/birefnet_ultra_node.png differ diff --git a/py/BiRefNet/__init__.py b/py/BiRefNet/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/py/BiRefNet/backbones/__init__.py b/py/BiRefNet/backbones/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/py/BiRefNet/backbones/build_backbone.py b/py/BiRefNet/backbones/build_backbone.py new file mode 100644 index 0000000..0ca4535 --- /dev/null +++ b/py/BiRefNet/backbones/build_backbone.py @@ -0,0 +1,44 @@ +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 BiRefNet.backbones.pvt_v2 import pvt_v2_b2, pvt_v2_b5 +from BiRefNet.backbones.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]) + 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 diff --git a/py/BiRefNet/backbones/pvt_v2.py b/py/BiRefNet/backbones/pvt_v2.py new file mode 100644 index 0000000..06b9704 --- /dev/null +++ b/py/BiRefNet/backbones/pvt_v2.py @@ -0,0 +1,434 @@ +import torch +import torch.nn as nn +from functools import partial + +from timm.models.layers import DropPath, to_2tuple, trunc_normal_ +from timm.models.registry import register_model + +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 = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # 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 = [x.item() for x in torch.linspace(0, drop_path_rate, sum(self.depths))] + 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 + + +## @register_model +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) + + + +## @register_model +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) + +## @register_model +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) + +## @register_model +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) + +## @register_model +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) + + +## @register_model +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) diff --git a/py/BiRefNet/backbones/swin_v1.py b/py/BiRefNet/backbones/swin_v1.py new file mode 100644 index 0000000..57501dc --- /dev/null +++ b/py/BiRefNet/backbones/swin_v1.py @@ -0,0 +1,652 @@ +# -------------------------------------------------------- +# 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.models.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 + Hp = int(np.ceil(H / self.window_size)) * self.window_size + Wp = int(np.ceil(W / self.window_size)) * 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)) + + 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 = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # 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 init_weights(self, pretrained=None): + """Initialize the weights in backbone. + + Args: + pretrained (str, optional): Path to pre-trained weights. + Defaults to None. + """ + + def _init_weights(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) + + if isinstance(pretrained, str): + self.apply(_init_weights) + logger = get_root_logger() + load_checkpoint(self, pretrained, strict=False, logger=logger) + elif pretrained is None: + self.apply(_init_weights) + else: + raise TypeError('pretrained must be a str or None') + + 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 diff --git a/py/BiRefNet/baseline.py b/py/BiRefNet/baseline.py new file mode 100644 index 0000000..8b762c5 --- /dev/null +++ b/py/BiRefNet/baseline.py @@ -0,0 +1,292 @@ +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 kornia.filters import laplacian + +from BiRefNet.backbones.build_backbone import build_backbone +from BiRefNet.modules.decoder_blocks import BasicDecBlk, ResBlk, HierarAttDecBlk +from BiRefNet.modules.lateral_blocks import BasicLatBlk +from BiRefNet.modules.aspp import ASPP, ASPPDeformable +from BiRefNet.modules.ing import * +from BiRefNet.refinement.refiner import Refiner, RefinerPVTInChannels4, RefUNet +from BiRefNet.refinement.stem_layer import StemLayer + +from .config import Config +from .dataset import class_labels_TR_sorted + + +class BiRefNet(nn.Module): + def __init__(self): + super(BiRefNet, self).__init__() + self.config = Config() + self.epoch = 1 + self.bb = build_backbone(self.config.bb, pretrained=True) + + 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.locate_head: + self.locate_header = nn.ModuleList([ + BasicDecBlk(channels[0], channels[-1]), + nn.Sequential( + nn.Conv2d(channels[-1], 1, 1, 1, 0), + ) + ]) + + 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) + 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 + + 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.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_ref(self, x, pred): + # refine patch-level segmentation + if pred.shape[2:] != x.shape[2:]: + pred = F.interpolate(pred, size=x.shape[2:], mode='bilinear', align_corners=True) + # pred = pred.sigmoid() + if self.config.refine == 'itself': + x = self.stem_layer(torch.cat([x, pred], dim=1)) + scaled_preds, class_preds = self.forward_ori(x) + else: + scaled_preds = self.refiner([x, pred]) + class_preds = None + return scaled_preds, class_preds + + def forward_ref_end(self, x): + # remove the grids of concatenated preds + return self.dec_end(x) if self.config.ender else x + + + 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_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], 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), nn.ReLU(inplace=True)) + self.gdt_convs_3 = nn.Sequential(nn.Conv2d(channels[2], _N, 3, 1, 1), nn.BatchNorm2d(_N), nn.ReLU(inplace=True)) + self.gdt_convs_2 = nn.Sequential(nn.Conv2d(channels[3], _N, 3, 1, 1), nn.BatchNorm2d(_N), 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 get_patches_batch(self, x, p): + _size_h, _size_w = p.shape[2:] + patches_batch = [] + for idx in range(x.shape[0]): + columns_x = torch.split(x[idx], split_size_or_sections=_size_w, dim=-1) + patches_x = [] + for column_x in columns_x: + patches_x += [p.unsqueeze(0) for p in torch.split(column_x, split_size_or_sections=_size_h, dim=-2)] + patch_sample = torch.cat(patches_x, dim=1) + patches_batch.append(patch_sample) + return torch.cat(patches_batch, dim=0) + + def forward(self, features): + if 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 = [] + p4 = self.decoder_block4(x4) + m4 = self.conv_ms_spvn_4(p4) if self.config.ms_supervision else None + _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 = self.get_patches_batch(x, _p3) 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 else None + if self.config.out_ref: + # >> 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 + p3_gdt = self.gdt_convs_3(p3) + 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 = self.get_patches_batch(x, _p2) 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 else None + if self.config.out_ref: + # >> 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: + p2_gdt = self.gdt_convs_2(p2) + 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 = self.get_patches_batch(x, _p1) 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 = self.get_patches_batch(x, _p1) 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: + 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)) diff --git a/py/BiRefNet/config.py b/py/BiRefNet/config.py new file mode 100644 index 0000000..e7ee155 --- /dev/null +++ b/py/BiRefNet/config.py @@ -0,0 +1,104 @@ +import os +import math +from folder_paths import models_dir + + +class Config(): + def __init__(self) -> None: + self.ms_supervision = True + self.out_ref = self.ms_supervision and True + self.dec_ipt = True + self.dec_ipt_split = True + self.locate_head = False + self.cxt_num = [0, 3][1] # multi-scale skip connections from encoder + self.mul_scl_ipt = ['', 'add', 'cat'][2] + 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.dec_att = ['', 'ASPP', 'ASPPDeformable'][2] + self.squeeze_block = ['', 'BasicDecBlk_x1', 'ResBlk_x4', 'ASPP_x3', 'ASPPDeformable_x3'][1] + self.dec_blk = ['BasicDecBlk', 'ResBlk', 'HierarAttDecBlk'][0] + self.auxiliary_classification = False + self.refine_iteration = 1 + self.freeze_bb = False + self.precisionHigh = True + self.compile = True + self.load_all = True + self.verbose_eval = True + + self.size = 1024 + self.batch_size = 2 + self.IoU_finetune_last_epochs = [0, -40][1] # choose 0 to skip + if self.dec_blk == 'HierarAttDecBlk': + self.batch_size = 2 ** [0, 1, 2, 3, 4][2] + self.model = [ + 'BiRefNet', + ][0] + + # Components + self.lat_blk = ['BasicLatBlk'][0] + self.dec_channels_inter = ['fixed', 'adap'][0] + + # Backbone + self.bb = [ + 'vgg16', 'vgg16bn', 'resnet50', # 0, 1, 2 + 'pvt_v2_b2', 'pvt_v2_b5', # 3-bs10, 4-bs5 + 'swin_v1_b', 'swin_v1_l' # 5-bs9, 6-bs6 + ][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], + }[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 [] + self.sys_home_dir = models_dir + self.weights_root_dir = os.path.join(self.sys_home_dir, "BiRefNet") + 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]), + } + + # Training + self.num_workers = 5 # will be decrease to min(it, batch_size) at the initialization of the data_loader + self.optimizer = ['Adam', 'AdamW'][0] + self.lr = 1e-5 * math.sqrt(self.batch_size / 5) # adapt the lr linearly + self.lr_decay_epochs = [1e4] # Set to negative N to decay the lr in the last N-th epoch. + self.lr_decay_rate = 0.5 + self.only_S_MAE = False + self.SDPA_enabled = False # Bug. Slower and errors occur in multi-GPUs + + # Data + self.data_root_dir = os.path.join(self.sys_home_dir, 'datasets/dis') + self.dataset = ['DIS5K', 'COD', 'SOD'][0] + self.preproc_methods = ['flip', 'enhance', 'rotate', 'pepper', 'crop'][:4] + + # 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 + } + 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) + + # others + self.device = [0, 'cpu'][0] # .to(0) = .to('cuda:0') + + self.batch_size_valid = 1 + self.rand_seed = 7 diff --git a/py/BiRefNet/dataset.py b/py/BiRefNet/dataset.py new file mode 100644 index 0000000..cd4280c --- /dev/null +++ b/py/BiRefNet/dataset.py @@ -0,0 +1,140 @@ +import os +import cv2 +from tqdm import tqdm +from PIL import Image +from torch.utils import data +from torchvision import transforms + +from .preproc import preproc +from .config import Config +from glob import glob + + +Image.MAX_IMAGE_PIXELS = None # remove DecompressionBombWarning +config = Config() +_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(', ') + + +class MyData(data.Dataset): + def __init__(self, data_root, image_size, is_train=True): + self.size_train = image_size + self.size_test = image_size + self.keep_size = not config.size + self.data_size = (config.size, config.size) + self.is_train = is_train + self.load_all = config.load_all + self.device = config.device + self.dataset = data_root.replace('\\', '/').split('/')[-1] + if self.is_train and config.auxiliary_classification: + self.cls_name2id = {_name: _id for _id, _name in enumerate(class_labels_TR_sorted)} + self.transform_image = transforms.Compose([ + transforms.Resize(self.data_size), + transforms.ToTensor(), + transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]), + ][self.load_all or self.keep_size:]) + self.transform_label = transforms.Compose([ + transforms.Resize(self.data_size), + transforms.ToTensor(), + ][self.load_all or self.keep_size:]) + ## 'im' and 'gt' need modifying + image_root = os.path.join(data_root, 'im') + self.image_paths = [os.path.join(image_root, p) for p in os.listdir(image_root)] + self.label_paths = [p.replace('/im/', '/gt/').replace('.jpg', '.png') for p in self.image_paths] + if self.load_all: + self.images_loaded, self.labels_loaded = [], [] + self.class_labels_loaded = [] + # for image_path, label_path in zip(self.image_paths, self.label_paths): + for image_path, label_path in tqdm(zip(self.image_paths, self.label_paths), total=len(self.image_paths)): + _image = cv2.imread(image_path) + _label = cv2.imread(label_path, cv2.IMREAD_GRAYSCALE) + if not self.keep_size: + _image_rs = cv2.resize(_image, (config.size, config.size), interpolation=cv2.INTER_LINEAR) + _label_rs = cv2.resize(_label, (config.size, config.size), interpolation=cv2.INTER_LINEAR) + self.images_loaded.append( + Image.fromarray(cv2.cvtColor(_image_rs, cv2.COLOR_BGR2RGB)).convert('RGB') + ) + self.labels_loaded.append( + Image.fromarray(_label_rs).convert('L') + ) + self.class_labels_loaded.append( + self.cls_name2id[label_path.split('/')[-1].split('#')[3]] if self.is_train and config.auxiliary_classification else -1 + ) + + + def __getitem__(self, index): + + if self.load_all: + image = self.images_loaded[index] + class_label = self.class_labels_loaded[index] if self.is_train and config.auxiliary_classification else -1 + else: + image = Image.open(self.image_paths[index]).convert('RGB') + + # loading image and label + if self.is_train: + image, label = preproc(image, image, preproc_methods=config.preproc_methods) + # else: + # if _label.shape[0] > 2048 or _label.shape[1] > 2048: + # _image = cv2.resize(_image, (2048, 2048), interpolation=cv2.INTER_LINEAR) + # _label = cv2.resize(_label, (2048, 2048), interpolation=cv2.INTER_LINEAR) + + image, label = self.transform_image(image), self.transform_label(label) + + if self.is_train: + return image, label, class_label + else: + return image, label, self.label_paths[index] + + def __len__(self): + return len(self.image_paths) + + +class YouData(data.Dataset): + def __init__(self, data_root, image_size, is_train=True): + self.size_train = image_size + self.size_test = image_size + self.keep_size = not config.size + self.data_size = (config.size, config.size) + self.is_train = is_train + self.load_all = config.load_all + self.device = config.device + self.dataset = data_root.replace('\\', '/').split('/')[-1] + if self.is_train and config.auxiliary_classification: + self.cls_name2id = {_name: _id for _id, _name in enumerate(class_labels_TR_sorted)} + self.transform_image = transforms.Compose([ + transforms.Resize(self.data_size), + transforms.ToTensor(), + transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]), + ][self.load_all or self.keep_size:]) + ## 'im' and 'gt' need modifying + self.image_paths = glob(data_root + "/*") + self.img_sizes = [] + if self.load_all: + self.images_loaded, self.labels_loaded = [], [] + for image_path in tqdm(self.image_paths, total=len(self.image_paths)): + _image = cv2.imread(image_path) + if not self.keep_size: + _image_rs = cv2.resize(_image, (config.size, config.size), interpolation=cv2.INTER_LINEAR) + self.images_loaded.append( + Image.fromarray(cv2.cvtColor(_image_rs, cv2.COLOR_BGR2RGB)).convert('RGB') + ) + self.img_sizes.append(_image.shape[:2]) + + + def __getitem__(self, index): + + if self.load_all: + image = self.images_loaded[index] + else: + image = Image.open(self.image_paths[index]).convert('RGB') + + # loading image and label + if self.is_train: + image, _ = preproc(image, image, preproc_methods=config.preproc_methods) + + image = self.transform_image(image) + size = self.img_sizes[index] + return image, size + + def __len__(self): + return len(self.image_paths) diff --git a/py/BiRefNet/modules/__init__.py b/py/BiRefNet/modules/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/py/BiRefNet/modules/aspp.py b/py/BiRefNet/modules/aspp.py new file mode 100644 index 0000000..997340a --- /dev/null +++ b/py/BiRefNet/modules/aspp.py @@ -0,0 +1,162 @@ +import torch +import torch.nn as nn +import torch.nn.functional as F +from BiRefNet.modules.deform_conv import DeformableConv2d +from ..config import Config + + +config = Config() + + +class ASPPComplex(nn.Module): + def __init__(self, in_channels=64, out_channels=None, output_stride=16): + super(ASPPComplex, 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) + 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) + + +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) + 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) + 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) + 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, num_parallel_block=1): + 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, 3, padding=1) for _ in range(num_parallel_block) + ]) + + 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) + 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) diff --git a/py/BiRefNet/modules/attentions.py b/py/BiRefNet/modules/attentions.py new file mode 100644 index 0000000..e1032af --- /dev/null +++ b/py/BiRefNet/modules/attentions.py @@ -0,0 +1,93 @@ +import numpy as np +import torch +from torch import nn +from torch.nn import init + + +class SEWeightModule(nn.Module): + def __init__(self, channels, reduction=16): + super(SEWeightModule, self).__init__() + self.avg_pool = nn.AdaptiveAvgPool2d(1) + self.fc1 = nn.Conv2d(channels, channels//reduction, kernel_size=1, padding=0) + self.relu = nn.ReLU(inplace=True) + self.fc2 = nn.Conv2d(channels//reduction, channels, kernel_size=1, padding=0) + self.sigmoid = nn.Sigmoid() + + def forward(self, x): + out = self.avg_pool(x) + out = self.fc1(out) + out = self.relu(out) + out = self.fc2(out) + weight = self.sigmoid(out) + return weight + + +class PSA(nn.Module): + + def __init__(self, in_channels, S=4, reduction=4): + super().__init__() + self.S = S + + _convs = [] + for i in range(S): + _convs.append(nn.Conv2d(in_channels//S, in_channels//S, kernel_size=2*(i+1)+1, padding=i+1)) + self.convs = nn.ModuleList(_convs) + + self.se_block = SEWeightModule(in_channels//S, reduction=S*reduction) + + self.softmax = nn.Softmax(dim=1) + + def forward(self, x): + b, c, h, w = x.size() + + # Step1: SPC module + SPC_out = x.view(b, self.S, c//self.S, h, w) #bs,s,ci,h,w + for idx, conv in enumerate(self.convs): + SPC_out[:,idx,:,:,:] = conv(SPC_out[:,idx,:,:,:].clone()) + + # Step2: SE weight + se_out=[] + for idx in range(self.S): + se_out.append(self.se_block(SPC_out[:, idx, :, :, :])) + SE_out = torch.stack(se_out, dim=1) + SE_out = SE_out.expand_as(SPC_out) + + # Step3: Softmax + softmax_out = self.softmax(SE_out) + + # Step4: SPA + PSA_out = SPC_out * softmax_out + PSA_out = PSA_out.view(b, -1, h, w) + + return PSA_out + + +class SGE(nn.Module): + + def __init__(self, groups): + super().__init__() + self.groups=groups + self.avg_pool = nn.AdaptiveAvgPool2d(1) + self.weight=nn.Parameter(torch.zeros(1,groups,1,1)) + self.bias=nn.Parameter(torch.zeros(1,groups,1,1)) + self.sig=nn.Sigmoid() + + def forward(self, x): + b, c, h,w=x.shape + x=x.view(b*self.groups,-1,h,w) #bs*g,dim//g,h,w + xn=x*self.avg_pool(x) #bs*g,dim//g,h,w + xn=xn.sum(dim=1,keepdim=True) #bs*g,1,h,w + t=xn.view(b*self.groups,-1) #bs*g,h*w + + t=t-t.mean(dim=1,keepdim=True) #bs*g,h*w + std=t.std(dim=1,keepdim=True)+1e-5 + t=t/std #bs*g,h*w + t=t.view(b,self.groups,h,w) #bs,g,h*w + + t=t*self.weight+self.bias #bs,g,h*w + t=t.view(b*self.groups,1,h,w) #bs*g,1,h*w + x=x*self.sig(t) + x=x.view(b,c,h,w) + + return x + diff --git a/py/BiRefNet/modules/decoder_blocks.py b/py/BiRefNet/modules/decoder_blocks.py new file mode 100644 index 0000000..1ece7c2 --- /dev/null +++ b/py/BiRefNet/modules/decoder_blocks.py @@ -0,0 +1,101 @@ +import torch +import torch.nn as nn +from BiRefNet.modules.aspp import ASPP, ASPPDeformable +from BiRefNet.modules.attentions import PSA, SGE +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) + self.bn_out = nn.BatchNorm2d(out_channels) + + 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) + 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) + + 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 + + +class HierarAttDecBlk(nn.Module): + def __init__(self, in_channels=64, out_channels=None, inter_channels=64): + super(HierarAttDecBlk, 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.split_y = 8 # must be divided by channels of all intermediate features + self.split_x = 8 + + self.conv_in = nn.Conv2d(in_channels, inter_channels, 3, 1, 1) + + self.psa = PSA(inter_channels*self.split_y*self.split_x, S=config.batch_size) + self.sge = SGE(groups=config.batch_size) + + 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, 1) + + def forward(self, x): + x = self.conv_in(x) + N, C, H, W = x.shape + x_patchs = x.reshape(N, -1, H//self.split_y, W//self.split_x) + + # Hierarchical attention: group attention X patch spatial attention + x_patchs = self.psa(x_patchs) # Group Channel Attention -- each group is a single image + x_patchs = self.sge(x_patchs) # Patch Spatial Attention + x = x.reshape(N, C, H, W) + if hasattr(self, 'dec_att'): + x = self.dec_att(x) + x = self.conv_out(x) + return x diff --git a/py/BiRefNet/modules/deform_conv.py b/py/BiRefNet/modules/deform_conv.py new file mode 100644 index 0000000..43f5e57 --- /dev/null +++ b/py/BiRefNet/modules/deform_conv.py @@ -0,0 +1,66 @@ +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 diff --git a/py/BiRefNet/modules/ing.py b/py/BiRefNet/modules/ing.py new file mode 100644 index 0000000..bdd3ef0 --- /dev/null +++ b/py/BiRefNet/modules/ing.py @@ -0,0 +1,29 @@ +import torch.nn as nn +from BiRefNet.modules.mlp import MLPLayer + + +class BlockA(nn.Module): + def __init__(self, in_channels=64, out_channels=64, inter_channels=64, mlp_ratio=4.): + super(BlockA, self).__init__() + inter_channels = in_channels + self.conv = nn.Conv2d(in_channels, inter_channels, 3, 1, 1) + self.norm1 = nn.LayerNorm(inter_channels) + self.ffn = MLPLayer(in_features=inter_channels, + hidden_features=int(inter_channels * mlp_ratio), + act_layer=nn.GELU, + drop=0.) + self.norm2 = nn.LayerNorm(inter_channels) + + def forward(self, x): + B, C, H, W = x.shape + _x = self.conv(x) + _x = _x.flatten(2).transpose(1, 2) + _x = self.norm1(_x) + x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous() + + x = x + _x + _x1 = self.ffn(x) + _x1 = self.norm2(_x1) + _x1 = _x1.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous() + x = x + _x1 + return x \ No newline at end of file diff --git a/py/BiRefNet/modules/lateral_blocks.py b/py/BiRefNet/modules/lateral_blocks.py new file mode 100644 index 0000000..a3022e6 --- /dev/null +++ b/py/BiRefNet/modules/lateral_blocks.py @@ -0,0 +1,21 @@ +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 diff --git a/py/BiRefNet/modules/mlp.py b/py/BiRefNet/modules/mlp.py new file mode 100644 index 0000000..39b3568 --- /dev/null +++ b/py/BiRefNet/modules/mlp.py @@ -0,0 +1,118 @@ +import torch +import torch.nn as nn +from functools import partial + +from timm.models.layers import DropPath, to_2tuple, trunc_normal_ +from timm.models.registry import register_model + +import math + + +class MLPLayer(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.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 + + +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 = 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) + + 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] + + 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 = MLPLayer(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) + + 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) + + 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 + diff --git a/py/BiRefNet/modules/utils.py b/py/BiRefNet/modules/utils.py new file mode 100644 index 0000000..59bd912 --- /dev/null +++ b/py/BiRefNet/modules/utils.py @@ -0,0 +1,54 @@ +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) diff --git a/py/BiRefNet/preproc.py b/py/BiRefNet/preproc.py new file mode 100644 index 0000000..a059c5d --- /dev/null +++ b/py/BiRefNet/preproc.py @@ -0,0 +1,85 @@ +from PIL import Image, ImageEnhance +import random +import numpy as np +import random + + +def preproc(image, label, preproc_methods=['flip']): + if 'flip' in preproc_methods: + image, label = cv_random_flip(image, label) + if 'crop' in preproc_methods: + image, label = random_crop(image, label) + if 'rotate' in preproc_methods: + image, label = random_rotate(image, label) + if 'enhance' in preproc_methods: + image = color_enhance(image) + if 'pepper' in preproc_methods: + label = random_pepper(label) + return image, label + + +def cv_random_flip(img, label): + if random.random() > 0.5: + img = img.transpose(Image.FLIP_LEFT_RIGHT) + label = label.transpose(Image.FLIP_LEFT_RIGHT) + return img, label + + +def random_crop(image, label): + border = 30 + image_width = image.size[0] + image_height = image.size[1] + border = int(min(image_width, image_height) * 0.1) + crop_win_width = np.random.randint(image_width - border, image_width) + crop_win_height = np.random.randint(image_height - border, image_height) + random_region = ( + (image_width - crop_win_width) >> 1, (image_height - crop_win_height) >> 1, (image_width + crop_win_width) >> 1, + (image_height + crop_win_height) >> 1) + return image.crop(random_region), label.crop(random_region) + + +def random_rotate(image, label, angle=15): + mode = Image.BICUBIC + if random.random() > 0.8: + random_angle = np.random.randint(-angle, angle) + image = image.rotate(random_angle, mode) + label = label.rotate(random_angle, mode) + return image, label + + +def color_enhance(image): + bright_intensity = random.randint(5, 15) / 10.0 + image = ImageEnhance.Brightness(image).enhance(bright_intensity) + contrast_intensity = random.randint(5, 15) / 10.0 + image = ImageEnhance.Contrast(image).enhance(contrast_intensity) + color_intensity = random.randint(0, 20) / 10.0 + image = ImageEnhance.Color(image).enhance(color_intensity) + sharp_intensity = random.randint(0, 30) / 10.0 + image = ImageEnhance.Sharpness(image).enhance(sharp_intensity) + return image + + +def random_gaussian(image, mean=0.1, sigma=0.35): + def gaussianNoisy(im, mean=mean, sigma=sigma): + for _i in range(len(im)): + im[_i] += random.gauss(mean, sigma) + return im + + img = np.asarray(image) + width, height = img.shape + img = gaussianNoisy(img[:].flatten(), mean, sigma) + img = img.reshape([width, height]) + return Image.fromarray(np.uint8(img)) + + +def random_pepper(img, N=0.0015): + img = np.array(img) + noiseNum = int(N * img.shape[0] * img.shape[1]) + for i in range(noiseNum): + randX = random.randint(0, img.shape[0] - 1) + randY = random.randint(0, img.shape[1] - 1) + if random.randint(0, 1) == 0: + img[randX, randY] = 0 + else: + img[randX, randY] = 255 + return Image.fromarray(img) diff --git a/py/BiRefNet/refinement/__init__.py b/py/BiRefNet/refinement/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/py/BiRefNet/refinement/refiner.py b/py/BiRefNet/refinement/refiner.py new file mode 100644 index 0000000..393d47b --- /dev/null +++ b/py/BiRefNet/refinement/refiner.py @@ -0,0 +1,253 @@ +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 BiRefNet.backbones.build_backbone import build_backbone +from BiRefNet.modules.decoder_blocks import BasicDecBlk +from BiRefNet.modules.lateral_blocks import BasicLatBlk +from BiRefNet.modules.ing import * +from BiRefNet.refinement.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) + 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 diff --git a/py/BiRefNet/refinement/stem_layer.py b/py/BiRefNet/refinement/stem_layer.py new file mode 100644 index 0000000..ecab538 --- /dev/null +++ b/py/BiRefNet/refinement/stem_layer.py @@ -0,0 +1,45 @@ +import torch.nn as nn +from BiRefNet.modules.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 diff --git a/py/birefnet_ultra.py b/py/birefnet_ultra.py new file mode 100644 index 0000000..e74d2a5 --- /dev/null +++ b/py/birefnet_ultra.py @@ -0,0 +1,148 @@ +import torch + +from .imagefunc import * + +import torch.nn as nn +from torchvision import transforms +from .BiRefNet.baseline import BiRefNet +# from .BiRefNet import config +from .BiRefNet.config import Config + +NODE_NAME = 'BiRefNetUltra' + + +config = Config() + +class BiRefNet_img_processor: + def __init__(self, config): + self.config = config + self.data_size = (config.size, config.size) + self.transform_image = transforms.Compose([ + transforms.Resize(self.data_size), + transforms.ToTensor(), + transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]), + ]) + + def __call__(self, _image: np.array): + _image_rs = cv2.resize(_image, (self.config.size, self.config.size), interpolation=cv2.INTER_LINEAR) + _image_rs = Image.fromarray(np.uint8(_image_rs*255)).convert('RGB') + image = self.transform_image(_image_rs) + return image + +class BiRefNetUltra: + def __init__(self): + self.ready = False + + def load(self, weight_path, device): + # load model + self.model = BiRefNet() + state_dict = torch.load(weight_path, map_location='cpu') + unwanted_prefix = '_orig_mod.' + for k, v in list(state_dict.items()): + if k.startswith(unwanted_prefix): + state_dict[k[len(unwanted_prefix):]] = state_dict.pop(k) + self.model.load_state_dict(state_dict) + self.model = self.model.to(device) + self.model.eval() + # load processor + self.processor = BiRefNet_img_processor(config) + self.ready = True + + + @classmethod + def INPUT_TYPES(cls): + + method_list = ['VITMatte', 'PyMatting', 'GuidedFilter'] + + + return { + "required": { + "image": ("IMAGE",), + "detail_method": (method_list,), + "detail_erode": ("INT", {"default": 6, "min": 1, "max": 255, "step": 1}), + "detail_dilate": ("INT", {"default": 6, "min": 1, "max": 255, "step": 1}), + "black_point": ("FLOAT", {"default": 0.01, "min": 0.01, "max": 0.98, "step": 0.01}), + "white_point": ("FLOAT", {"default": 0.99, "min": 0.02, "max": 0.99, "step": 0.01}), + "process_detail": ("BOOLEAN", {"default": True}), + }, + "optional": { + } + } + + RETURN_TYPES = ("IMAGE", "MASK", ) + RETURN_NAMES = ("image", "mask", ) + FUNCTION = "birefnet_ultra" + CATEGORY = '😺dzNodes/LayerMask' + + def birefnet_ultra(self, image, detail_method, detail_erode, detail_dilate, + black_point, white_point, process_detail): + ret_images = [] + ret_masks = [] + + if torch.backends.mps.is_available(): + device = "mps" + elif torch.cuda.is_available(): + device = "cuda" + else: + device = "cpu" + + if not self.ready: + model_folder_name = 'BiRefNet' + model_name = 'BiRefNet-ep480.pth' + model_file_path = "" + try: + model_file_path = os.path.join( + os.path.normpath(folder_paths.folder_names_and_paths[model_folder_name][0][0]), model_name) + except: + pass + if not os.path.exists(model_file_path): + model_file_path = os.path.join(folder_paths.models_dir, model_folder_name, model_name) + self.load(model_file_path, device=device) + + for i in image: + i = torch.unsqueeze(i, 0) + orig_image = tensor2pil(i).convert('RGB') + np_image = i.squeeze().numpy() + img = self.processor(np_image) + inputs = img[None, ...].to(device) + with torch.no_grad(): + scaled_preds = self.model(inputs)[-1].sigmoid() + _mask = nn.functional.interpolate( + scaled_preds[0].unsqueeze(0), + size=np_image.shape[:2], + mode='bilinear', + align_corners=True + )[0] + + detail_range = detail_erode + detail_dilate + if process_detail: + if detail_method == 'GuidedFilter': + _mask = guided_filter_alpha(i, _mask, detail_range // 6 + 1) + _mask = tensor2pil(histogram_remap(_mask, black_point, white_point)) + elif detail_method == 'PyMatting': + _mask = tensor2pil(mask_edge_detail(i, _mask, detail_range // 8 + 1, black_point, white_point)) + else: + brightness_image = ImageEnhance.Brightness(tensor2pil(_mask)) + _mask = brightness_image.enhance(factor=1.01) + _mask = pil2tensor(_mask) + _trimap = generate_VITMatte_trimap(_mask, detail_erode, detail_dilate) + _mask = generate_VITMatte(orig_image, _trimap) + _mask = tensor2pil(histogram_remap(pil2tensor(_mask), black_point, white_point)) + else: + _mask = tensor2pil(_mask) + + ret_image = RGB2RGBA(orig_image, _mask.convert('L')) + ret_images.append(pil2tensor(ret_image)) + ret_masks.append(image2mask(_mask)) + + log(f"{NODE_NAME} Processed {len(ret_masks)} image(s).", message_type='finish') + return (torch.cat(ret_images, dim=0), torch.cat(ret_masks, dim=0),) + # return (None, torch.cat(ret_masks, dim=0),) + +NODE_CLASS_MAPPINGS = { + "LayerMask: BiRefNetUltra": BiRefNetUltra, +} + +NODE_DISPLAY_NAME_MAPPINGS = { + "LayerMask: BiRefNetUltra": "LayerMask: BiRefNetUltra", +} diff --git a/py/imagefunc.py b/py/imagefunc.py index 6549602..08ce657 100644 --- a/py/imagefunc.py +++ b/py/imagefunc.py @@ -113,6 +113,13 @@ def tensor2np(tensor: torch.Tensor) -> List[np.ndarray]: def tensor2pil(t_image: torch.Tensor) -> Image: return Image.fromarray(np.clip(255.0 * t_image.cpu().numpy().squeeze(), 0, 255).astype(np.uint8)) +def tensor2cv2(image:torch.Tensor) -> np.array: + if image.dim() == 4: + image = image.squeeze() + npimage = image.numpy() + cv2image = np.uint8(npimage * 255 / npimage.max()) + return cv2.cvtColor(cv2image, cv2.COLOR_RGB2BGR) + def image2mask(image:Image) -> torch.Tensor: _image = image.convert('RGBA') alpha = _image.split() [0] diff --git a/requirements.txt b/requirements.txt index 7d1a73c..3e50f76 100644 --- a/requirements.txt +++ b/requirements.txt @@ -21,4 +21,5 @@ google-generativeai diffusers omegaconf tqdm -transformers \ No newline at end of file +transformers +kornia \ No newline at end of file