新增图片裁剪

.gitignore

+*.pth

app/api/routes.py

@@ -43,10 +43,12 @@ def search():
     image = request.args.get("image")
     top_k = request.args.get("top_k", type=int)
     bucket = request.args.get("bucket")
+    part = request.args.get("part")
+
 
     milvus = MilvusClient().connect()
 
-    result = ImageSearch(get_feature_extractor(), milvus).image_to_image_search(bucket, image, top_k)
+    result = ImageSearch(get_feature_extractor(), milvus).image_to_image_search(bucket, image, part, top_k)
 
     return jsonify({
         'code': 0,

app/models/schp/.gitignore

+**/__pycache__
+
+data/
+log/
+pretrain_model/

app/models/schp/LICENSE

+MIT License
+
+Copyright (c) 2020 Peike Li
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

app/models/schp/README.md

+# Self Correction for Human Parsing
+
+![Python 3.6](https://img.shields.io/badge/python-3.6-green.svg)
+[![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](https://opensource.org/licenses/MIT)
+
+An out-of-box human parsing representation extractor.
+
+Our solution ranks 1st for all human parsing tracks (including single, multiple and video) in the third LIP challenge!
+
+![lip-visualization](./demo/lip-visualization.jpg) 
+
+Features:
+- [x] Out-of-box human parsing extractor for other downstream applications.
+- [x] Pretrained model on three popular single person human parsing datasets.
+- [x] Training and inferecne code.
+- [x] Simple yet effective extension on multi-person and video human parsing tasks.
+
+## Requirements
+
+```
+conda env create -f environment.yaml
+conda activate schp
+pip install -r requirements.txt
+```
+
+## Simple Out-of-Box Extractor
+
+The easiest way to get started is to use our trained SCHP models on your own images to extract human parsing representations. Here we provided state-of-the-art [trained models](https://drive.google.com/drive/folders/1uOaQCpNtosIjEL2phQKEdiYd0Td18jNo?usp=sharing) on three popular datasets. Theses three datasets have different label system, you can choose the best one to fit on your own task.
+
+**LIP** ([exp-schp-201908261155-lip.pth](https://drive.google.com/file/d/1k4dllHpu0bdx38J7H28rVVLpU-kOHmnH/view?usp=sharing))
+
+* mIoU on LIP validation: **59.36 %**.
+
+* LIP is the largest single person human parsing dataset with 50000+ images. This dataset focus more on the complicated real scenarios. LIP has 20 labels, including 'Background', 'Hat', 'Hair', 'Glove', 'Sunglasses', 'Upper-clothes', 'Dress', 'Coat', 'Socks', 'Pants', 'Jumpsuits', 'Scarf', 'Skirt', 'Face', 'Left-arm', 'Right-arm', 'Left-leg', 'Right-leg', 'Left-shoe', 'Right-shoe'.
+
+**ATR** ([exp-schp-201908301523-atr.pth](https://drive.google.com/file/d/1ruJg4lqR_jgQPj-9K0PP-L2vJERYOxLP/view?usp=sharing))
+
+* mIoU on ATR test: **82.29%**.
+
+* ATR is a large single person human parsing dataset with 17000+ images. This dataset focus more on fashion AI. ATR has 18 labels, including 'Background', 'Hat', 'Hair', 'Sunglasses', 'Upper-clothes', 'Skirt', 'Pants', 'Dress', 'Belt', 'Left-shoe', 'Right-shoe', 'Face', 'Left-leg', 'Right-leg', 'Left-arm', 'Right-arm', 'Bag', 'Scarf'.
+
+**Pascal-Person-Part** ([exp-schp-201908270938-pascal-person-part.pth](https://drive.google.com/file/d/1E5YwNKW2VOEayK9mWCS3Kpsxf-3z04ZE/view?usp=sharing))
+
+* mIoU on Pascal-Person-Part validation: **71.46** %.
+
+* Pascal Person Part is a tiny single person human parsing dataset with 3000+ images. This dataset focus more on body parts segmentation. Pascal Person Part has 7 labels, including 'Background', 'Head', 'Torso', 'Upper Arms', 'Lower Arms', 'Upper Legs', 'Lower Legs'.
+
+Choose one and have fun on your own task!
+
+To extract the human parsing representation, simply put your own image in the `INPUT_PATH` folder, then download a pretrained model and run the following command. The output images with the same file name will be saved in `OUTPUT_PATH`
+
+```
+python simple_extractor.py --dataset [DATASET] --model-restore [CHECKPOINT_PATH] --input-dir [INPUT_PATH] --output-dir [OUTPUT_PATH]
+```
+
+**[Updated]** Here is also a [colab demo example](https://colab.research.google.com/drive/1JOwOPaChoc9GzyBi5FUEYTSaP2qxJl10?usp=sharing) for quick inference provided by [@levindabhi](https://github.com/levindabhi).
+
+The `DATASET` command has three options, including 'lip', 'atr' and 'pascal'. Note each pixel in the output images denotes the predicted label number. The output images have the same size as the input ones. To better visualization, we put a palette with the output images. We suggest you to read the image with `PIL`.
+
+If you need not only the final parsing images, but also the feature map representations. Add `--logits` command to save the output feature maps. These feature maps are the logits before softmax layer.
+
+## Dataset Preparation
+
+Please download the [LIP](http://sysu-hcp.net/lip/) dataset following the below structure.
+
+```commandline
+data/LIP
+|--- train_imgaes # 30462 training single person images
+|--- val_images # 10000 validation single person images
+|--- train_segmentations # 30462 training annotations
+|--- val_segmentations # 10000 training annotations
+|--- train_id.txt # training image list
+|--- val_id.txt # validation image list
+```
+
+## Training
+
+```
+python train.py 
+```
+By default, the trained model will be saved in `./log` directory. Please read the arguments for more details.
+
+## Evaluation
+```
+python evaluate.py --model-restore [CHECKPOINT_PATH]
+```
+CHECKPOINT_PATH should be the path of trained model.
+
+## Extension on Multiple Human Parsing
+
+Please read [MultipleHumanParsing.md](./mhp_extension/README.md) for more details.
+
+## Citation
+
+Please cite our work if you find this repo useful in your research.
+
+```latex
+@article{li2020self,
+  title={Self-Correction for Human Parsing}, 
+  author={Li, Peike and Xu, Yunqiu and Wei, Yunchao and Yang, Yi},
+  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
+  year={2020},
+  doi={10.1109/TPAMI.2020.3048039}}
+```
+
+## Visualization
+
+* Source Image.
+![demo](./demo/demo.jpg)
+* LIP Parsing Result.
+![demo-lip](./demo/demo_lip.png)
+* ATR Parsing Result.
+![demo-atr](./demo/demo_atr.png)
+* Pascal-Person-Part Parsing Result.
+![demo-pascal](./demo/demo_pascal.png)
+* Source Image.
+![demo](./mhp_extension/demo/demo.jpg)
+* Instance Human Mask.
+![demo-lip](./mhp_extension/demo/demo_instance_human_mask.png)
+* Global Human Parsing Result.
+![demo-lip](./mhp_extension/demo/demo_global_human_parsing.png)
+* Multiple Human Parsing Result.
+![demo-lip](./mhp_extension/demo/demo_multiple_human_parsing.png)
+
+
+## Related
+Our code adopts the [InplaceSyncBN](https://github.com/mapillary/inplace_abn) to save gpu memory cost.
+
+There is also a [PaddlePaddle](https://github.com/PaddlePaddle/PaddleSeg/tree/develop/contrib/ACE2P) Implementation of this project.

app/models/schp/datasets/datasets.py

+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+
+"""
+@Author  :   Peike Li
+@Contact :   peike.li@yahoo.com
+@File    :   datasets.py
+@Time    :   8/4/19 3:35 PM
+@Desc    :
+@License :   This source code is licensed under the license found in the
+             LICENSE file in the root directory of this source tree.
+"""
+
+import os
+import numpy as np
+import random
+import torch
+import cv2
+from torch.utils import data
+from utils.transforms import get_affine_transform
+
+
+class LIPDataSet(data.Dataset):
+    def __init__(self, root, dataset, crop_size=[473, 473], scale_factor=0.25,
+                 rotation_factor=30, ignore_label=255, transform=None):
+        self.root = root
+        self.aspect_ratio = crop_size[1] * 1.0 / crop_size[0]
+        self.crop_size = np.asarray(crop_size)
+        self.ignore_label = ignore_label
+        self.scale_factor = scale_factor
+        self.rotation_factor = rotation_factor
+        self.flip_prob = 0.5
+        self.transform = transform
+        self.dataset = dataset
+
+        list_path = os.path.join(self.root, self.dataset + '_id.txt')
+        train_list = [i_id.strip() for i_id in open(list_path)]
+
+        self.train_list = train_list
+        self.number_samples = len(self.train_list)
+
+    def __len__(self):
+        return self.number_samples
+
+    def _box2cs(self, box):
+        x, y, w, h = box[:4]
+        return self._xywh2cs(x, y, w, h)
+
+    def _xywh2cs(self, x, y, w, h):
+        center = np.zeros((2), dtype=np.float32)
+        center[0] = x + w * 0.5
+        center[1] = y + h * 0.5
+        if w > self.aspect_ratio * h:
+            h = w * 1.0 / self.aspect_ratio
+        elif w < self.aspect_ratio * h:
+            w = h * self.aspect_ratio
+        scale = np.array([w * 1.0, h * 1.0], dtype=np.float32)
+        return center, scale
+
+    def __getitem__(self, index):
+        train_item = self.train_list[index]
+
+        im_path = os.path.join(self.root, self.dataset + '_images', train_item + '.jpg')
+        parsing_anno_path = os.path.join(self.root, self.dataset + '_segmentations', train_item + '.png')
+
+        im = cv2.imread(im_path, cv2.IMREAD_COLOR)
+        h, w, _ = im.shape
+        parsing_anno = np.zeros((h, w), dtype=np.long)
+
+        # Get person center and scale
+        person_center, s = self._box2cs([0, 0, w - 1, h - 1])
+        r = 0
+
+        if self.dataset != 'test':
+            # Get pose annotation
+            parsing_anno = cv2.imread(parsing_anno_path, cv2.IMREAD_GRAYSCALE)
+            if self.dataset == 'train' or self.dataset == 'trainval':
+                sf = self.scale_factor
+                rf = self.rotation_factor
+                s = s * np.clip(np.random.randn() * sf + 1, 1 - sf, 1 + sf)
+                r = np.clip(np.random.randn() * rf, -rf * 2, rf * 2) if random.random() <= 0.6 else 0
+
+                if random.random() <= self.flip_prob:
+                    im = im[:, ::-1, :]
+                    parsing_anno = parsing_anno[:, ::-1]
+                    person_center[0] = im.shape[1] - person_center[0] - 1
+                    right_idx = [15, 17, 19]
+                    left_idx = [14, 16, 18]
+                    for i in range(0, 3):
+                        right_pos = np.where(parsing_anno == right_idx[i])
+                        left_pos = np.where(parsing_anno == left_idx[i])
+                        parsing_anno[right_pos[0], right_pos[1]] = left_idx[i]
+                        parsing_anno[left_pos[0], left_pos[1]] = right_idx[i]
+
+        trans = get_affine_transform(person_center, s, r, self.crop_size)
+        input = cv2.warpAffine(
+            im,
+            trans,
+            (int(self.crop_size[1]), int(self.crop_size[0])),
+            flags=cv2.INTER_LINEAR,
+            borderMode=cv2.BORDER_CONSTANT,
+            borderValue=(0, 0, 0))
+
+        if self.transform:
+            input = self.transform(input)
+
+        meta = {
+            'name': train_item,
+            'center': person_center,
+            'height': h,
+            'width': w,
+            'scale': s,
+            'rotation': r
+        }
+
+        if self.dataset == 'val' or self.dataset == 'test':
+            return input, meta
+        else:
+            label_parsing = cv2.warpAffine(
+                parsing_anno,
+                trans,
+                (int(self.crop_size[1]), int(self.crop_size[0])),
+                flags=cv2.INTER_NEAREST,
+                borderMode=cv2.BORDER_CONSTANT,
+                borderValue=(255))
+
+            label_parsing = torch.from_numpy(label_parsing)
+
+            return input, label_parsing, meta
+
+
+class LIPDataValSet(data.Dataset):
+    def __init__(self, root, dataset='val', crop_size=[473, 473], transform=None, flip=False):
+        self.root = root
+        self.crop_size = crop_size
+        self.transform = transform
+        self.flip = flip
+        self.dataset = dataset
+        self.root = root
+        self.aspect_ratio = crop_size[1] * 1.0 / crop_size[0]
+        self.crop_size = np.asarray(crop_size)
+
+        list_path = os.path.join(self.root, self.dataset + '_id.txt')
+        val_list = [i_id.strip() for i_id in open(list_path)]
+
+        self.val_list = val_list
+        self.number_samples = len(self.val_list)
+
+    def __len__(self):
+        return len(self.val_list)
+
+    def _box2cs(self, box):
+        x, y, w, h = box[:4]
+        return self._xywh2cs(x, y, w, h)
+
+    def _xywh2cs(self, x, y, w, h):
+        center = np.zeros((2), dtype=np.float32)
+        center[0] = x + w * 0.5
+        center[1] = y + h * 0.5
+        if w > self.aspect_ratio * h:
+            h = w * 1.0 / self.aspect_ratio
+        elif w < self.aspect_ratio * h:
+            w = h * self.aspect_ratio
+        scale = np.array([w * 1.0, h * 1.0], dtype=np.float32)
+
+        return center, scale
+
+    def __getitem__(self, index):
+        val_item = self.val_list[index]
+        # Load training image
+        im_path = os.path.join(self.root, self.dataset + '_images', val_item + '.jpg')
+        im = cv2.imread(im_path, cv2.IMREAD_COLOR)
+        h, w, _ = im.shape
+        # Get person center and scale
+        person_center, s = self._box2cs([0, 0, w - 1, h - 1])
+        r = 0
+        trans = get_affine_transform(person_center, s, r, self.crop_size)
+        input = cv2.warpAffine(
+            im,
+            trans,
+            (int(self.crop_size[1]), int(self.crop_size[0])),
+            flags=cv2.INTER_LINEAR,
+            borderMode=cv2.BORDER_CONSTANT,
+            borderValue=(0, 0, 0))
+        input = self.transform(input)
+        flip_input = input.flip(dims=[-1])
+        if self.flip:
+            batch_input_im = torch.stack([input, flip_input])
+        else:
+            batch_input_im = input
+
+        meta = {
+            'name': val_item,
+            'center': person_center,
+            'height': h,
+            'width': w,
+            'scale': s,
+            'rotation': r
+        }
+
+        return batch_input_im, meta

app/models/schp/datasets/simple_extractor_dataset.py

+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+
+"""
+@Author  :   Peike Li
+@Contact :   peike.li@yahoo.com
+@File    :   dataset.py
+@Time    :   8/30/19 9:12 PM
+@Desc    :   Dataset Definition
+@License :   This source code is licensed under the license found in the
+             LICENSE file in the root directory of this source tree.
+"""
+
+import os
+import cv2
+import numpy as np
+
+from torch.utils import data
+from utils.transforms import get_affine_transform
+
+
+class SimpleFolderDataset(data.Dataset):
+    def __init__(self, root, input_size=[512, 512], transform=None):
+        self.root = root
+        self.input_size = input_size
+        self.transform = transform
+        self.aspect_ratio = input_size[1] * 1.0 / input_size[0]
+        self.input_size = np.asarray(input_size)
+
+        self.file_list = os.listdir(self.root)
+
+    def __len__(self):
+        return len(self.file_list)
+
+    def _box2cs(self, box):
+        x, y, w, h = box[:4]
+        return self._xywh2cs(x, y, w, h)
+
+    def _xywh2cs(self, x, y, w, h):
+        center = np.zeros((2), dtype=np.float32)
+        center[0] = x + w * 0.5
+        center[1] = y + h * 0.5
+        if w > self.aspect_ratio * h:
+            h = w * 1.0 / self.aspect_ratio
+        elif w < self.aspect_ratio * h:
+            w = h * self.aspect_ratio
+        scale = np.array([w, h], dtype=np.float32)
+        return center, scale
+
+    def __getitem__(self, index):
+        img_name = self.file_list[index]
+        img_path = os.path.join(self.root, img_name)
+        img = cv2.imread(img_path, cv2.IMREAD_COLOR)
+        h, w, _ = img.shape
+
+        # Get person center and scale
+        person_center, s = self._box2cs([0, 0, w - 1, h - 1])
+        r = 0
+        trans = get_affine_transform(person_center, s, r, self.input_size)
+        input = cv2.warpAffine(
+            img,
+            trans,
+            (int(self.input_size[1]), int(self.input_size[0])),
+            flags=cv2.INTER_LINEAR,
+            borderMode=cv2.BORDER_CONSTANT,
+            borderValue=(0, 0, 0))
+
+        input = self.transform(input)
+        meta = {
+            'name': img_name,
+            'center': person_center,
+            'height': h,
+            'width': w,
+            'scale': s,
+            'rotation': r
+        }
+
+        return input, meta

app/models/schp/datasets/target_generation.py

+import torch
+from torch.nn import functional as F
+
+
+def generate_edge_tensor(label, edge_width=3):
+    label = label.type(torch.cuda.FloatTensor)
+    if len(label.shape) == 2:
+        label = label.unsqueeze(0)
+    n, h, w = label.shape
+    edge = torch.zeros(label.shape, dtype=torch.float).cuda()
+    # right
+    edge_right = edge[:, 1:h, :]
+    edge_right[(label[:, 1:h, :] != label[:, :h - 1, :]) & (label[:, 1:h, :] != 255)
+               & (label[:, :h - 1, :] != 255)] = 1
+
+    # up
+    edge_up = edge[:, :, :w - 1]
+    edge_up[(label[:, :, :w - 1] != label[:, :, 1:w])
+            & (label[:, :, :w - 1] != 255)
+            & (label[:, :, 1:w] != 255)] = 1
+
+    # upright
+    edge_upright = edge[:, :h - 1, :w - 1]
+    edge_upright[(label[:, :h - 1, :w - 1] != label[:, 1:h, 1:w])
+                 & (label[:, :h - 1, :w - 1] != 255)
+                 & (label[:, 1:h, 1:w] != 255)] = 1
+
+    # bottomright
+    edge_bottomright = edge[:, :h - 1, 1:w]
+    edge_bottomright[(label[:, :h - 1, 1:w] != label[:, 1:h, :w - 1])
+                     & (label[:, :h - 1, 1:w] != 255)
+                     & (label[:, 1:h, :w - 1] != 255)] = 1
+
+    kernel = torch.ones((1, 1, edge_width, edge_width), dtype=torch.float).cuda()
+    with torch.no_grad():
+        edge = edge.unsqueeze(1)
+        edge = F.conv2d(edge, kernel, stride=1, padding=1)
+    edge[edge!=0] = 1
+    edge = edge.squeeze()
+    return edge

app/models/schp/environment.yaml

+name: schp
+channels:
+  - pytorch
+  - defaults
+dependencies:
+  - _libgcc_mutex=0.1=main
+  - blas=1.0=mkl
+  - ca-certificates=2020.12.8=h06a4308_0
+  - certifi=2020.12.5=py38h06a4308_0
+  - cudatoolkit=10.1.243=h6bb024c_0
+  - freetype=2.10.4=h5ab3b9f_0
+  - intel-openmp=2020.2=254
+  - jpeg=9b=h024ee3a_2
+  - lcms2=2.11=h396b838_0
+  - ld_impl_linux-64=2.33.1=h53a641e_7
+  - libedit=3.1.20191231=h14c3975_1
+  - libffi=3.3=he6710b0_2
+  - libgcc-ng=9.1.0=hdf63c60_0
+  - libpng=1.6.37=hbc83047_0
+  - libstdcxx-ng=9.1.0=hdf63c60_0
+  - libtiff=4.1.0=h2733197_1
+  - lz4-c=1.9.2=heb0550a_3
+  - mkl=2020.2=256
+  - mkl-service=2.3.0=py38he904b0f_0
+  - mkl_fft=1.2.0=py38h23d657b_0
+  - mkl_random=1.1.1=py38h0573a6f_0
+  - ncurses=6.2=he6710b0_1
+  - ninja=1.10.2=py38hff7bd54_0
+  - numpy=1.19.2=py38h54aff64_0
+  - numpy-base=1.19.2=py38hfa32c7d_0
+  - olefile=0.46=py_0
+  - openssl=1.1.1i=h27cfd23_0
+  - pillow=8.0.1=py38he98fc37_0
+  - pip=20.3.3=py38h06a4308_0
+  - python=3.8.5=h7579374_1
+  - readline=8.0=h7b6447c_0
+  - setuptools=51.0.0=py38h06a4308_2
+  - six=1.15.0=py38h06a4308_0
+  - sqlite=3.33.0=h62c20be_0
+  - tk=8.6.10=hbc83047_0
+  - tqdm=4.55.0=pyhd3eb1b0_0
+  - wheel=0.36.2=pyhd3eb1b0_0
+  - xz=5.2.5=h7b6447c_0
+  - zlib=1.2.11=h7b6447c_3
+  - zstd=1.4.5=h9ceee32_0
+  - pytorch=1.5.1=py3.8_cuda10.1.243_cudnn7.6.3_0
+  - torchvision=0.6.1=py38_cu101
+prefix: /home/peike/opt/anaconda3/envs/schp
+

app/models/schp/evaluate.py

+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+
+"""
+@Author  :   Peike Li
+@Contact :   peike.li@yahoo.com
+@File    :   evaluate.py
+@Time    :   8/4/19 3:36 PM
+@Desc    :
+@License :   This source code is licensed under the license found in the
+             LICENSE file in the root directory of this source tree.
+"""
+
+import os
+import argparse
+import numpy as np
+import torch
+
+from torch.utils import data
+from tqdm import tqdm
+from PIL import Image as PILImage
+import torchvision.transforms as transforms
+import torch.backends.cudnn as cudnn
+
+import networks
+from datasets.datasets import LIPDataValSet
+from utils.miou import compute_mean_ioU
+from utils.transforms import BGR2RGB_transform
+from utils.transforms import transform_parsing
+
+
+def get_arguments():
+    """Parse all the arguments provided from the CLI.
+
+    Returns:
+      A list of parsed arguments.
+    """
+    parser = argparse.ArgumentParser(description="Self Correction for Human Parsing")
+
+    # Network Structure
+    parser.add_argument("--arch", type=str, default='resnet101')
+    # Data Preference
+    parser.add_argument("--data-dir", type=str, default='./data/LIP')
+    parser.add_argument("--batch-size", type=int, default=1)
+    parser.add_argument("--input-size", type=str, default='473,473')
+    parser.add_argument("--num-classes", type=int, default=20)
+    parser.add_argument("--ignore-label", type=int, default=255)
+    parser.add_argument("--random-mirror", action="store_true")
+    parser.add_argument("--random-scale", action="store_true")
+    # Evaluation Preference
+    parser.add_argument("--log-dir", type=str, default='./log')
+    parser.add_argument("--model-restore", type=str, default='./log/checkpoint.pth.tar')
+    parser.add_argument("--gpu", type=str, default='0', help="choose gpu device.")
+    parser.add_argument("--save-results", action="store_true", help="whether to save the results.")
+    parser.add_argument("--flip", action="store_true", help="random flip during the test.")
+    parser.add_argument("--multi-scales", type=str, default='1', help="multiple scales during the test")
+    return parser.parse_args()
+
+
+def get_palette(num_cls):
+    """ Returns the color map for visualizing the segmentation mask.
+    Args:
+        num_cls: Number of classes
+    Returns:
+        The color map
+    """
+    n = num_cls
+    palette = [0] * (n * 3)
+    for j in range(0, n):
+        lab = j
+        palette[j * 3 + 0] = 0
+        palette[j * 3 + 1] = 0
+        palette[j * 3 + 2] = 0
+        i = 0
+        while lab:
+            palette[j * 3 + 0] |= (((lab >> 0) & 1) << (7 - i))
+            palette[j * 3 + 1] |= (((lab >> 1) & 1) << (7 - i))
+            palette[j * 3 + 2] |= (((lab >> 2) & 1) << (7 - i))
+            i += 1
+            lab >>= 3
+    return palette
+
+
+def multi_scale_testing(model, batch_input_im, crop_size=[473, 473], flip=True, multi_scales=[1]):
+    flipped_idx = (15, 14, 17, 16, 19, 18)
+    if len(batch_input_im.shape) > 4:
+        batch_input_im = batch_input_im.squeeze()
+    if len(batch_input_im.shape) == 3:
+        batch_input_im = batch_input_im.unsqueeze(0)
+
+    interp = torch.nn.Upsample(size=crop_size, mode='bilinear', align_corners=True)
+    ms_outputs = []
+    for s in multi_scales:
+        interp_im = torch.nn.Upsample(scale_factor=s, mode='bilinear', align_corners=True)
+        scaled_im = interp_im(batch_input_im)
+        parsing_output = model(scaled_im)
+        parsing_output = parsing_output[0][-1]
+        output = parsing_output[0]
+        if flip:
+            flipped_output = parsing_output[1]
+            flipped_output[14:20, :, :] = flipped_output[flipped_idx, :, :]
+            output += flipped_output.flip(dims=[-1])
+            output *= 0.5
+        output = interp(output.unsqueeze(0))
+        ms_outputs.append(output[0])
+    ms_fused_parsing_output = torch.stack(ms_outputs)
+    ms_fused_parsing_output = ms_fused_parsing_output.mean(0)
+    ms_fused_parsing_output = ms_fused_parsing_output.permute(1, 2, 0)  # HWC
+    parsing = torch.argmax(ms_fused_parsing_output, dim=2)
+    parsing = parsing.data.cpu().numpy()
+    ms_fused_parsing_output = ms_fused_parsing_output.data.cpu().numpy()
+    return parsing, ms_fused_parsing_output
+
+
+def main():
+    """Create the model and start the evaluation process."""
+    args = get_arguments()
+    multi_scales = [float(i) for i in args.multi_scales.split(',')]
+    gpus = [int(i) for i in args.gpu.split(',')]
+    assert len(gpus) == 1
+    if not args.gpu == 'None':
+        os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu
+
+    cudnn.benchmark = True
+    cudnn.enabled = True
+
+    h, w = map(int, args.input_size.split(','))
+    input_size = [h, w]
+
+    model = networks.init_model(args.arch, num_classes=args.num_classes, pretrained=None)
+
+    IMAGE_MEAN = model.mean
+    IMAGE_STD = model.std
+    INPUT_SPACE = model.input_space
+    print('image mean: {}'.format(IMAGE_MEAN))
+    print('image std: {}'.format(IMAGE_STD))
+    print('input space:{}'.format(INPUT_SPACE))
+    if INPUT_SPACE == 'BGR':
+        print('BGR Transformation')
+        transform = transforms.Compose([
+            transforms.ToTensor(),
+            transforms.Normalize(mean=IMAGE_MEAN,
+                                 std=IMAGE_STD),
+
+        ])
+    if INPUT_SPACE == 'RGB':
+        print('RGB Transformation')
+        transform = transforms.Compose([
+            transforms.ToTensor(),
+            BGR2RGB_transform(),
+            transforms.Normalize(mean=IMAGE_MEAN,
+                                 std=IMAGE_STD),
+        ])
+
+    # Data loader
+    lip_test_dataset = LIPDataValSet(args.data_dir, 'val', crop_size=input_size, transform=transform, flip=args.flip)
+    num_samples = len(lip_test_dataset)
+    print('Totoal testing sample numbers: {}'.format(num_samples))
+    testloader = data.DataLoader(lip_test_dataset, batch_size=args.batch_size, shuffle=False, pin_memory=True)
+
+    # Load model weight
+    state_dict = torch.load(args.model_restore)['state_dict']
+    from collections import OrderedDict
+    new_state_dict = OrderedDict()
+    for k, v in state_dict.items():
+        name = k[7:]  # remove `module.`
+        new_state_dict[name] = v
+    model.load_state_dict(new_state_dict)
+    model.cuda()
+    model.eval()
+
+    sp_results_dir = os.path.join(args.log_dir, 'sp_results')
+    if not os.path.exists(sp_results_dir):
+        os.makedirs(sp_results_dir)
+
+    palette = get_palette(20)
+    parsing_preds = []
+    scales = np.zeros((num_samples, 2), dtype=np.float32)
+    centers = np.zeros((num_samples, 2), dtype=np.int32)
+    with torch.no_grad():
+        for idx, batch in enumerate(tqdm(testloader)):
+            image, meta = batch
+            if (len(image.shape) > 4):
+                image = image.squeeze()
+            im_name = meta['name'][0]
+            c = meta['center'].numpy()[0]
+            s = meta['scale'].numpy()[0]
+            w = meta['width'].numpy()[0]
+            h = meta['height'].numpy()[0]
+            scales[idx, :] = s
+            centers[idx, :] = c
+            parsing, logits = multi_scale_testing(model, image.cuda(), crop_size=input_size, flip=args.flip,
+                                                  multi_scales=multi_scales)
+            if args.save_results:
+                parsing_result = transform_parsing(parsing, c, s, w, h, input_size)
+                parsing_result_path = os.path.join(sp_results_dir, im_name + '.png')
+                output_im = PILImage.fromarray(np.asarray(parsing_result, dtype=np.uint8))
+                output_im.putpalette(palette)
+                output_im.save(parsing_result_path)
+
+            parsing_preds.append(parsing)
+    assert len(parsing_preds) == num_samples
+    mIoU = compute_mean_ioU(parsing_preds, scales, centers, args.num_classes, args.data_dir, input_size)
+    print(mIoU)
+    return
+
+
+if __name__ == '__main__':
+    main()

app/models/schp/mask.py

+import cv2
+import numpy as np
+from PIL import Image
+
+
+def extract_color_region_simple(original_image, mask_image, target_color, margin=10):
+    """
+    简化版本：提取指定颜色区域，白色背景
+
+    Args:
+        original_image: 原始图片 PIL Image 对象
+        mask_image: 色块图 PIL Image 对象
+        target_color: 目标颜色 RGB值，例如 [192, 0, 0] (红色)
+        margin: 裁剪边距，默认 10
+
+    Returns:
+        PIL.Image: 提取的颜色区域图片，失败返回 None
+    """
+
+    if original_image is None:
+        print("原始图片为空")
+        return None
+
+    if mask_image is None:
+        print("掩码图片为空")
+        return None
+
+    # 将PIL Image转换为numpy数组
+    original_rgb = np.array(original_image.convert('RGB'))
+    mask_rgb = np.array(mask_image.convert('RGB'))
+
+    # 创建目标颜色掩码 - 精确匹配
+    target_color = np.array(target_color)
+    color_mask = np.all(mask_rgb == target_color, axis=2).astype(np.uint8) * 255
+
+    # 形态学处理
+    kernel = np.ones((3, 3), np.uint8)
+    color_mask = cv2.morphologyEx(color_mask, cv2.MORPH_CLOSE, kernel)
+    color_mask = cv2.morphologyEx(color_mask, cv2.MORPH_OPEN, kernel)
+
+    # 调整掩码尺寸
+    if color_mask.shape != original_rgb.shape[:2]:
+        color_mask = cv2.resize(color_mask, (original_rgb.shape[1], original_rgb.shape[0]))
+
+    # 找边界框并裁剪
+    coords = np.column_stack(np.where(color_mask > 0))
+    if len(coords) == 0:
+        print(f"未找到目标颜色区域 RGB{target_color.tolist()}！")
+        return None
+
+    y_min, x_min = coords.min(axis=0)
+    y_max, x_max = coords.max(axis=0)
+
+    # 添加边距
+    y_min = max(0, y_min - margin)
+    x_min = max(0, x_min - margin)
+    y_max = min(original_rgb.shape[0], y_max + margin)
+    x_max = min(original_rgb.shape[1], x_max + margin)
+
+    # 裁剪
+    cropped_original = original_rgb[y_min:y_max, x_min:x_max]
+    cropped_mask = color_mask[y_min:y_max, x_min:x_max]
+
+    # 创建白色背景结果
+    result = np.full_like(cropped_original, 255, dtype=np.uint8)
+    result[cropped_mask > 0] = cropped_original[cropped_mask > 0]
+
+    # 统计信息
+    mask_pixels = np.sum(cropped_mask > 0)
+    total_pixels = cropped_mask.shape[0] * cropped_mask.shape[1]
+
+    print(f"提取完成")
+    print(f"目标颜色: RGB{target_color.tolist()}")
+    print(f"裁剪区域尺寸: {result.shape[1]} x {result.shape[0]}")
+    print(f"目标颜色像素数量: {mask_pixels}")
+    print(f"占裁剪区域的比例: {mask_pixels / total_pixels * 100:.2f}%")
+
+    # 转换为PIL Image并返回
+    pil_image = Image.fromarray(result)
+    return pil_image
+
+
+# 使用示例
+if __name__ == "__main__":
+    # 从文件加载PIL Image
+    original_image_path = r"D:\work\image_search\img1.jpg"
+    mask_image_path = r"D:\work\image_search\output.png"
+
+    original_image = Image.open(original_image_path)
+    mask_image = Image.open(mask_image_path)
+
+    # 示例: 提取红色区域
+    print("=== 提取红色区域 ===")
+    result_image_red = extract_color_region_simple(
+        original_image,
+        mask_image,
+        target_color=[192, 0, 0]
+    )
+
+    if result_image_red:
+        result_image_red.save(r"D:\work\image_search\extracted_red.png")
+        print("红色区域提取成功并已保存")
+    else:
+        print("红色区域提取失败")

app/models/schp/modules/__init__.py

+from .bn import ABN, InPlaceABN, InPlaceABNSync
+from .functions import ACT_RELU, ACT_LEAKY_RELU, ACT_ELU, ACT_NONE
+from .misc import GlobalAvgPool2d, SingleGPU
+from .residual import IdentityResidualBlock
+from .dense import DenseModule

app/models/schp/modules/bn.py

+import torch
+import torch.nn as nn
+import torch.nn.functional as functional
+
+try:
+    from queue import Queue
+except ImportError:
+    from Queue import Queue
+
+from .functions import *
+
+
+class ABN(nn.Module):
+    """Activated Batch Normalization
+
+    This gathers a `BatchNorm2d` and an activation function in a single module
+    """
+
+    def __init__(self, num_features, eps=1e-5, momentum=0.1, affine=True, activation="leaky_relu", slope=0.01):
+        """Creates an Activated Batch Normalization module
+
+        Parameters
+        ----------
+        num_features : int
+            Number of feature channels in the input and output.
+        eps : float
+            Small constant to prevent numerical issues.
+        momentum : float
+            Momentum factor applied to compute running statistics as.
+        affine : bool
+            If `True` apply learned scale and shift transformation after normalization.
+        activation : str
+            Name of the activation functions, one of: `leaky_relu`, `elu` or `none`.
+        slope : float
+            Negative slope for the `leaky_relu` activation.
+        """
+        super(ABN, self).__init__()
+        self.num_features = num_features
+        self.affine = affine
+        self.eps = eps
+        self.momentum = momentum
+        self.activation = activation
+        self.slope = slope
+        if self.affine:
+            self.weight = nn.Parameter(torch.ones(num_features))
+            self.bias = nn.Parameter(torch.zeros(num_features))
+        else:
+            self.register_parameter('weight', None)
+            self.register_parameter('bias', None)
+        self.register_buffer('running_mean', torch.zeros(num_features))
+        self.register_buffer('running_var', torch.ones(num_features))
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        nn.init.constant_(self.running_mean, 0)
+        nn.init.constant_(self.running_var, 1)
+        if self.affine:
+            nn.init.constant_(self.weight, 1)
+            nn.init.constant_(self.bias, 0)
+
+    def forward(self, x):
+        x = functional.batch_norm(x, self.running_mean, self.running_var, self.weight, self.bias,
+                                  self.training, self.momentum, self.eps)
+
+        if self.activation == ACT_RELU:
+            return functional.relu(x, inplace=True)
+        elif self.activation == ACT_LEAKY_RELU:
+            return functional.leaky_relu(x, negative_slope=self.slope, inplace=True)
+        elif self.activation == ACT_ELU:
+            return functional.elu(x, inplace=True)
+        else:
+            return x
+
+    def __repr__(self):
+        rep = '{name}({num_features}, eps={eps}, momentum={momentum},' \
+              ' affine={affine}, activation={activation}'
+        if self.activation == "leaky_relu":
+            rep += ', slope={slope})'
+        else:
+            rep += ')'
+        return rep.format(name=self.__class__.__name__, **self.__dict__)
+
+
+class InPlaceABN(ABN):
+    """InPlace Activated Batch Normalization"""
+
+    def __init__(self, num_features, eps=1e-5, momentum=0.1, affine=True, activation="leaky_relu", slope=0.01):
+        """Creates an InPlace Activated Batch Normalization module
+
+        Parameters
+        ----------
+        num_features : int
+            Number of feature channels in the input and output.
+        eps : float
+            Small constant to prevent numerical issues.
+        momentum : float
+            Momentum factor applied to compute running statistics as.
+        affine : bool
+            If `True` apply learned scale and shift transformation after normalization.
+        activation : str
+            Name of the activation functions, one of: `leaky_relu`, `elu` or `none`.
+        slope : float
+            Negative slope for the `leaky_relu` activation.
+        """
+        super(InPlaceABN, self).__init__(num_features, eps, momentum, affine, activation, slope)
+
+    def forward(self, x):
+        x, _, _ = inplace_abn(x, self.weight, self.bias, self.running_mean, self.running_var,
+                           self.training, self.momentum, self.eps, self.activation, self.slope)
+        return x
+
+
+class InPlaceABNSync(ABN):
+    """InPlace Activated Batch Normalization with cross-GPU synchronization
+    This assumes that it will be replicated across GPUs using the same mechanism as in `nn.DistributedDataParallel`.
+    """
+
+    def forward(self, x):
+        x, _, _ =  inplace_abn_sync(x, self.weight, self.bias, self.running_mean, self.running_var,
+                                   self.training, self.momentum, self.eps, self.activation, self.slope)
+        return x
+
+    def __repr__(self):
+        rep = '{name}({num_features}, eps={eps}, momentum={momentum},' \
+              ' affine={affine}, activation={activation}'
+        if self.activation == "leaky_relu":
+            rep += ', slope={slope})'
+        else:
+            rep += ')'
+        return rep.format(name=self.__class__.__name__, **self.__dict__)
+
+

app/models/schp/modules/deeplab.py

+import torch
+import torch.nn as nn
+import torch.nn.functional as functional
+
+from models._util import try_index
+from .bn import ABN
+
+
+class DeeplabV3(nn.Module):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 hidden_channels=256,
+                 dilations=(12, 24, 36),
+                 norm_act=ABN,
+                 pooling_size=None):
+        super(DeeplabV3, self).__init__()
+        self.pooling_size = pooling_size
+
+        self.map_convs = nn.ModuleList([
+            nn.Conv2d(in_channels, hidden_channels, 1, bias=False),
+            nn.Conv2d(in_channels, hidden_channels, 3, bias=False, dilation=dilations[0], padding=dilations[0]),
+            nn.Conv2d(in_channels, hidden_channels, 3, bias=False, dilation=dilations[1], padding=dilations[1]),
+            nn.Conv2d(in_channels, hidden_channels, 3, bias=False, dilation=dilations[2], padding=dilations[2])
+        ])
+        self.map_bn = norm_act(hidden_channels * 4)
+
+        self.global_pooling_conv = nn.Conv2d(in_channels, hidden_channels, 1, bias=False)
+        self.global_pooling_bn = norm_act(hidden_channels)
+
+        self.red_conv = nn.Conv2d(hidden_channels * 4, out_channels, 1, bias=False)
+        self.pool_red_conv = nn.Conv2d(hidden_channels, out_channels, 1, bias=False)
+        self.red_bn = norm_act(out_channels)
+
+        self.reset_parameters(self.map_bn.activation, self.map_bn.slope)
+
+    def reset_parameters(self, activation, slope):
+        gain = nn.init.calculate_gain(activation, slope)
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.xavier_normal_(m.weight.data, gain)
+                if hasattr(m, "bias") and m.bias is not None:
+                    nn.init.constant_(m.bias, 0)
+            elif isinstance(m, ABN):
+                if hasattr(m, "weight") and m.weight is not None:
+                    nn.init.constant_(m.weight, 1)
+                if hasattr(m, "bias") and m.bias is not None:
+                    nn.init.constant_(m.bias, 0)
+
+    def forward(self, x):
+        # Map convolutions
+        out = torch.cat([m(x) for m in self.map_convs], dim=1)
+        out = self.map_bn(out)
+        out = self.red_conv(out)
+
+        # Global pooling
+        pool = self._global_pooling(x)
+        pool = self.global_pooling_conv(pool)
+        pool = self.global_pooling_bn(pool)
+        pool = self.pool_red_conv(pool)
+        if self.training or self.pooling_size is None:
+            pool = pool.repeat(1, 1, x.size(2), x.size(3))
+
+        out += pool
+        out = self.red_bn(out)
+        return out
+
+    def _global_pooling(self, x):
+        if self.training or self.pooling_size is None:
+            pool = x.view(x.size(0), x.size(1), -1).mean(dim=-1)
+            pool = pool.view(x.size(0), x.size(1), 1, 1)
+        else:
+            pooling_size = (min(try_index(self.pooling_size, 0), x.shape[2]),
+                            min(try_index(self.pooling_size, 1), x.shape[3]))
+            padding = (
+                (pooling_size[1] - 1) // 2,
+                (pooling_size[1] - 1) // 2 if pooling_size[1] % 2 == 1 else (pooling_size[1] - 1) // 2 + 1,
+                (pooling_size[0] - 1) // 2,
+                (pooling_size[0] - 1) // 2 if pooling_size[0] % 2 == 1 else (pooling_size[0] - 1) // 2 + 1
+            )
+
+            pool = functional.avg_pool2d(x, pooling_size, stride=1)
+            pool = functional.pad(pool, pad=padding, mode="replicate")
+        return pool

app/models/schp/modules/dense.py

+from collections import OrderedDict
+
+import torch
+import torch.nn as nn
+
+from .bn import ABN
+
+
+class DenseModule(nn.Module):
+    def __init__(self, in_channels, growth, layers, bottleneck_factor=4, norm_act=ABN, dilation=1):
+        super(DenseModule, self).__init__()
+        self.in_channels = in_channels
+        self.growth = growth
+        self.layers = layers
+
+        self.convs1 = nn.ModuleList()
+        self.convs3 = nn.ModuleList()
+        for i in range(self.layers):
+            self.convs1.append(nn.Sequential(OrderedDict([
+                ("bn", norm_act(in_channels)),
+                ("conv", nn.Conv2d(in_channels, self.growth * bottleneck_factor, 1, bias=False))
+            ])))
+            self.convs3.append(nn.Sequential(OrderedDict([
+                ("bn", norm_act(self.growth * bottleneck_factor)),
+                ("conv", nn.Conv2d(self.growth * bottleneck_factor, self.growth, 3, padding=dilation, bias=False,
+                                   dilation=dilation))
+            ])))
+            in_channels += self.growth
+
+    @property
+    def out_channels(self):
+        return self.in_channels + self.growth * self.layers
+
+    def forward(self, x):
+        inputs = [x]
+        for i in range(self.layers):
+            x = torch.cat(inputs, dim=1)
+            x = self.convs1[i](x)
+            x = self.convs3[i](x)
+            inputs += [x]
+
+        return torch.cat(inputs, dim=1)

app/models/schp/modules/functions.py

+from os import path
+import torch
+import torch.distributed as dist
+import torch.autograd as autograd
+import torch.cuda.comm as comm
+from torch.autograd.function import once_differentiable
+from torch.utils.cpp_extension import load
+
+_src_path = path.join(path.dirname(path.abspath(__file__)), "src")
+# 检查CUDA是否可用
+if torch.cuda.is_available():
+    sources = [path.join(_src_path, f) for f in [
+        "inplace_abn.cpp",
+        "inplace_abn_cpu.cpp",
+        "inplace_abn_cuda.cu",
+        "inplace_abn_cuda_half.cu"
+    ]]
+    extra_cuda_cflags = ["--expt-extended-lambda"]
+else:
+    # CPU-only版本
+    sources = [path.join(_src_path, f) for f in [
+        "inplace_abn.cpp",
+        "inplace_abn_cpu.cpp"
+    ]]
+    extra_cuda_cflags = []
+
+_backend = load(name="inplace_abn",
+                extra_cflags=["-O3"],
+                sources=sources,
+                extra_cuda_cflags=extra_cuda_cflags)
+
+# Activation names
+ACT_RELU = "relu"
+ACT_LEAKY_RELU = "leaky_relu"
+ACT_ELU = "elu"
+ACT_NONE = "none"
+
+
+def _check(fn, *args, **kwargs):
+    success = fn(*args, **kwargs)
+    if not success:
+        raise RuntimeError("CUDA Error encountered in {}".format(fn))
+
+
+def _broadcast_shape(x):
+    out_size = []
+    for i, s in enumerate(x.size()):
+        if i != 1:
+            out_size.append(1)
+        else:
+            out_size.append(s)
+    return out_size
+
+
+def _reduce(x):
+    if len(x.size()) == 2:
+        return x.sum(dim=0)
+    else:
+        n, c = x.size()[0:2]
+        return x.contiguous().view((n, c, -1)).sum(2).sum(0)
+
+
+def _count_samples(x):
+    count = 1
+    for i, s in enumerate(x.size()):
+        if i != 1:
+            count *= s
+    return count
+
+
+def _act_forward(ctx, x):
+    if ctx.activation == ACT_LEAKY_RELU:
+        _backend.leaky_relu_forward(x, ctx.slope)
+    elif ctx.activation == ACT_ELU:
+        _backend.elu_forward(x)
+    elif ctx.activation == ACT_NONE:
+        pass
+
+
+def _act_backward(ctx, x, dx):
+    if ctx.activation == ACT_LEAKY_RELU:
+        _backend.leaky_relu_backward(x, dx, ctx.slope)
+    elif ctx.activation == ACT_ELU:
+        _backend.elu_backward(x, dx)
+    elif ctx.activation == ACT_NONE:
+        pass
+
+
+class InPlaceABN(autograd.Function):
+    @staticmethod
+    def forward(ctx, x, weight, bias, running_mean, running_var,
+                training=True, momentum=0.1, eps=1e-05, activation=ACT_LEAKY_RELU, slope=0.01):
+        # Save context
+        ctx.training = training
+        ctx.momentum = momentum
+        ctx.eps = eps
+        ctx.activation = activation
+        ctx.slope = slope
+        ctx.affine = weight is not None and bias is not None
+
+        # Prepare inputs
+        count = _count_samples(x)
+        x = x.contiguous()
+        weight = weight.contiguous() if ctx.affine else x.new_empty(0)
+        bias = bias.contiguous() if ctx.affine else x.new_empty(0)
+
+        if ctx.training:
+            mean, var = _backend.mean_var(x)
+
+            # Update running stats
+            running_mean.mul_((1 - ctx.momentum)).add_(ctx.momentum * mean)
+            running_var.mul_((1 - ctx.momentum)).add_(ctx.momentum * var * count / (count - 1))
+
+            # Mark in-place modified tensors
+            ctx.mark_dirty(x, running_mean, running_var)
+        else:
+            mean, var = running_mean.contiguous(), running_var.contiguous()
+            ctx.mark_dirty(x)
+
+        # BN forward + activation
+        _backend.forward(x, mean, var, weight, bias, ctx.affine, ctx.eps)
+        _act_forward(ctx, x)
+
+        # Output
+        ctx.var = var
+        ctx.save_for_backward(x, var, weight, bias)
+        ctx.mark_non_differentiable(running_mean, running_var)
+        return x, running_mean, running_var
+
+    @staticmethod
+    @once_differentiable
+    def backward(ctx, dz, _drunning_mean, _drunning_var):
+        z, var, weight, bias = ctx.saved_tensors
+        dz = dz.contiguous()
+
+        # Undo activation
+        _act_backward(ctx, z, dz)
+
+        if ctx.training:
+            edz, eydz = _backend.edz_eydz(z, dz, weight, bias, ctx.affine, ctx.eps)
+        else:
+            # TODO: implement simplified CUDA backward for inference mode
+            edz = dz.new_zeros(dz.size(1))
+            eydz = dz.new_zeros(dz.size(1))
+
+        dx = _backend.backward(z, dz, var, weight, bias, edz, eydz, ctx.affine, ctx.eps)
+        # dweight = eydz * weight.sign() if ctx.affine else None
+        dweight = eydz if ctx.affine else None
+        if dweight is not None:
+            dweight[weight < 0] *= -1
+        dbias = edz if ctx.affine else None
+
+        return dx, dweight, dbias, None, None, None, None, None, None, None
+
+
+class InPlaceABNSync(autograd.Function):
+    @classmethod
+    def forward(cls, ctx, x, weight, bias, running_mean, running_var,
+                training=True, momentum=0.1, eps=1e-05, activation=ACT_LEAKY_RELU, slope=0.01, equal_batches=True):
+        # Save context
+        ctx.training = training
+        ctx.momentum = momentum
+        ctx.eps = eps
+        ctx.activation = activation
+        ctx.slope = slope
+        ctx.affine = weight is not None and bias is not None
+
+        # Prepare inputs
+        ctx.world_size = dist.get_world_size() if dist.is_initialized() else 1
+
+        # count = _count_samples(x)
+        batch_size = x.new_tensor([x.shape[0]], dtype=torch.long)
+
+        x = x.contiguous()
+        weight = weight.contiguous() if ctx.affine else x.new_empty(0)
+        bias = bias.contiguous() if ctx.affine else x.new_empty(0)
+
+        if ctx.training:
+            mean, var = _backend.mean_var(x)
+            if ctx.world_size > 1:
+                # get global batch size
+                if equal_batches:
+                    batch_size *= ctx.world_size
+                else:
+                    dist.all_reduce(batch_size, dist.ReduceOp.SUM)
+
+                ctx.factor = x.shape[0] / float(batch_size.item())
+
+                mean_all = mean.clone() * ctx.factor
+                dist.all_reduce(mean_all, dist.ReduceOp.SUM)
+
+                var_all = (var + (mean - mean_all) ** 2) * ctx.factor
+                dist.all_reduce(var_all, dist.ReduceOp.SUM)
+
+                mean = mean_all
+                var = var_all
+
+            # Update running stats
+            running_mean.mul_((1 - ctx.momentum)).add_(ctx.momentum * mean)
+            count = batch_size.item() * x.view(x.shape[0], x.shape[1], -1).shape[-1]
+            running_var.mul_((1 - ctx.momentum)).add_(ctx.momentum * var * (float(count) / (count - 1)))
+
+            # Mark in-place modified tensors
+            ctx.mark_dirty(x, running_mean, running_var)
+        else:
+            mean, var = running_mean.contiguous(), running_var.contiguous()
+            ctx.mark_dirty(x)
+
+        # BN forward + activation
+        _backend.forward(x, mean, var, weight, bias, ctx.affine, ctx.eps)
+        _act_forward(ctx, x)
+
+        # Output
+        ctx.var = var
+        ctx.save_for_backward(x, var, weight, bias)
+        ctx.mark_non_differentiable(running_mean, running_var)
+        return x, running_mean, running_var
+
+    @staticmethod
+    @once_differentiable
+    def backward(ctx, dz, _drunning_mean, _drunning_var):
+        z, var, weight, bias = ctx.saved_tensors
+        dz = dz.contiguous()
+
+        # Undo activation
+        _act_backward(ctx, z, dz)
+
+        if ctx.training:
+            edz, eydz = _backend.edz_eydz(z, dz, weight, bias, ctx.affine, ctx.eps)
+            edz_local = edz.clone()
+            eydz_local = eydz.clone()
+
+            if ctx.world_size > 1:
+                edz *= ctx.factor
+                dist.all_reduce(edz, dist.ReduceOp.SUM)
+
+                eydz *= ctx.factor
+                dist.all_reduce(eydz, dist.ReduceOp.SUM)
+        else:
+            edz_local = edz = dz.new_zeros(dz.size(1))
+            eydz_local = eydz = dz.new_zeros(dz.size(1))
+
+        dx = _backend.backward(z, dz, var, weight, bias, edz, eydz, ctx.affine, ctx.eps)
+        # dweight = eydz_local * weight.sign() if ctx.affine else None
+        dweight = eydz_local if ctx.affine else None
+        if dweight is not None:
+            dweight[weight < 0] *= -1
+        dbias = edz_local if ctx.affine else None
+
+        return dx, dweight, dbias, None, None, None, None, None, None, None
+
+
+inplace_abn = InPlaceABN.apply
+inplace_abn_sync = InPlaceABNSync.apply
+
+__all__ = ["inplace_abn", "inplace_abn_sync", "ACT_RELU", "ACT_LEAKY_RELU", "ACT_ELU", "ACT_NONE"]

app/models/schp/modules/misc.py

+import torch.nn as nn
+import torch
+import torch.distributed as dist
+
+class GlobalAvgPool2d(nn.Module):
+    def __init__(self):
+        """Global average pooling over the input's spatial dimensions"""
+        super(GlobalAvgPool2d, self).__init__()
+
+    def forward(self, inputs):
+        in_size = inputs.size()
+        return inputs.view((in_size[0], in_size[1], -1)).mean(dim=2)
+
+class SingleGPU(nn.Module):
+    def __init__(self, module):
+        super(SingleGPU, self).__init__()
+        self.module=module
+
+    def forward(self, input):
+        return self.module(input.cuda(non_blocking=True))
+

app/models/schp/modules/residual.py

+from collections import OrderedDict
+
+import torch.nn as nn
+
+from .bn import ABN, ACT_LEAKY_RELU, ACT_ELU, ACT_NONE
+import torch.nn.functional as functional
+
+
+class ResidualBlock(nn.Module):
+    """Configurable residual block
+
+    Parameters
+    ----------
+    in_channels : int
+        Number of input channels.
+    channels : list of int
+        Number of channels in the internal feature maps. Can either have two or three elements: if three construct
+        a residual block with two `3 x 3` convolutions, otherwise construct a bottleneck block with `1 x 1`, then
+        `3 x 3` then `1 x 1` convolutions.
+    stride : int
+        Stride of the first `3 x 3` convolution
+    dilation : int
+        Dilation to apply to the `3 x 3` convolutions.
+    groups : int
+        Number of convolution groups. This is used to create ResNeXt-style blocks and is only compatible with
+        bottleneck blocks.
+    norm_act : callable
+        Function to create normalization / activation Module.
+    dropout: callable
+        Function to create Dropout Module.
+    """
+
+    def __init__(self,
+                 in_channels,
+                 channels,
+                 stride=1,
+                 dilation=1,
+                 groups=1,
+                 norm_act=ABN,
+                 dropout=None):
+        super(ResidualBlock, self).__init__()
+
+        # Check parameters for inconsistencies
+        if len(channels) != 2 and len(channels) != 3:
+            raise ValueError("channels must contain either two or three values")
+        if len(channels) == 2 and groups != 1:
+            raise ValueError("groups > 1 are only valid if len(channels) == 3")
+
+        is_bottleneck = len(channels) == 3
+        need_proj_conv = stride != 1 or in_channels != channels[-1]
+
+        if not is_bottleneck:
+            bn2 = norm_act(channels[1])
+            bn2.activation = ACT_NONE
+            layers = [
+                ("conv1", nn.Conv2d(in_channels, channels[0], 3, stride=stride, padding=dilation, bias=False,
+                                    dilation=dilation)),
+                ("bn1", norm_act(channels[0])),
+                ("conv2", nn.Conv2d(channels[0], channels[1], 3, stride=1, padding=dilation, bias=False,
+                                    dilation=dilation)),
+                ("bn2", bn2)
+            ]
+            if dropout is not None:
+                layers = layers[0:2] + [("dropout", dropout())] + layers[2:]
+        else:
+            bn3 = norm_act(channels[2])
+            bn3.activation = ACT_NONE
+            layers = [
+                ("conv1", nn.Conv2d(in_channels, channels[0], 1, stride=1, padding=0, bias=False)),
+                ("bn1", norm_act(channels[0])),
+                ("conv2", nn.Conv2d(channels[0], channels[1], 3, stride=stride, padding=dilation, bias=False,
+                                    groups=groups, dilation=dilation)),
+                ("bn2", norm_act(channels[1])),
+                ("conv3", nn.Conv2d(channels[1], channels[2], 1, stride=1, padding=0, bias=False)),
+                ("bn3", bn3)
+            ]
+            if dropout is not None:
+                layers = layers[0:4] + [("dropout", dropout())] + layers[4:]
+        self.convs = nn.Sequential(OrderedDict(layers))
+
+        if need_proj_conv:
+            self.proj_conv = nn.Conv2d(in_channels, channels[-1], 1, stride=stride, padding=0, bias=False)
+            self.proj_bn = norm_act(channels[-1])
+            self.proj_bn.activation = ACT_NONE
+
+    def forward(self, x):
+        if hasattr(self, "proj_conv"):
+            residual = self.proj_conv(x)
+            residual = self.proj_bn(residual)
+        else:
+            residual = x
+        x = self.convs(x) + residual
+
+        if self.convs.bn1.activation == ACT_LEAKY_RELU:
+            return functional.leaky_relu(x, negative_slope=self.convs.bn1.slope, inplace=True)
+        elif self.convs.bn1.activation == ACT_ELU:
+            return functional.elu(x, inplace=True)
+        else:
+            return x
+
+
+class IdentityResidualBlock(nn.Module):
+    def __init__(self,
+                 in_channels,
+                 channels,
+                 stride=1,
+                 dilation=1,
+                 groups=1,
+                 norm_act=ABN,
+                 dropout=None):
+        """Configurable identity-mapping residual block
+
+        Parameters
+        ----------
+        in_channels : int
+            Number of input channels.
+        channels : list of int
+            Number of channels in the internal feature maps. Can either have two or three elements: if three construct
+            a residual block with two `3 x 3` convolutions, otherwise construct a bottleneck block with `1 x 1`, then
+            `3 x 3` then `1 x 1` convolutions.
+        stride : int
+            Stride of the first `3 x 3` convolution
+        dilation : int
+            Dilation to apply to the `3 x 3` convolutions.
+        groups : int
+            Number of convolution groups. This is used to create ResNeXt-style blocks and is only compatible with
+            bottleneck blocks.
+        norm_act : callable
+            Function to create normalization / activation Module.
+        dropout: callable
+            Function to create Dropout Module.
+        """
+        super(IdentityResidualBlock, self).__init__()
+
+        # Check parameters for inconsistencies
+        if len(channels) != 2 and len(channels) != 3:
+            raise ValueError("channels must contain either two or three values")
+        if len(channels) == 2 and groups != 1:
+            raise ValueError("groups > 1 are only valid if len(channels) == 3")
+
+        is_bottleneck = len(channels) == 3
+        need_proj_conv = stride != 1 or in_channels != channels[-1]
+
+        self.bn1 = norm_act(in_channels)
+        if not is_bottleneck:
+            layers = [
+                ("conv1", nn.Conv2d(in_channels, channels[0], 3, stride=stride, padding=dilation, bias=False,
+                                    dilation=dilation)),
+                ("bn2", norm_act(channels[0])),
+                ("conv2", nn.Conv2d(channels[0], channels[1], 3, stride=1, padding=dilation, bias=False,
+                                    dilation=dilation))
+            ]
+            if dropout is not None:
+                layers = layers[0:2] + [("dropout", dropout())] + layers[2:]
+        else:
+            layers = [
+                ("conv1", nn.Conv2d(in_channels, channels[0], 1, stride=stride, padding=0, bias=False)),
+                ("bn2", norm_act(channels[0])),
+                ("conv2", nn.Conv2d(channels[0], channels[1], 3, stride=1, padding=dilation, bias=False,
+                                    groups=groups, dilation=dilation)),
+                ("bn3", norm_act(channels[1])),
+                ("conv3", nn.Conv2d(channels[1], channels[2], 1, stride=1, padding=0, bias=False))
+            ]
+            if dropout is not None:
+                layers = layers[0:4] + [("dropout", dropout())] + layers[4:]
+        self.convs = nn.Sequential(OrderedDict(layers))
+
+        if need_proj_conv:
+            self.proj_conv = nn.Conv2d(in_channels, channels[-1], 1, stride=stride, padding=0, bias=False)
+
+    def forward(self, x):
+        if hasattr(self, "proj_conv"):
+            bn1 = self.bn1(x)
+            shortcut = self.proj_conv(bn1)
+        else:
+            shortcut = x.clone()
+            bn1 = self.bn1(x)
+
+        out = self.convs(bn1)
+        out.add_(shortcut)
+
+        return out

app/models/schp/modules/src/checks.h

+#pragma once
+
+#include <ATen/ATen.h>
+
+// Define AT_CHECK for old version of ATen where the same function was called AT_ASSERT
+#ifndef AT_CHECK
+#define AT_CHECK AT_ASSERT
+#endif
+
+#define CHECK_CUDA(x) AT_CHECK((x).type().is_cuda(), #x " must be a CUDA tensor")
+#define CHECK_CPU(x) AT_CHECK(!(x).type().is_cuda(), #x " must be a CPU tensor")
+#define CHECK_CONTIGUOUS(x) AT_CHECK((x).is_contiguous(), #x " must be contiguous")
+
+#define CHECK_CUDA_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x)
+#define CHECK_CPU_INPUT(x) CHECK_CPU(x); CHECK_CONTIGUOUS(x)
\ No newline at end of file

app/models/schp/modules/src/inplace_abn.cpp

+#include <torch/extension.h>
+#include <vector>
+#include "inplace_abn.h"
+
+std::vector<at::Tensor> mean_var(at::Tensor x) {
+  if (x.is_cuda()) {
+#ifdef WITH_CUDA
+    if (x.scalar_type() == at::ScalarType::Half) {
+      return mean_var_cuda_h(x);
+    } else {
+      return mean_var_cuda(x);
+    }
+#else
+    AT_ERROR("Not compiled with CUDA support");
+#endif
+  } else {
+    return mean_var_cpu(x);
+  }
+}
+
+at::Tensor forward(at::Tensor x, at::Tensor mean, at::Tensor var, at::Tensor weight, at::Tensor bias,
+                   bool affine, float eps) {
+  if (x.is_cuda()) {
+#ifdef WITH_CUDA
+    if (x.scalar_type() == at::ScalarType::Half) {
+      return forward_cuda_h(x, mean, var, weight, bias, affine, eps);
+    } else {
+      return forward_cuda(x, mean, var, weight, bias, affine, eps);
+    }
+#else
+    AT_ERROR("Not compiled with CUDA support");
+#endif
+  } else {
+    return forward_cpu(x, mean, var, weight, bias, affine, eps);
+  }
+}
+
+std::vector<at::Tensor> edz_eydz(at::Tensor z, at::Tensor dz, at::Tensor weight, at::Tensor bias,
+                                 bool affine, float eps) {
+  if (z.is_cuda()) {
+#ifdef WITH_CUDA
+    if (z.scalar_type() == at::ScalarType::Half) {
+      return edz_eydz_cuda_h(z, dz, weight, bias, affine, eps);
+    } else {
+      return edz_eydz_cuda(z, dz, weight, bias, affine, eps);
+    }
+#else
+    AT_ERROR("Not compiled with CUDA support");
+#endif
+  } else {
+    return edz_eydz_cpu(z, dz, weight, bias, affine, eps);
+  }
+}
+
+at::Tensor backward(at::Tensor z, at::Tensor dz, at::Tensor var, at::Tensor weight, at::Tensor bias,
+                    at::Tensor edz, at::Tensor eydz, bool affine, float eps) {
+  if (z.is_cuda()) {
+#ifdef WITH_CUDA
+    if (z.scalar_type() == at::ScalarType::Half) {
+      return backward_cuda_h(z, dz, var, weight, bias, edz, eydz, affine, eps);
+    } else {
+      return backward_cuda(z, dz, var, weight, bias, edz, eydz, affine, eps);
+    }
+#else
+    AT_ERROR("Not compiled with CUDA support");
+#endif
+  } else {
+    return backward_cpu(z, dz, var, weight, bias, edz, eydz, affine, eps);
+  }
+}
+
+void leaky_relu_forward(at::Tensor z, float slope) {
+  at::leaky_relu_(z, slope);
+}
+
+void leaky_relu_backward(at::Tensor z, at::Tensor dz, float slope) {
+  if (z.is_cuda()) {
+#ifdef WITH_CUDA
+    if (z.scalar_type() == at::ScalarType::Half) {
+      return leaky_relu_backward_cuda_h(z, dz, slope);
+    } else {
+      return leaky_relu_backward_cuda(z, dz, slope);
+    }
+#else
+    AT_ERROR("Not compiled with CUDA support");
+#endif
+  } else {
+    return leaky_relu_backward_cpu(z, dz, slope);
+  }
+}
+
+void elu_forward(at::Tensor z) {
+  at::elu_(z);
+}
+
+void elu_backward(at::Tensor z, at::Tensor dz) {
+  if (z.is_cuda()) {
+#ifdef WITH_CUDA
+    return elu_backward_cuda(z, dz);
+#else
+    AT_ERROR("Not compiled with CUDA support");
+#endif
+  } else {
+    return elu_backward_cpu(z, dz);
+  }
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def("mean_var", &mean_var, "Mean and variance computation");
+  m.def("forward", &forward, "In-place forward computation");
+  m.def("edz_eydz", &edz_eydz, "First part of backward computation");
+  m.def("backward", &backward, "Second part of backward computation");
+  m.def("leaky_relu_forward", &leaky_relu_forward, "Leaky relu forward computation");
+  m.def("leaky_relu_backward", &leaky_relu_backward, "Leaky relu backward computation and inversion");
+  m.def("elu_forward", &elu_forward, "Elu forward computation");
+  m.def("elu_backward", &elu_backward, "Elu backward computation and inversion");
+}

app/models/schp/modules/src/inplace_abn.h

+#pragma once
+
+#include <ATen/ATen.h>
+
+#include <vector>
+
+std::vector<at::Tensor> mean_var_cpu(at::Tensor x);
+std::vector<at::Tensor> mean_var_cuda(at::Tensor x);
+std::vector<at::Tensor> mean_var_cuda_h(at::Tensor x);
+
+at::Tensor forward_cpu(at::Tensor x, at::Tensor mean, at::Tensor var, at::Tensor weight, at::Tensor bias,
+                       bool affine, float eps);
+at::Tensor forward_cuda(at::Tensor x, at::Tensor mean, at::Tensor var, at::Tensor weight, at::Tensor bias,
+                        bool affine, float eps);
+at::Tensor forward_cuda_h(at::Tensor x, at::Tensor mean, at::Tensor var, at::Tensor weight, at::Tensor bias,
+                          bool affine, float eps);
+
+std::vector<at::Tensor> edz_eydz_cpu(at::Tensor z, at::Tensor dz, at::Tensor weight, at::Tensor bias,
+                                     bool affine, float eps);
+std::vector<at::Tensor> edz_eydz_cuda(at::Tensor z, at::Tensor dz, at::Tensor weight, at::Tensor bias,
+                                      bool affine, float eps);
+std::vector<at::Tensor> edz_eydz_cuda_h(at::Tensor z, at::Tensor dz, at::Tensor weight, at::Tensor bias,
+                                        bool affine, float eps);
+
+at::Tensor backward_cpu(at::Tensor z, at::Tensor dz, at::Tensor var, at::Tensor weight, at::Tensor bias,
+                                     at::Tensor edz, at::Tensor eydz, bool affine, float eps);
+at::Tensor backward_cuda(at::Tensor z, at::Tensor dz, at::Tensor var, at::Tensor weight, at::Tensor bias,
+                                      at::Tensor edz, at::Tensor eydz, bool affine, float eps);
+at::Tensor backward_cuda_h(at::Tensor z, at::Tensor dz, at::Tensor var, at::Tensor weight, at::Tensor bias,
+                                        at::Tensor edz, at::Tensor eydz, bool affine, float eps);
+
+void leaky_relu_backward_cpu(at::Tensor z, at::Tensor dz, float slope);
+void leaky_relu_backward_cuda(at::Tensor z, at::Tensor dz, float slope);
+void leaky_relu_backward_cuda_h(at::Tensor z, at::Tensor dz, float slope);
+
+void elu_backward_cpu(at::Tensor z, at::Tensor dz);
+void elu_backward_cuda(at::Tensor z, at::Tensor dz);
+
+static void get_dims(at::Tensor x, int64_t& num, int64_t& chn, int64_t& sp) {
+  num = x.size(0);
+  chn = x.size(1);
+  sp = 1;
+  for (int64_t i = 2; i < x.ndimension(); ++i)
+    sp *= x.size(i);
+}
+
+/*
+ * Specialized CUDA reduction functions for BN
+ */
+#ifdef __CUDACC__
+
+#include "utils/cuda.cuh"
+
+template <typename T, typename Op>
+__device__ T reduce(Op op, int plane, int N, int S) {
+  T sum = (T)0;
+  for (int batch = 0; batch < N; ++batch) {
+    for (int x = threadIdx.x; x < S; x += blockDim.x) {
+      sum += op(batch, plane, x);
+    }
+  }
+
+  // sum over NumThreads within a warp
+  sum = warpSum(sum);
+
+  // 'transpose', and reduce within warp again
+  __shared__ T shared[32];
+  __syncthreads();
+  if (threadIdx.x % WARP_SIZE == 0) {
+    shared[threadIdx.x / WARP_SIZE] = sum;
+  }
+  if (threadIdx.x >= blockDim.x / WARP_SIZE && threadIdx.x < WARP_SIZE) {
+    // zero out the other entries in shared
+    shared[threadIdx.x] = (T)0;
+  }
+  __syncthreads();
+  if (threadIdx.x / WARP_SIZE == 0) {
+    sum = warpSum(shared[threadIdx.x]);
+    if (threadIdx.x == 0) {
+      shared[0] = sum;
+    }
+  }
+  __syncthreads();
+
+  // Everyone picks it up, should be broadcast into the whole gradInput
+  return shared[0];
+}
+#endif

app/models/schp/modules/src/inplace_abn_cpu.cpp

+#include <ATen/ATen.h>
+
+#include <vector>
+
+#include "utils/checks.h"
+#include "inplace_abn.h"
+
+at::Tensor reduce_sum(at::Tensor x) {
+  if (x.ndimension() == 2) {
+    return x.sum(0);
+  } else {
+    auto x_view = x.view({x.size(0), x.size(1), -1});
+    return x_view.sum(-1).sum(0);
+  }
+}
+
+at::Tensor broadcast_to(at::Tensor v, at::Tensor x) {
+  if (x.ndimension() == 2) {
+    return v;
+  } else {
+    std::vector<int64_t> broadcast_size = {1, -1};
+    for (int64_t i = 2; i < x.ndimension(); ++i)
+      broadcast_size.push_back(1);
+
+    return v.view(broadcast_size);
+  }
+}
+
+int64_t count(at::Tensor x) {
+  int64_t count = x.size(0);
+  for (int64_t i = 2; i < x.ndimension(); ++i)
+    count *= x.size(i);
+
+  return count;
+}
+
+at::Tensor invert_affine(at::Tensor z, at::Tensor weight, at::Tensor bias, bool affine, float eps) {
+  if (affine) {
+    return (z - broadcast_to(bias, z)) / broadcast_to(at::abs(weight) + eps, z);
+  } else {
+    return z;
+  }
+}
+
+std::vector<at::Tensor> mean_var_cpu(at::Tensor x) {
+  auto num = count(x);
+  auto mean = reduce_sum(x) / num;
+  auto diff = x - broadcast_to(mean, x);
+  auto var = reduce_sum(diff.pow(2)) / num;
+
+  return {mean, var};
+}
+
+at::Tensor forward_cpu(at::Tensor x, at::Tensor mean, at::Tensor var, at::Tensor weight, at::Tensor bias,
+                       bool affine, float eps) {
+  auto gamma = affine ? at::abs(weight) + eps : at::ones_like(var);
+  auto mul = at::rsqrt(var + eps) * gamma;
+
+  x.sub_(broadcast_to(mean, x));
+  x.mul_(broadcast_to(mul, x));
+  if (affine) x.add_(broadcast_to(bias, x));
+
+  return x;
+}
+
+std::vector<at::Tensor> edz_eydz_cpu(at::Tensor z, at::Tensor dz, at::Tensor weight, at::Tensor bias,
+                                     bool affine, float eps) {
+  auto edz = reduce_sum(dz);
+  auto y = invert_affine(z, weight, bias, affine, eps);
+  auto eydz = reduce_sum(y * dz);
+
+  return {edz, eydz};
+}
+
+at::Tensor backward_cpu(at::Tensor z, at::Tensor dz, at::Tensor var, at::Tensor weight, at::Tensor bias,
+                                     at::Tensor edz, at::Tensor eydz, bool affine, float eps) {
+  auto y = invert_affine(z, weight, bias, affine, eps);
+  auto mul = affine ? at::rsqrt(var + eps) * (at::abs(weight) + eps) : at::rsqrt(var + eps);
+
+  auto num = count(z);
+  auto dx = (dz - broadcast_to(edz / num, dz) - y * broadcast_to(eydz / num, dz)) * broadcast_to(mul, dz);
+  return dx;
+}
+
+void leaky_relu_backward_cpu(at::Tensor z, at::Tensor dz, float slope) {
+  CHECK_CPU_INPUT(z);
+  CHECK_CPU_INPUT(dz);
+
+  AT_DISPATCH_FLOATING_TYPES(z.scalar_type(), "leaky_relu_backward_cpu", ([&] {
+    int64_t count = z.numel();
+    auto *_z = z.data<scalar_t>();
+    auto *_dz = dz.data<scalar_t>();
+
+    for (int64_t i = 0; i < count; ++i) {
+      if (_z[i] < 0) {
+        _z[i] *= 1 / slope;
+        _dz[i] *= slope;
+      }
+    }
+  }));
+}
+
+void elu_backward_cpu(at::Tensor z, at::Tensor dz) {
+  CHECK_CPU_INPUT(z);
+  CHECK_CPU_INPUT(dz);
+
+  AT_DISPATCH_FLOATING_TYPES(z.scalar_type(), "elu_backward_cpu", ([&] {
+    int64_t count = z.numel();
+    auto *_z = z.data<scalar_t>();
+    auto *_dz = dz.data<scalar_t>();
+
+    for (int64_t i = 0; i < count; ++i) {
+      if (_z[i] < 0) {
+        _z[i] = log1p(_z[i]);
+        _dz[i] *= (_z[i] + 1.f);
+      }
+    }
+  }));
+}

app/models/schp/modules/src/inplace_abn_cuda.cu

+#include <ATen/ATen.h>
+
+#include <thrust/device_ptr.h>
+#include <thrust/transform.h>
+
+#include <vector>
+
+#include "utils/checks.h"
+#include "utils/cuda.cuh"
+#include "inplace_abn.h"
+
+#include <ATen/cuda/CUDAContext.h>
+
+// Operations for reduce
+template<typename T>
+struct SumOp {
+  __device__ SumOp(const T *t, int c, int s)
+      : tensor(t), chn(c), sp(s) {}
+  __device__ __forceinline__ T operator()(int batch, int plane, int n) {
+    return tensor[(batch * chn + plane) * sp + n];
+  }
+  const T *tensor;
+  const int chn;
+  const int sp;
+};
+
+template<typename T>
+struct VarOp {
+  __device__ VarOp(T m, const T *t, int c, int s)
+      : mean(m), tensor(t), chn(c), sp(s) {}
+  __device__ __forceinline__ T operator()(int batch, int plane, int n) {
+    T val = tensor[(batch * chn + plane) * sp + n];
+    return (val - mean) * (val - mean);
+  }
+  const T mean;
+  const T *tensor;
+  const int chn;
+  const int sp;
+};
+
+template<typename T>
+struct GradOp {
+  __device__ GradOp(T _weight, T _bias, const T *_z, const T *_dz, int c, int s)
+      : weight(_weight), bias(_bias), z(_z), dz(_dz), chn(c), sp(s) {}
+  __device__ __forceinline__ Pair<T> operator()(int batch, int plane, int n) {
+    T _y = (z[(batch * chn + plane) * sp + n] - bias) / weight;
+    T _dz = dz[(batch * chn + plane) * sp + n];
+    return Pair<T>(_dz, _y * _dz);
+  }
+  const T weight;
+  const T bias;
+  const T *z;
+  const T *dz;
+  const int chn;
+  const int sp;
+};
+
+/***********
+ * mean_var
+ ***********/
+
+template<typename T>
+__global__ void mean_var_kernel(const T *x, T *mean, T *var, int num, int chn, int sp) {
+  int plane = blockIdx.x;
+  T norm = T(1) / T(num * sp);
+
+  T _mean = reduce<T, SumOp<T>>(SumOp<T>(x, chn, sp), plane, num, sp) * norm;
+  __syncthreads();
+  T _var = reduce<T, VarOp<T>>(VarOp<T>(_mean, x, chn, sp), plane, num, sp) * norm;
+
+  if (threadIdx.x == 0) {
+    mean[plane] = _mean;
+    var[plane] = _var;
+  }
+}
+
+std::vector<at::Tensor> mean_var_cuda(at::Tensor x) {
+  CHECK_CUDA_INPUT(x);
+
+  // Extract dimensions
+  int64_t num, chn, sp;
+  get_dims(x, num, chn, sp);
+
+  // Prepare output tensors
+  auto mean = at::empty({chn}, x.options());
+  auto var = at::empty({chn}, x.options());
+
+  // Run kernel
+  dim3 blocks(chn);
+  dim3 threads(getNumThreads(sp));
+  auto stream = at::cuda::getCurrentCUDAStream();
+  AT_DISPATCH_FLOATING_TYPES(x.type(), "mean_var_cuda", ([&] {
+    mean_var_kernel<scalar_t><<<blocks, threads, 0, stream>>>(
+        x.data<scalar_t>(),
+        mean.data<scalar_t>(),
+        var.data<scalar_t>(),
+        num, chn, sp);
+  }));
+
+  return {mean, var};
+}
+
+/**********
+ * forward
+ **********/
+
+template<typename T>
+__global__ void forward_kernel(T *x, const T *mean, const T *var, const T *weight, const T *bias,
+                               bool affine, float eps, int num, int chn, int sp) {
+  int plane = blockIdx.x;
+
+  T _mean = mean[plane];
+  T _var = var[plane];
+  T _weight = affine ? abs(weight[plane]) + eps : T(1);
+  T _bias = affine ? bias[plane] : T(0);
+
+  T mul = rsqrt(_var + eps) * _weight;
+
+  for (int batch = 0; batch < num; ++batch) {
+    for (int n = threadIdx.x; n < sp; n += blockDim.x) {
+      T _x = x[(batch * chn + plane) * sp + n];
+      T _y = (_x - _mean) * mul + _bias;
+
+      x[(batch * chn + plane) * sp + n] = _y;
+    }
+  }
+}
+
+at::Tensor forward_cuda(at::Tensor x, at::Tensor mean, at::Tensor var, at::Tensor weight, at::Tensor bias,
+                        bool affine, float eps) {
+  CHECK_CUDA_INPUT(x);
+  CHECK_CUDA_INPUT(mean);
+  CHECK_CUDA_INPUT(var);
+  CHECK_CUDA_INPUT(weight);
+  CHECK_CUDA_INPUT(bias);
+
+  // Extract dimensions
+  int64_t num, chn, sp;
+  get_dims(x, num, chn, sp);
+
+  // Run kernel
+  dim3 blocks(chn);
+  dim3 threads(getNumThreads(sp));
+  auto stream = at::cuda::getCurrentCUDAStream();
+  AT_DISPATCH_FLOATING_TYPES(x.type(), "forward_cuda", ([&] {
+    forward_kernel<scalar_t><<<blocks, threads, 0, stream>>>(
+        x.data<scalar_t>(),
+        mean.data<scalar_t>(),
+        var.data<scalar_t>(),
+        weight.data<scalar_t>(),
+        bias.data<scalar_t>(),
+        affine, eps, num, chn, sp);
+  }));
+
+  return x;
+}
+
+/***********
+ * edz_eydz
+ ***********/
+
+template<typename T>
+__global__ void edz_eydz_kernel(const T *z, const T *dz, const T *weight, const T *bias,
+                                T *edz, T *eydz, bool affine, float eps, int num, int chn, int sp) {
+  int plane = blockIdx.x;
+
+  T _weight = affine ? abs(weight[plane]) + eps : 1.f;
+  T _bias = affine ? bias[plane] : 0.f;
+
+  Pair<T> res = reduce<Pair<T>, GradOp<T>>(GradOp<T>(_weight, _bias, z, dz, chn, sp), plane, num, sp);
+  __syncthreads();
+
+  if (threadIdx.x == 0) {
+    edz[plane] = res.v1;
+    eydz[plane] = res.v2;
+  }
+}
+
+std::vector<at::Tensor> edz_eydz_cuda(at::Tensor z, at::Tensor dz, at::Tensor weight, at::Tensor bias,
+                                      bool affine, float eps) {
+  CHECK_CUDA_INPUT(z);
+  CHECK_CUDA_INPUT(dz);
+  CHECK_CUDA_INPUT(weight);
+  CHECK_CUDA_INPUT(bias);
+
+  // Extract dimensions
+  int64_t num, chn, sp;
+  get_dims(z, num, chn, sp);
+
+  auto edz = at::empty({chn}, z.options());
+  auto eydz = at::empty({chn}, z.options());
+
+  // Run kernel
+  dim3 blocks(chn);
+  dim3 threads(getNumThreads(sp));
+  auto stream = at::cuda::getCurrentCUDAStream();
+  AT_DISPATCH_FLOATING_TYPES(z.type(), "edz_eydz_cuda", ([&] {
+    edz_eydz_kernel<scalar_t><<<blocks, threads, 0, stream>>>(
+        z.data<scalar_t>(),
+        dz.data<scalar_t>(),
+        weight.data<scalar_t>(),
+        bias.data<scalar_t>(),
+        edz.data<scalar_t>(),
+        eydz.data<scalar_t>(),
+        affine, eps, num, chn, sp);
+  }));
+
+  return {edz, eydz};
+}
+
+/***********
+ * backward
+ ***********/
+
+template<typename T>
+__global__ void backward_kernel(const T *z, const T *dz, const T *var, const T *weight, const T *bias, const T *edz,
+	                        const T *eydz, T *dx, bool affine, float eps, int num, int chn, int sp) {
+  int plane = blockIdx.x;
+
+  T _weight = affine ? abs(weight[plane]) + eps : 1.f;
+  T _bias = affine ? bias[plane] : 0.f;
+  T _var = var[plane];
+  T _edz = edz[plane];
+  T _eydz = eydz[plane];
+
+  T _mul = _weight * rsqrt(_var + eps);
+  T count = T(num * sp);
+
+  for (int batch = 0; batch < num; ++batch) {
+    for (int n = threadIdx.x; n < sp; n += blockDim.x) {
+      T _dz = dz[(batch * chn + plane) * sp + n];
+      T _y = (z[(batch * chn + plane) * sp + n] - _bias) / _weight;
+
+      dx[(batch * chn + plane) * sp + n] = (_dz - _edz / count - _y * _eydz / count) * _mul;
+    }
+  }
+}
+
+at::Tensor backward_cuda(at::Tensor z, at::Tensor dz, at::Tensor var, at::Tensor weight, at::Tensor bias,
+                                      at::Tensor edz, at::Tensor eydz, bool affine, float eps) {
+  CHECK_CUDA_INPUT(z);
+  CHECK_CUDA_INPUT(dz);
+  CHECK_CUDA_INPUT(var);
+  CHECK_CUDA_INPUT(weight);
+  CHECK_CUDA_INPUT(bias);
+  CHECK_CUDA_INPUT(edz);
+  CHECK_CUDA_INPUT(eydz);
+
+  // Extract dimensions
+  int64_t num, chn, sp;
+  get_dims(z, num, chn, sp);
+
+  auto dx = at::zeros_like(z);
+
+  // Run kernel
+  dim3 blocks(chn);
+  dim3 threads(getNumThreads(sp));
+  auto stream = at::cuda::getCurrentCUDAStream();
+  AT_DISPATCH_FLOATING_TYPES(z.type(), "backward_cuda", ([&] {
+    backward_kernel<scalar_t><<<blocks, threads, 0, stream>>>(
+        z.data<scalar_t>(),
+        dz.data<scalar_t>(),
+        var.data<scalar_t>(),
+        weight.data<scalar_t>(),
+        bias.data<scalar_t>(),
+        edz.data<scalar_t>(),
+        eydz.data<scalar_t>(),
+        dx.data<scalar_t>(),
+        affine, eps, num, chn, sp);
+  }));
+
+  return dx;
+}
+
+/**************
+ * activations
+ **************/
+
+template<typename T>
+inline void leaky_relu_backward_impl(T *z, T *dz, float slope, int64_t count) {
+  // Create thrust pointers
+  thrust::device_ptr<T> th_z = thrust::device_pointer_cast(z);
+  thrust::device_ptr<T> th_dz = thrust::device_pointer_cast(dz);
+
+  auto stream = at::cuda::getCurrentCUDAStream();
+  thrust::transform_if(thrust::cuda::par.on(stream),
+                       th_dz, th_dz + count, th_z, th_dz,
+                       [slope] __device__ (const T& dz) { return dz * slope; },
+                       [] __device__ (const T& z) { return z < 0; });
+  thrust::transform_if(thrust::cuda::par.on(stream),
+                       th_z, th_z + count, th_z,
+                       [slope] __device__ (const T& z) { return z / slope; },
+                       [] __device__ (const T& z) { return z < 0; });
+}
+
+void leaky_relu_backward_cuda(at::Tensor z, at::Tensor dz, float slope) {
+  CHECK_CUDA_INPUT(z);
+  CHECK_CUDA_INPUT(dz);
+
+  int64_t count = z.numel();
+
+  AT_DISPATCH_FLOATING_TYPES(z.type(), "leaky_relu_backward_cuda", ([&] {
+    leaky_relu_backward_impl<scalar_t>(z.data<scalar_t>(), dz.data<scalar_t>(), slope, count);
+  }));
+}
+
+template<typename T>
+inline void elu_backward_impl(T *z, T *dz, int64_t count) {
+  // Create thrust pointers
+  thrust::device_ptr<T> th_z = thrust::device_pointer_cast(z);
+  thrust::device_ptr<T> th_dz = thrust::device_pointer_cast(dz);
+
+  auto stream = at::cuda::getCurrentCUDAStream();
+  thrust::transform_if(thrust::cuda::par.on(stream),
+                       th_dz, th_dz + count, th_z, th_z, th_dz,
+                       [] __device__ (const T& dz, const T& z) { return dz * (z + 1.); },
+                       [] __device__ (const T& z) { return z < 0; });
+  thrust::transform_if(thrust::cuda::par.on(stream),
+                       th_z, th_z + count, th_z,
+                       [] __device__ (const T& z) { return log1p(z); },
+                       [] __device__ (const T& z) { return z < 0; });
+}
+
+void elu_backward_cuda(at::Tensor z, at::Tensor dz) {
+  CHECK_CUDA_INPUT(z);
+  CHECK_CUDA_INPUT(dz);
+
+  int64_t count = z.numel();
+
+  AT_DISPATCH_FLOATING_TYPES(z.type(), "leaky_relu_backward_cuda", ([&] {
+    elu_backward_impl<scalar_t>(z.data<scalar_t>(), dz.data<scalar_t>(), count);
+  }));
+}

app/models/schp/modules/src/inplace_abn_cuda_half.cu

+#include <ATen/ATen.h>
+
+#include <cuda_fp16.h>
+
+#include <vector>
+
+#include "utils/checks.h"
+#include "utils/cuda.cuh"
+#include "inplace_abn.h"
+
+#include <ATen/cuda/CUDAContext.h>
+
+// Operations for reduce
+struct SumOpH {
+  __device__ SumOpH(const half *t, int c, int s)
+      : tensor(t), chn(c), sp(s) {}
+  __device__ __forceinline__ float operator()(int batch, int plane, int n) {
+    return __half2float(tensor[(batch * chn + plane) * sp + n]);
+  }
+  const half *tensor;
+  const int chn;
+  const int sp;
+};
+
+struct VarOpH {
+  __device__ VarOpH(float m, const half *t, int c, int s)
+      : mean(m), tensor(t), chn(c), sp(s) {}
+  __device__ __forceinline__ float operator()(int batch, int plane, int n) {
+    const auto t = __half2float(tensor[(batch * chn + plane) * sp + n]);
+    return (t - mean) * (t - mean);
+  }
+  const float mean;
+  const half *tensor;
+  const int chn;
+  const int sp;
+};
+
+struct GradOpH {
+  __device__ GradOpH(float _weight, float _bias, const half *_z, const half *_dz, int c, int s)
+      : weight(_weight), bias(_bias), z(_z), dz(_dz), chn(c), sp(s) {}
+  __device__ __forceinline__ Pair<float> operator()(int batch, int plane, int n) {
+    float _y = (__half2float(z[(batch * chn + plane) * sp + n]) - bias) / weight;
+    float _dz = __half2float(dz[(batch * chn + plane) * sp + n]);
+    return Pair<float>(_dz, _y * _dz);
+  }
+  const float weight;
+  const float bias;
+  const half *z;
+  const half *dz;
+  const int chn;
+  const int sp;
+};
+
+/***********
+ * mean_var
+ ***********/
+
+__global__ void mean_var_kernel_h(const half *x, float *mean, float *var, int num, int chn, int sp) {
+  int plane = blockIdx.x;
+  float norm = 1.f / static_cast<float>(num * sp);
+
+  float _mean = reduce<float, SumOpH>(SumOpH(x, chn, sp), plane, num, sp) * norm;
+  __syncthreads();
+  float _var = reduce<float, VarOpH>(VarOpH(_mean, x, chn, sp), plane, num, sp) * norm;
+
+  if (threadIdx.x == 0) {
+    mean[plane] = _mean;
+    var[plane] = _var;
+  }
+}
+
+std::vector<at::Tensor> mean_var_cuda_h(at::Tensor x) {
+  CHECK_CUDA_INPUT(x);
+
+  // Extract dimensions
+  int64_t num, chn, sp;
+  get_dims(x, num, chn, sp);
+
+  // Prepare output tensors
+  auto mean = at::empty({chn},x.options().dtype(at::kFloat));
+  auto var = at::empty({chn},x.options().dtype(at::kFloat));
+
+  // Run kernel
+  dim3 blocks(chn);
+  dim3 threads(getNumThreads(sp));
+  auto stream = at::cuda::getCurrentCUDAStream();
+  mean_var_kernel_h<<<blocks, threads, 0, stream>>>(
+      reinterpret_cast<half*>(x.data<at::Half>()),
+      mean.data<float>(),
+      var.data<float>(),
+      num, chn, sp);
+
+  return {mean, var};
+}
+
+/**********
+ * forward
+ **********/
+
+__global__ void forward_kernel_h(half *x, const float *mean, const float *var, const float *weight, const float *bias,
+                                 bool affine, float eps, int num, int chn, int sp) {
+  int plane = blockIdx.x;
+
+  const float _mean = mean[plane];
+  const float _var = var[plane];
+  const float _weight = affine ? abs(weight[plane]) + eps : 1.f;
+  const float _bias = affine ? bias[plane] : 0.f;
+
+  const float mul = rsqrt(_var + eps) * _weight;
+
+  for (int batch = 0; batch < num; ++batch) {
+    for (int n = threadIdx.x; n < sp; n += blockDim.x) {
+      half *x_ptr = x + (batch * chn + plane) * sp + n;
+      float _x = __half2float(*x_ptr);
+      float _y = (_x - _mean) * mul + _bias;
+
+      *x_ptr = __float2half(_y);
+    }
+  }
+}
+
+at::Tensor forward_cuda_h(at::Tensor x, at::Tensor mean, at::Tensor var, at::Tensor weight, at::Tensor bias,
+                        bool affine, float eps) {
+  CHECK_CUDA_INPUT(x);
+  CHECK_CUDA_INPUT(mean);
+  CHECK_CUDA_INPUT(var);
+  CHECK_CUDA_INPUT(weight);
+  CHECK_CUDA_INPUT(bias);
+
+  // Extract dimensions
+  int64_t num, chn, sp;
+  get_dims(x, num, chn, sp);
+
+  // Run kernel
+  dim3 blocks(chn);
+  dim3 threads(getNumThreads(sp));
+  auto stream = at::cuda::getCurrentCUDAStream();
+  forward_kernel_h<<<blocks, threads, 0, stream>>>(
+      reinterpret_cast<half*>(x.data<at::Half>()),
+      mean.data<float>(),
+      var.data<float>(),
+      weight.data<float>(),
+      bias.data<float>(),
+      affine, eps, num, chn, sp);
+
+  return x;
+}
+
+__global__ void edz_eydz_kernel_h(const half *z, const half *dz, const float *weight, const float *bias,
+                                float *edz, float *eydz, bool affine, float eps, int num, int chn, int sp) {
+  int plane = blockIdx.x;
+
+  float _weight = affine ? abs(weight[plane]) + eps : 1.f;
+  float _bias = affine ? bias[plane] : 0.f;
+
+  Pair<float> res = reduce<Pair<float>, GradOpH>(GradOpH(_weight, _bias, z, dz, chn, sp), plane, num, sp);
+  __syncthreads();
+
+  if (threadIdx.x == 0) {
+    edz[plane] = res.v1;
+    eydz[plane] = res.v2;
+  }
+}
+
+std::vector<at::Tensor> edz_eydz_cuda_h(at::Tensor z, at::Tensor dz, at::Tensor weight, at::Tensor bias,
+                                      bool affine, float eps) {
+  CHECK_CUDA_INPUT(z);
+  CHECK_CUDA_INPUT(dz);
+  CHECK_CUDA_INPUT(weight);
+  CHECK_CUDA_INPUT(bias);
+
+  // Extract dimensions
+  int64_t num, chn, sp;
+  get_dims(z, num, chn, sp);
+
+  auto edz = at::empty({chn},z.options().dtype(at::kFloat));
+  auto eydz = at::empty({chn},z.options().dtype(at::kFloat));
+
+  // Run kernel
+  dim3 blocks(chn);
+  dim3 threads(getNumThreads(sp));
+  auto stream = at::cuda::getCurrentCUDAStream();
+  edz_eydz_kernel_h<<<blocks, threads, 0, stream>>>(
+        reinterpret_cast<half*>(z.data<at::Half>()),
+        reinterpret_cast<half*>(dz.data<at::Half>()),
+        weight.data<float>(),
+        bias.data<float>(),
+        edz.data<float>(),
+        eydz.data<float>(),
+        affine, eps, num, chn, sp);
+ 
+  return {edz, eydz};
+}
+
+__global__ void backward_kernel_h(const half *z, const half *dz, const float *var, const float *weight, const float *bias, const float *edz,
+                                  const float *eydz, half *dx, bool affine, float eps, int num, int chn, int sp) {
+  int plane = blockIdx.x;
+
+  float _weight = affine ? abs(weight[plane]) + eps : 1.f;
+  float _bias = affine ? bias[plane] : 0.f;
+  float _var = var[plane];
+  float _edz = edz[plane];
+  float _eydz = eydz[plane];
+
+  float _mul = _weight * rsqrt(_var + eps);
+  float count = float(num * sp);
+
+  for (int batch = 0; batch < num; ++batch) {
+    for (int n = threadIdx.x; n < sp; n += blockDim.x) {
+      float _dz = __half2float(dz[(batch * chn + plane) * sp + n]);
+      float _y = (__half2float(z[(batch * chn + plane) * sp + n]) - _bias) / _weight;
+
+      dx[(batch * chn + plane) * sp + n] = __float2half((_dz - _edz / count - _y * _eydz / count) * _mul);
+    }
+  }
+}
+
+at::Tensor backward_cuda_h(at::Tensor z, at::Tensor dz, at::Tensor var, at::Tensor weight, at::Tensor bias,
+                                      at::Tensor edz, at::Tensor eydz, bool affine, float eps) {
+  CHECK_CUDA_INPUT(z);
+  CHECK_CUDA_INPUT(dz);
+  CHECK_CUDA_INPUT(var);
+  CHECK_CUDA_INPUT(weight);
+  CHECK_CUDA_INPUT(bias);
+  CHECK_CUDA_INPUT(edz);
+  CHECK_CUDA_INPUT(eydz);
+
+  // Extract dimensions
+  int64_t num, chn, sp;
+  get_dims(z, num, chn, sp);
+
+  auto dx = at::zeros_like(z);
+
+  // Run kernel
+  dim3 blocks(chn);
+  dim3 threads(getNumThreads(sp));
+  auto stream = at::cuda::getCurrentCUDAStream();
+  backward_kernel_h<<<blocks, threads, 0, stream>>>(
+        reinterpret_cast<half*>(z.data<at::Half>()),
+        reinterpret_cast<half*>(dz.data<at::Half>()),
+        var.data<float>(),
+        weight.data<float>(),
+        bias.data<float>(),
+        edz.data<float>(),
+        eydz.data<float>(),
+        reinterpret_cast<half*>(dx.data<at::Half>()),
+        affine, eps, num, chn, sp);
+
+  return dx;
+}
+
+__global__ void leaky_relu_backward_impl_h(half *z, half *dz, float slope, int64_t count) {
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < count;  i += blockDim.x * gridDim.x){
+    float _z = __half2float(z[i]);
+    if (_z < 0) {
+      dz[i] = __float2half(__half2float(dz[i]) * slope);
+      z[i] = __float2half(_z / slope);
+    }
+  }
+}
+
+void leaky_relu_backward_cuda_h(at::Tensor z, at::Tensor dz, float slope) {
+  CHECK_CUDA_INPUT(z);
+  CHECK_CUDA_INPUT(dz);
+
+  int64_t count = z.numel();
+  dim3 threads(getNumThreads(count));
+  dim3 blocks = (count + threads.x - 1) / threads.x;
+  auto stream = at::cuda::getCurrentCUDAStream();
+  leaky_relu_backward_impl_h<<<blocks, threads, 0, stream>>>(
+      reinterpret_cast<half*>(z.data<at::Half>()),
+      reinterpret_cast<half*>(dz.data<at::Half>()),
+      slope, count);
+}
+

app/models/schp/modules/src/utils/checks.h

+#pragma once
+
+#include <ATen/ATen.h>
+
+// Define AT_CHECK for old version of ATen where the same function was called AT_ASSERT
+#ifndef AT_CHECK
+#define AT_CHECK AT_ASSERT
+#endif
+
+#define CHECK_CUDA(x) AT_CHECK((x).type().is_cuda(), #x " must be a CUDA tensor")
+#define CHECK_CPU(x) AT_CHECK(!(x).type().is_cuda(), #x " must be a CPU tensor")
+#define CHECK_CONTIGUOUS(x) AT_CHECK((x).is_contiguous(), #x " must be contiguous")
+
+#define CHECK_CUDA_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x)
+#define CHECK_CPU_INPUT(x) CHECK_CPU(x); CHECK_CONTIGUOUS(x)
\ No newline at end of file

app/models/schp/modules/src/utils/common.h

+#pragma once
+
+#include <ATen/ATen.h>
+
+/*
+ * Functions to share code between CPU and GPU
+ */
+
+#ifdef __CUDACC__
+// CUDA versions
+
+#define HOST_DEVICE __host__ __device__
+#define INLINE_HOST_DEVICE __host__ __device__ inline
+#define FLOOR(x) floor(x)
+
+#if __CUDA_ARCH__ >= 600
+// Recent compute capabilities have block-level atomicAdd for all data types, so we use that
+#define ACCUM(x,y) atomicAdd_block(&(x),(y))
+#else
+// Older architectures don't have block-level atomicAdd, nor atomicAdd for doubles, so we defer to atomicAdd for float
+// and use the known atomicCAS-based implementation for double
+template<typename data_t>
+__device__ inline data_t atomic_add(data_t *address, data_t val) {
+  return atomicAdd(address, val);
+}
+
+template<>
+__device__ inline double atomic_add(double *address, double val) {
+  unsigned long long int* address_as_ull = (unsigned long long int*)address;
+  unsigned long long int old = *address_as_ull, assumed;
+  do {
+    assumed = old;
+    old = atomicCAS(address_as_ull, assumed, __double_as_longlong(val + __longlong_as_double(assumed)));
+  } while (assumed != old);
+  return __longlong_as_double(old);
+}
+
+#define ACCUM(x,y) atomic_add(&(x),(y))
+#endif // #if __CUDA_ARCH__ >= 600
+
+#else
+// CPU versions
+
+#define HOST_DEVICE
+#define INLINE_HOST_DEVICE inline
+#define FLOOR(x) std::floor(x)
+#define ACCUM(x,y) (x) += (y)
+
+#endif // #ifdef __CUDACC__
\ No newline at end of file

app/models/schp/modules/src/utils/cuda.cuh

+#pragma once
+
+/*
+ * General settings and functions
+ */
+const int WARP_SIZE = 32;
+const int MAX_BLOCK_SIZE = 1024;
+
+static int getNumThreads(int nElem) {
+  int threadSizes[6] = {32, 64, 128, 256, 512, MAX_BLOCK_SIZE};
+  for (int i = 0; i < 6; ++i) {
+    if (nElem <= threadSizes[i]) {
+      return threadSizes[i];
+    }
+  }
+  return MAX_BLOCK_SIZE;
+}
+
+/*
+ * Reduction utilities
+ */
+template <typename T>
+__device__ __forceinline__ T WARP_SHFL_XOR(T value, int laneMask, int width = warpSize,
+                                           unsigned int mask = 0xffffffff) {
+#if CUDART_VERSION >= 9000
+  return __shfl_xor_sync(mask, value, laneMask, width);
+#else
+  return __shfl_xor(value, laneMask, width);
+#endif
+}
+
+__device__ __forceinline__ int getMSB(int val) { return 31 - __clz(val); }
+
+template<typename T>
+struct Pair {
+  T v1, v2;
+  __device__ Pair() {}
+  __device__ Pair(T _v1, T _v2) : v1(_v1), v2(_v2) {}
+  __device__ Pair(T v) : v1(v), v2(v) {}
+  __device__ Pair(int v) : v1(v), v2(v) {}
+  __device__ Pair &operator+=(const Pair<T> &a) {
+    v1 += a.v1;
+    v2 += a.v2;
+    return *this;
+  }
+};
+
+template<typename T>
+static __device__ __forceinline__ T warpSum(T val) {
+#if __CUDA_ARCH__ >= 300
+  for (int i = 0; i < getMSB(WARP_SIZE); ++i) {
+    val += WARP_SHFL_XOR(val, 1 << i, WARP_SIZE);
+  }
+#else
+  __shared__ T values[MAX_BLOCK_SIZE];
+  values[threadIdx.x] = val;
+  __threadfence_block();
+  const int base = (threadIdx.x / WARP_SIZE) * WARP_SIZE;
+  for (int i = 1; i < WARP_SIZE; i++) {
+    val += values[base + ((i + threadIdx.x) % WARP_SIZE)];
+  }
+#endif
+  return val;
+}
+
+template<typename T>
+static __device__ __forceinline__ Pair<T> warpSum(Pair<T> value) {
+  value.v1 = warpSum(value.v1);
+  value.v2 = warpSum(value.v2);
+  return value;
+}
\ No newline at end of file

app/models/schp/networks/__init__.py

+from __future__ import absolute_import
+
+from app.models.schp.networks.AugmentCE2P import resnet101
+
+__factory = {
+    'resnet101': resnet101,
+}
+
+
+def init_model(name, *args, **kwargs):
+    if name not in __factory.keys():
+        raise KeyError("Unknown model arch: {}".format(name))
+    return __factory[name](*args, **kwargs)
\ No newline at end of file

app/models/schp/networks/backbone/mobilenetv2.py

+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+
+"""
+@Author  :   Peike Li
+@Contact :   peike.li@yahoo.com
+@File    :   mobilenetv2.py
+@Time    :   8/4/19 3:35 PM
+@Desc    :   
+@License :   This source code is licensed under the license found in the 
+             LICENSE file in the root directory of this source tree.
+"""
+
+import torch.nn as nn
+import math
+import functools
+
+from modules import InPlaceABN, InPlaceABNSync
+
+BatchNorm2d = functools.partial(InPlaceABNSync, activation='none')
+
+__all__ = ['mobilenetv2']
+
+
+def conv_bn(inp, oup, stride):
+    return nn.Sequential(
+        nn.Conv2d(inp, oup, 3, stride, 1, bias=False),
+        BatchNorm2d(oup),
+        nn.ReLU6(inplace=True)
+    )
+
+
+def conv_1x1_bn(inp, oup):
+    return nn.Sequential(
+        nn.Conv2d(inp, oup, 1, 1, 0, bias=False),
+        BatchNorm2d(oup),
+        nn.ReLU6(inplace=True)
+    )
+
+
+class InvertedResidual(nn.Module):
+    def __init__(self, inp, oup, stride, expand_ratio):
+        super(InvertedResidual, self).__init__()
+        self.stride = stride
+        assert stride in [1, 2]
+
+        hidden_dim = round(inp * expand_ratio)
+        self.use_res_connect = self.stride == 1 and inp == oup
+
+        if expand_ratio == 1:
+            self.conv = nn.Sequential(
+                # dw
+                nn.Conv2d(hidden_dim, hidden_dim, 3, stride, 1, groups=hidden_dim, bias=False),
+                BatchNorm2d(hidden_dim),
+                nn.ReLU6(inplace=True),
+                # pw-linear
+                nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
+                BatchNorm2d(oup),
+            )
+        else:
+            self.conv = nn.Sequential(
+                # pw
+                nn.Conv2d(inp, hidden_dim, 1, 1, 0, bias=False),
+                BatchNorm2d(hidden_dim),
+                nn.ReLU6(inplace=True),
+                # dw
+                nn.Conv2d(hidden_dim, hidden_dim, 3, stride, 1, groups=hidden_dim, bias=False),
+                BatchNorm2d(hidden_dim),
+                nn.ReLU6(inplace=True),
+                # pw-linear
+                nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
+                BatchNorm2d(oup),
+            )
+
+    def forward(self, x):
+        if self.use_res_connect:
+            return x + self.conv(x)
+        else:
+            return self.conv(x)
+
+
+class MobileNetV2(nn.Module):
+    def __init__(self, n_class=1000, input_size=224, width_mult=1.):
+        super(MobileNetV2, self).__init__()
+        block = InvertedResidual
+        input_channel = 32
+        last_channel = 1280
+        interverted_residual_setting = [
+            # t, c, n, s
+            [1, 16, 1, 1],
+            [6, 24, 2, 2],  # layer 2
+            [6, 32, 3, 2],  # layer 3
+            [6, 64, 4, 2],
+            [6, 96, 3, 1],  # layer 4
+            [6, 160, 3, 2],
+            [6, 320, 1, 1],  # layer 5
+        ]
+
+        # building first layer
+        assert input_size % 32 == 0
+        input_channel = int(input_channel * width_mult)
+        self.last_channel = int(last_channel * width_mult) if width_mult > 1.0 else last_channel
+        self.features = [conv_bn(3, input_channel, 2)]
+        # building inverted residual blocks
+        for t, c, n, s in interverted_residual_setting:
+            output_channel = int(c * width_mult)
+            for i in range(n):
+                if i == 0:
+                    self.features.append(block(input_channel, output_channel, s, expand_ratio=t))
+                else:
+                    self.features.append(block(input_channel, output_channel, 1, expand_ratio=t))
+                input_channel = output_channel
+        # building last several layers
+        self.features.append(conv_1x1_bn(input_channel, self.last_channel))
+        # make it nn.Sequential
+        self.features = nn.Sequential(*self.features)
+
+        # building classifier
+        self.classifier = nn.Sequential(
+            nn.Dropout(0.2),
+            nn.Linear(self.last_channel, n_class),
+        )
+
+        self._initialize_weights()
+
+    def forward(self, x):
+        x = self.features(x)
+        x = x.mean(3).mean(2)
+        x = self.classifier(x)
+        return x
+
+    def _initialize_weights(self):
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+                m.weight.data.normal_(0, math.sqrt(2. / n))
+                if m.bias is not None:
+                    m.bias.data.zero_()
+            elif isinstance(m, BatchNorm2d):
+                m.weight.data.fill_(1)
+                m.bias.data.zero_()
+            elif isinstance(m, nn.Linear):
+                n = m.weight.size(1)
+                m.weight.data.normal_(0, 0.01)
+                m.bias.data.zero_()
+
+
+def mobilenetv2(pretrained=False, **kwargs):
+    """Constructs a MobileNet_V2 model.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
+    model = MobileNetV2(n_class=1000, **kwargs)
+    if pretrained:
+        model.load_state_dict(load_url(model_urls['mobilenetv2']), strict=False)
+    return model

app/models/schp/networks/backbone/resnet.py

+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+
+"""
+@Author  :   Peike Li
+@Contact :   peike.li@yahoo.com
+@File    :   resnet.py
+@Time    :   8/4/19 3:35 PM
+@Desc    :   
+@License :   This source code is licensed under the license found in the 
+             LICENSE file in the root directory of this source tree.
+"""
+
+import functools
+import torch.nn as nn
+import math
+from torch.utils.model_zoo import load_url
+
+from modules import InPlaceABNSync
+
+BatchNorm2d = functools.partial(InPlaceABNSync, activation='none')
+
+__all__ = ['ResNet', 'resnet18', 'resnet50', 'resnet101']  # resnet101 is coming soon!
+
+model_urls = {
+    'resnet18': 'http://sceneparsing.csail.mit.edu/model/pretrained_resnet/resnet18-imagenet.pth',
+    'resnet50': 'http://sceneparsing.csail.mit.edu/model/pretrained_resnet/resnet50-imagenet.pth',
+    'resnet101': 'http://sceneparsing.csail.mit.edu/model/pretrained_resnet/resnet101-imagenet.pth'
+}
+
+
+def conv3x3(in_planes, out_planes, stride=1):
+    "3x3 convolution with padding"
+    return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
+                     padding=1, bias=False)
+
+
+class BasicBlock(nn.Module):
+    expansion = 1
+
+    def __init__(self, inplanes, planes, stride=1, downsample=None):
+        super(BasicBlock, self).__init__()
+        self.conv1 = conv3x3(inplanes, planes, stride)
+        self.bn1 = BatchNorm2d(planes)
+        self.relu = nn.ReLU(inplace=True)
+        self.conv2 = conv3x3(planes, planes)
+        self.bn2 = BatchNorm2d(planes)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        residual = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+
+        if self.downsample is not None:
+            residual = self.downsample(x)
+
+        out += residual
+        out = self.relu(out)
+
+        return out
+
+
+class Bottleneck(nn.Module):
+    expansion = 4
+
+    def __init__(self, inplanes, planes, stride=1, downsample=None):
+        super(Bottleneck, self).__init__()
+        self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False)
+        self.bn1 = BatchNorm2d(planes)
+        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride,
+                               padding=1, bias=False)
+        self.bn2 = BatchNorm2d(planes)
+        self.conv3 = nn.Conv2d(planes, planes * 4, kernel_size=1, bias=False)
+        self.bn3 = BatchNorm2d(planes * 4)
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        residual = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+        out = self.relu(out)
+
+        out = self.conv3(out)
+        out = self.bn3(out)
+
+        if self.downsample is not None:
+            residual = self.downsample(x)
+
+        out += residual
+        out = self.relu(out)
+
+        return out
+
+
+class ResNet(nn.Module):
+
+    def __init__(self, block, layers, num_classes=1000):
+        self.inplanes = 128
+        super(ResNet, self).__init__()
+        self.conv1 = conv3x3(3, 64, stride=2)
+        self.bn1 = BatchNorm2d(64)
+        self.relu1 = nn.ReLU(inplace=True)
+        self.conv2 = conv3x3(64, 64)
+        self.bn2 = BatchNorm2d(64)
+        self.relu2 = nn.ReLU(inplace=True)
+        self.conv3 = conv3x3(64, 128)
+        self.bn3 = BatchNorm2d(128)
+        self.relu3 = nn.ReLU(inplace=True)
+        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+
+        self.layer1 = self._make_layer(block, 64, layers[0])
+        self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
+        self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
+        self.layer4 = self._make_layer(block, 512, layers[3], stride=2)
+        self.avgpool = nn.AvgPool2d(7, stride=1)
+        self.fc = nn.Linear(512 * block.expansion, num_classes)
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+                m.weight.data.normal_(0, math.sqrt(2. / n))
+            elif isinstance(m, BatchNorm2d):
+                m.weight.data.fill_(1)
+                m.bias.data.zero_()
+
+    def _make_layer(self, block, planes, blocks, stride=1):
+        downsample = None
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                nn.Conv2d(self.inplanes, planes * block.expansion,
+                          kernel_size=1, stride=stride, bias=False),
+                BatchNorm2d(planes * block.expansion),
+            )
+
+        layers = []
+        layers.append(block(self.inplanes, planes, stride, downsample))
+        self.inplanes = planes * block.expansion
+        for i in range(1, blocks):
+            layers.append(block(self.inplanes, planes))
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        x = self.relu1(self.bn1(self.conv1(x)))
+        x = self.relu2(self.bn2(self.conv2(x)))
+        x = self.relu3(self.bn3(self.conv3(x)))
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+
+        x = self.avgpool(x)
+        x = x.view(x.size(0), -1)
+        x = self.fc(x)
+
+        return x
+
+
+def resnet18(pretrained=False, **kwargs):
+    """Constructs a ResNet-18 model.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
+    model = ResNet(BasicBlock, [2, 2, 2, 2], **kwargs)
+    if pretrained:
+        model.load_state_dict(load_url(model_urls['resnet18']))
+    return model
+
+
+def resnet50(pretrained=False, **kwargs):
+    """Constructs a ResNet-50 model.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
+    model = ResNet(Bottleneck, [3, 4, 6, 3], **kwargs)
+    if pretrained:
+        model.load_state_dict(load_url(model_urls['resnet50']), strict=False)
+    return model
+
+
+def resnet101(pretrained=False, **kwargs):
+    """Constructs a ResNet-101 model.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
+    model = ResNet(Bottleneck, [3, 4, 23, 3], **kwargs)
+    if pretrained:
+        model.load_state_dict(load_url(model_urls['resnet101']), strict=False)
+    return model

app/models/schp/networks/backbone/resnext.py

+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+
+"""
+@Author  :   Peike Li
+@Contact :   peike.li@yahoo.com
+@File    :   resnext.py.py
+@Time    :   8/11/19 8:58 PM
+@Desc    :   
+@License :   This source code is licensed under the license found in the 
+             LICENSE file in the root directory of this source tree.
+"""
+import functools
+import torch.nn as nn
+import math
+from torch.utils.model_zoo import load_url
+
+from modules import InPlaceABNSync
+
+BatchNorm2d = functools.partial(InPlaceABNSync, activation='none')
+
+__all__ = ['ResNeXt', 'resnext101']  # support resnext 101
+
+model_urls = {
+    'resnext50': 'http://sceneparsing.csail.mit.edu/model/pretrained_resnet/resnext50-imagenet.pth',
+    'resnext101': 'http://sceneparsing.csail.mit.edu/model/pretrained_resnet/resnext101-imagenet.pth'
+}
+
+
+def conv3x3(in_planes, out_planes, stride=1):
+    "3x3 convolution with padding"
+    return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
+                     padding=1, bias=False)
+
+
+class GroupBottleneck(nn.Module):
+    expansion = 2
+
+    def __init__(self, inplanes, planes, stride=1, groups=1, downsample=None):
+        super(GroupBottleneck, self).__init__()
+        self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False)
+        self.bn1 = BatchNorm2d(planes)
+        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride,
+                               padding=1, groups=groups, bias=False)
+        self.bn2 = BatchNorm2d(planes)
+        self.conv3 = nn.Conv2d(planes, planes * 2, kernel_size=1, bias=False)
+        self.bn3 = BatchNorm2d(planes * 2)
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        residual = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+        out = self.relu(out)
+
+        out = self.conv3(out)
+        out = self.bn3(out)
+
+        if self.downsample is not None:
+            residual = self.downsample(x)
+
+        out += residual
+        out = self.relu(out)
+
+        return out
+
+
+class ResNeXt(nn.Module):
+
+    def __init__(self, block, layers, groups=32, num_classes=1000):
+        self.inplanes = 128
+        super(ResNeXt, self).__init__()
+        self.conv1 = conv3x3(3, 64, stride=2)
+        self.bn1 = BatchNorm2d(64)
+        self.relu1 = nn.ReLU(inplace=True)
+        self.conv2 = conv3x3(64, 64)
+        self.bn2 = BatchNorm2d(64)
+        self.relu2 = nn.ReLU(inplace=True)
+        self.conv3 = conv3x3(64, 128)
+        self.bn3 = BatchNorm2d(128)
+        self.relu3 = nn.ReLU(inplace=True)
+        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+
+        self.layer1 = self._make_layer(block, 128, layers[0], groups=groups)
+        self.layer2 = self._make_layer(block, 256, layers[1], stride=2, groups=groups)
+        self.layer3 = self._make_layer(block, 512, layers[2], stride=2, groups=groups)
+        self.layer4 = self._make_layer(block, 1024, layers[3], stride=2, groups=groups)
+        self.avgpool = nn.AvgPool2d(7, stride=1)
+        self.fc = nn.Linear(1024 * block.expansion, num_classes)
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels // m.groups
+                m.weight.data.normal_(0, math.sqrt(2. / n))
+            elif isinstance(m, BatchNorm2d):
+                m.weight.data.fill_(1)
+                m.bias.data.zero_()
+
+    def _make_layer(self, block, planes, blocks, stride=1, groups=1):
+        downsample = None
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                nn.Conv2d(self.inplanes, planes * block.expansion,
+                          kernel_size=1, stride=stride, bias=False),
+                BatchNorm2d(planes * block.expansion),
+            )
+
+        layers = []
+        layers.append(block(self.inplanes, planes, stride, groups, downsample))
+        self.inplanes = planes * block.expansion
+        for i in range(1, blocks):
+            layers.append(block(self.inplanes, planes, groups=groups))
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        x = self.relu1(self.bn1(self.conv1(x)))
+        x = self.relu2(self.bn2(self.conv2(x)))
+        x = self.relu3(self.bn3(self.conv3(x)))
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+
+        x = self.avgpool(x)
+        x = x.view(x.size(0), -1)
+        x = self.fc(x)
+
+        return x
+
+
+def resnext101(pretrained=False, **kwargs):
+    """Constructs a ResNet-101 model.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on Places
+    """
+    model = ResNeXt(GroupBottleneck, [3, 4, 23, 3], **kwargs)
+    if pretrained:
+        model.load_state_dict(load_url(model_urls['resnext101']), strict=False)
+    return model

app/models/schp/networks/context_encoding/aspp.py

+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+
+"""
+@Author  :   Peike Li
+@Contact :   peike.li@yahoo.com
+@File    :   aspp.py
+@Time    :   8/4/19 3:36 PM
+@Desc    :   
+@License :   This source code is licensed under the license found in the 
+             LICENSE file in the root directory of this source tree.
+"""
+
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+
+from modules import InPlaceABNSync
+
+
+class ASPPModule(nn.Module):
+    """
+    Reference:
+        Chen, Liang-Chieh, et al. *"Rethinking Atrous Convolution for Semantic Image Segmentation."*
+    """
+    def __init__(self, features, out_features=512, inner_features=256, dilations=(12, 24, 36)):
+        super(ASPPModule, self).__init__()
+
+        self.conv1 = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)),
+                                   nn.Conv2d(features, inner_features, kernel_size=1, padding=0, dilation=1,
+                                             bias=False),
+                                   InPlaceABNSync(inner_features))
+        self.conv2 = nn.Sequential(
+            nn.Conv2d(features, inner_features, kernel_size=1, padding=0, dilation=1, bias=False),
+            InPlaceABNSync(inner_features))
+        self.conv3 = nn.Sequential(
+            nn.Conv2d(features, inner_features, kernel_size=3, padding=dilations[0], dilation=dilations[0], bias=False),
+            InPlaceABNSync(inner_features))
+        self.conv4 = nn.Sequential(
+            nn.Conv2d(features, inner_features, kernel_size=3, padding=dilations[1], dilation=dilations[1], bias=False),
+            InPlaceABNSync(inner_features))
+        self.conv5 = nn.Sequential(
+            nn.Conv2d(features, inner_features, kernel_size=3, padding=dilations[2], dilation=dilations[2], bias=False),
+            InPlaceABNSync(inner_features))
+
+        self.bottleneck = nn.Sequential(
+            nn.Conv2d(inner_features * 5, out_features, kernel_size=1, padding=0, dilation=1, bias=False),
+            InPlaceABNSync(out_features),
+            nn.Dropout2d(0.1)
+        )
+
+    def forward(self, x):
+        _, _, h, w = x.size()
+
+        feat1 = F.interpolate(self.conv1(x), size=(h, w), mode='bilinear', align_corners=True)
+
+        feat2 = self.conv2(x)
+        feat3 = self.conv3(x)
+        feat4 = self.conv4(x)
+        feat5 = self.conv5(x)
+        out = torch.cat((feat1, feat2, feat3, feat4, feat5), 1)
+
+        bottle = self.bottleneck(out)
+        return bottle
\ No newline at end of file

app/models/schp/networks/context_encoding/ocnet.py

+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+
+"""
+@Author  :   Peike Li
+@Contact :   peike.li@yahoo.com
+@File    :   ocnet.py
+@Time    :   8/4/19 3:36 PM
+@Desc    :   
+@License :   This source code is licensed under the license found in the 
+             LICENSE file in the root directory of this source tree.
+"""
+
+import functools
+
+import torch
+import torch.nn as nn
+from torch.autograd import Variable
+from torch.nn import functional as F
+
+from modules import InPlaceABNSync
+BatchNorm2d = functools.partial(InPlaceABNSync, activation='none')
+
+
+class _SelfAttentionBlock(nn.Module):
+    '''
+    The basic implementation for self-attention block/non-local block
+    Input:
+        N X C X H X W
+    Parameters:
+        in_channels       : the dimension of the input feature map
+        key_channels      : the dimension after the key/query transform
+        value_channels    : the dimension after the value transform
+        scale             : choose the scale to downsample the input feature maps (save memory cost)
+    Return:
+        N X C X H X W
+        position-aware context features.(w/o concate or add with the input)
+    '''
+
+    def __init__(self, in_channels, key_channels, value_channels, out_channels=None, scale=1):
+        super(_SelfAttentionBlock, self).__init__()
+        self.scale = scale
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.key_channels = key_channels
+        self.value_channels = value_channels
+        if out_channels == None:
+            self.out_channels = in_channels
+        self.pool = nn.MaxPool2d(kernel_size=(scale, scale))
+        self.f_key = nn.Sequential(
+            nn.Conv2d(in_channels=self.in_channels, out_channels=self.key_channels,
+                      kernel_size=1, stride=1, padding=0),
+            InPlaceABNSync(self.key_channels),
+        )
+        self.f_query = self.f_key
+        self.f_value = nn.Conv2d(in_channels=self.in_channels, out_channels=self.value_channels,
+                                 kernel_size=1, stride=1, padding=0)
+        self.W = nn.Conv2d(in_channels=self.value_channels, out_channels=self.out_channels,
+                           kernel_size=1, stride=1, padding=0)
+        nn.init.constant(self.W.weight, 0)
+        nn.init.constant(self.W.bias, 0)
+
+    def forward(self, x):
+        batch_size, h, w = x.size(0), x.size(2), x.size(3)
+        if self.scale > 1:
+            x = self.pool(x)
+
+        value = self.f_value(x).view(batch_size, self.value_channels, -1)
+        value = value.permute(0, 2, 1)
+        query = self.f_query(x).view(batch_size, self.key_channels, -1)
+        query = query.permute(0, 2, 1)
+        key = self.f_key(x).view(batch_size, self.key_channels, -1)
+
+        sim_map = torch.matmul(query, key)
+        sim_map = (self.key_channels ** -.5) * sim_map
+        sim_map = F.softmax(sim_map, dim=-1)
+
+        context = torch.matmul(sim_map, value)
+        context = context.permute(0, 2, 1).contiguous()
+        context = context.view(batch_size, self.value_channels, *x.size()[2:])
+        context = self.W(context)
+        if self.scale > 1:
+            context = F.upsample(input=context, size=(h, w), mode='bilinear', align_corners=True)
+        return context
+
+
+class SelfAttentionBlock2D(_SelfAttentionBlock):
+    def __init__(self, in_channels, key_channels, value_channels, out_channels=None, scale=1):
+        super(SelfAttentionBlock2D, self).__init__(in_channels,
+                                                   key_channels,
+                                                   value_channels,
+                                                   out_channels,
+                                                   scale)
+
+
+class BaseOC_Module(nn.Module):
+    """
+    Implementation of the BaseOC module
+    Parameters:
+        in_features / out_features: the channels of the input / output feature maps.
+        dropout: we choose 0.05 as the default value.
+        size: you can apply multiple sizes. Here we only use one size.
+    Return:
+        features fused with Object context information.
+    """
+
+    def __init__(self, in_channels, out_channels, key_channels, value_channels, dropout, sizes=([1])):
+        super(BaseOC_Module, self).__init__()
+        self.stages = []
+        self.stages = nn.ModuleList(
+            [self._make_stage(in_channels, out_channels, key_channels, value_channels, size) for size in sizes])
+        self.conv_bn_dropout = nn.Sequential(
+            nn.Conv2d(2 * in_channels, out_channels, kernel_size=1, padding=0),
+            InPlaceABNSync(out_channels),
+            nn.Dropout2d(dropout)
+        )
+
+    def _make_stage(self, in_channels, output_channels, key_channels, value_channels, size):
+        return SelfAttentionBlock2D(in_channels,
+                                    key_channels,
+                                    value_channels,
+                                    output_channels,
+                                    size)
+
+    def forward(self, feats):
+        priors = [stage(feats) for stage in self.stages]
+        context = priors[0]
+        for i in range(1, len(priors)):
+            context += priors[i]
+        output = self.conv_bn_dropout(torch.cat([context, feats], 1))
+        return output
+
+
+class BaseOC_Context_Module(nn.Module):
+    """
+    Output only the context features.
+    Parameters:
+        in_features / out_features: the channels of the input / output feature maps.
+        dropout: specify the dropout ratio
+        fusion: We provide two different fusion method, "concat" or "add"
+        size: we find that directly learn the attention weights on even 1/8 feature maps is hard.
+    Return:
+        features after "concat" or "add"
+    """
+
+    def __init__(self, in_channels, out_channels, key_channels, value_channels, dropout, sizes=([1])):
+        super(BaseOC_Context_Module, self).__init__()
+        self.stages = []
+        self.stages = nn.ModuleList(
+            [self._make_stage(in_channels, out_channels, key_channels, value_channels, size) for size in sizes])
+        self.conv_bn_dropout = nn.Sequential(
+            nn.Conv2d(in_channels, out_channels, kernel_size=1, padding=0),
+            InPlaceABNSync(out_channels),
+        )
+
+    def _make_stage(self, in_channels, output_channels, key_channels, value_channels, size):
+        return SelfAttentionBlock2D(in_channels,
+                                    key_channels,
+                                    value_channels,
+                                    output_channels,
+                                    size)
+
+    def forward(self, feats):
+        priors = [stage(feats) for stage in self.stages]
+        context = priors[0]
+        for i in range(1, len(priors)):
+            context += priors[i]
+        output = self.conv_bn_dropout(context)
+        return output
+
+
+class ASP_OC_Module(nn.Module):
+    def __init__(self, features, out_features=256, dilations=(12, 24, 36)):
+        super(ASP_OC_Module, self).__init__()
+        self.context = nn.Sequential(nn.Conv2d(features, out_features, kernel_size=3, padding=1, dilation=1, bias=True),
+                                     InPlaceABNSync(out_features),
+                                     BaseOC_Context_Module(in_channels=out_features, out_channels=out_features,
+                                                           key_channels=out_features // 2, value_channels=out_features,
+                                                           dropout=0, sizes=([2])))
+        self.conv2 = nn.Sequential(nn.Conv2d(features, out_features, kernel_size=1, padding=0, dilation=1, bias=False),
+                                   InPlaceABNSync(out_features))
+        self.conv3 = nn.Sequential(
+            nn.Conv2d(features, out_features, kernel_size=3, padding=dilations[0], dilation=dilations[0], bias=False),
+            InPlaceABNSync(out_features))
+        self.conv4 = nn.Sequential(
+            nn.Conv2d(features, out_features, kernel_size=3, padding=dilations[1], dilation=dilations[1], bias=False),
+            InPlaceABNSync(out_features))
+        self.conv5 = nn.Sequential(
+            nn.Conv2d(features, out_features, kernel_size=3, padding=dilations[2], dilation=dilations[2], bias=False),
+            InPlaceABNSync(out_features))
+
+        self.conv_bn_dropout = nn.Sequential(
+            nn.Conv2d(out_features * 5, out_features, kernel_size=1, padding=0, dilation=1, bias=False),
+            InPlaceABNSync(out_features),
+            nn.Dropout2d(0.1)
+        )
+
+    def _cat_each(self, feat1, feat2, feat3, feat4, feat5):
+        assert (len(feat1) == len(feat2))
+        z = []
+        for i in range(len(feat1)):
+            z.append(torch.cat((feat1[i], feat2[i], feat3[i], feat4[i], feat5[i]), 1))
+        return z
+
+    def forward(self, x):
+        if isinstance(x, Variable):
+            _, _, h, w = x.size()
+        elif isinstance(x, tuple) or isinstance(x, list):
+            _, _, h, w = x[0].size()
+        else:
+            raise RuntimeError('unknown input type')
+
+        feat1 = self.context(x)
+        feat2 = self.conv2(x)
+        feat3 = self.conv3(x)
+        feat4 = self.conv4(x)
+        feat5 = self.conv5(x)
+
+        if isinstance(x, Variable):
+            out = torch.cat((feat1, feat2, feat3, feat4, feat5), 1)
+        elif isinstance(x, tuple) or isinstance(x, list):
+            out = self._cat_each(feat1, feat2, feat3, feat4, feat5)
+        else:
+            raise RuntimeError('unknown input type')
+        output = self.conv_bn_dropout(out)
+        return output

app/models/schp/networks/context_encoding/psp.py

+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+
+"""
+@Author  :   Peike Li
+@Contact :   peike.li@yahoo.com
+@File    :   psp.py
+@Time    :   8/4/19 3:36 PM
+@Desc    :   
+@License :   This source code is licensed under the license found in the 
+             LICENSE file in the root directory of this source tree.
+"""
+
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+
+from modules import InPlaceABNSync
+
+
+class PSPModule(nn.Module):
+    """
+    Reference:
+        Zhao, Hengshuang, et al. *"Pyramid scene parsing network."*
+    """
+    def __init__(self, features, out_features=512, sizes=(1, 2, 3, 6)):
+        super(PSPModule, self).__init__()
+
+        self.stages = []
+        self.stages = nn.ModuleList([self._make_stage(features, out_features, size) for size in sizes])
+        self.bottleneck = nn.Sequential(
+            nn.Conv2d(features + len(sizes) * out_features, out_features, kernel_size=3, padding=1, dilation=1,
+                      bias=False),
+            InPlaceABNSync(out_features),
+        )
+
+    def _make_stage(self, features, out_features, size):
+        prior = nn.AdaptiveAvgPool2d(output_size=(size, size))
+        conv = nn.Conv2d(features, out_features, kernel_size=1, bias=False)
+        bn = InPlaceABNSync(out_features)
+        return nn.Sequential(prior, conv, bn)
+
+    def forward(self, feats):
+        h, w = feats.size(2), feats.size(3)
+        priors = [F.interpolate(input=stage(feats), size=(h, w), mode='bilinear', align_corners=True) for stage in
+                  self.stages] + [feats]
+        bottle = self.bottleneck(torch.cat(priors, 1))
+        return bottle
\ No newline at end of file

app/models/schp/requirements.txt

+opencv-python==4.4.0.46

app/models/schp/simple_extractor.py

+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+
+import os
+from typing import List
+
+import cv2
+import torch
+import numpy as np
+from PIL import Image
+import torchvision.transforms as transforms
+
+from app.models.schp import networks
+from app.models.schp.utils.transforms import get_affine_transform, transform_logits
+
+# 数据集设置
+dataset_settings = {
+    'lip': {
+        'input_size': [473, 473],
+        'num_classes': 20,
+        'label': ['Background', 'Hat', 'Hair', 'Glove', 'Sunglasses', 'Upper-clothes', 'Dress', 'Coat',
+                  'Socks', 'Pants', 'Jumpsuits', 'Scarf', 'Skirt', 'Face', 'Left-arm', 'Right-arm',
+                  'Left-leg', 'Right-leg', 'Left-shoe', 'Right-shoe']
+    },
+    'atr': {
+        'input_size': [512, 512],
+        'num_classes': 18,
+        'label': ['Background', 'Hat', 'Hair', 'Sunglasses', 'Upper-clothes', 'Skirt', 'Pants', 'Dress', 'Belt',
+                  'Left-shoe', 'Right-shoe', 'Face', 'Left-leg', 'Right-leg', 'Left-arm', 'Right-arm', 'Bag', 'Scarf']
+    },
+    'pascal': {
+        'input_size': [512, 512],
+        'num_classes': 7,
+        'label': ['Background', 'Head', 'Torso', 'Upper Arms', 'Lower Arms', 'Upper Legs', 'Lower Legs'],
+    }
+}
+
+
+def get_color_by_label(label: str) -> List[int]:
+    """
+    根据标签名称获取对应的RGB颜色值
+
+    Args:
+        label (str): 标签名称，如 'Face', 'Hair' 等
+
+    Returns:
+        List[int]: RGB颜色值列表，格式为 [R, G, B]，值范围0-255
+    """
+    # LIP数据集标签
+    labels = ['Background', 'Hat', 'Hair', 'Glove', 'Sunglasses', 'Upper-clothes',
+              'Dress', 'Coat', 'Socks', 'Pants', 'Jumpsuits', 'Scarf', 'Skirt',
+              'Face', 'Left-arm', 'Right-arm', 'Left-leg', 'Right-leg', 'Left-shoe', 'Right-shoe']
+
+    # 检查标签是否存在
+    if label not in labels:
+        return []
+
+    # 获取标签索引
+    label_index = labels.index(label)
+
+    # 生成调色板
+    palette = get_palette(len(labels))
+
+    # 获取对应颜色的RGB值并返回列表
+    r = palette[label_index * 3 + 0]
+    g = palette[label_index * 3 + 1]
+    b = palette[label_index * 3 + 2]
+
+    return [r, g, b]
+
+
+def get_palette(num_cls):
+    """返回用于可视化分割掩码的颜色映射"""
+    n = num_cls
+    palette = [0] * (n * 3)
+    for j in range(0, n):
+        lab = j
+        palette[j * 3 + 0] = 0
+        palette[j * 3 + 1] = 0
+        palette[j * 3 + 2] = 0
+        i = 0
+        while lab:
+            palette[j * 3 + 0] |= (((lab >> 0) & 1) << (7 - i))
+            palette[j * 3 + 1] |= (((lab >> 1) & 1) << (7 - i))
+            palette[j * 3 + 2] |= (((lab >> 2) & 1) << (7 - i))
+            i += 1
+            lab >>= 3
+    return palette
+
+
+def _box2cs(box, aspect_ratio):
+    """将边界框转换为中心点和尺度"""
+    x, y, w, h = box[:4]
+    return _xywh2cs(x, y, w, h, aspect_ratio)
+
+
+def _xywh2cs(x, y, w, h, aspect_ratio):
+    """将xywh格式转换为中心点和尺度"""
+    center = np.zeros((2), dtype=np.float32)
+    center[0] = x + w * 0.5
+    center[1] = y + h * 0.5
+    if w > aspect_ratio * h:
+        h = w * 1.0 / aspect_ratio
+    elif w < aspect_ratio * h:
+        w = h * aspect_ratio
+    scale = np.array([w, h], dtype=np.float32)
+    return center, scale
+
+
+class HumanParsingModel:
+    def __init__(self, model_path, dataset='atr', device=None):
+        """
+        初始化人体解析模型
+
+        Args:
+            model_path: 预训练模型路径
+            dataset: 数据集类型 ('lip', 'atr', 'pascal')
+            device: 计算设备 (None表示自动选择)
+        """
+        self.dataset = dataset
+        self.device = device if device else torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+
+        # 获取数据集设置
+        self.num_classes = dataset_settings[dataset]['num_classes']
+        self.input_size = dataset_settings[dataset]['input_size']
+        self.label = dataset_settings[dataset]['label']
+        self.aspect_ratio = self.input_size[1] * 1.0 / self.input_size[0]
+        self.input_size_array = np.asarray(self.input_size)
+
+        # 初始化模型
+        self.model = networks.init_model('resnet101', num_classes=self.num_classes, pretrained=None)
+
+        # 加载预训练权重
+        state_dict = torch.load(model_path, map_location=self.device)['state_dict']
+        from collections import OrderedDict
+        new_state_dict = OrderedDict()
+        for k, v in state_dict.items():
+            name = k[7:] if k.startswith('module.') else k  # 移除 'module.' 前缀
+            new_state_dict[name] = v
+        self.model.load_state_dict(new_state_dict)
+
+        # 将模型移动到指定设备并设置为评估模式
+        self.model.to(self.device)
+        self.model.eval()
+
+        # 图像预处理变换
+        self.transform = transforms.Compose([
+            transforms.ToTensor(),
+            transforms.Normalize(mean=[0.406, 0.456, 0.485], std=[0.225, 0.224, 0.229])
+        ])
+
+        # 获取调色板
+        self.palette = get_palette(self.num_classes)
+
+        print(f"模型已加载，使用设备: {self.device}")
+        print(f"数据集: {dataset}, 类别数: {self.num_classes}")
+
+
+def process_single_image(model, input_image):
+    """
+    处理单张图片
+
+    Args:
+        model: HumanParsingModel实例
+        input_image: 输入图片，可以是：
+                    - numpy数组 (H, W, C) BGR格式
+                    - PIL Image对象
+                    - 图片文件路径字符串
+
+    Returns:
+        PIL Image对象，包含分割结果的彩色图像
+    """
+
+    # 处理不同类型的输入
+    if isinstance(input_image, str):
+        # 如果是文件路径
+        img = cv2.imread(input_image, cv2.IMREAD_COLOR)
+        if img is None:
+            raise ValueError(f"无法读取图片: {input_image}")
+    elif isinstance(input_image, Image.Image):
+        # 如果是PIL Image，转换为BGR numpy数组
+        img = cv2.cvtColor(np.array(input_image), cv2.COLOR_RGB2BGR)
+    elif isinstance(input_image, np.ndarray):
+        # 如果是numpy数组，直接使用
+        img = input_image.copy()
+    else:
+        raise ValueError("输入图片格式不支持，请使用numpy数组、PIL Image或文件路径")
+
+    h, w, _ = img.shape
+
+    # 获取人体中心点和尺度
+    person_center, s = _box2cs([0, 0, w - 1, h - 1], model.aspect_ratio)
+    r = 0
+
+    # 获取仿射变换矩阵
+    trans = get_affine_transform(person_center, s, r, model.input_size_array)
+
+    # 应用仿射变换
+    input_tensor = cv2.warpAffine(
+        img,
+        trans,
+        (int(model.input_size[1]), int(model.input_size[0])),
+        flags=cv2.INTER_LINEAR,
+        borderMode=cv2.BORDER_CONSTANT,
+        borderValue=(0, 0, 0)
+    )
+
+    # 预处理
+    input_tensor = model.transform(input_tensor)
+    input_tensor = input_tensor.unsqueeze(0)  # 添加batch维度
+    input_tensor = input_tensor.to(model.device)
+
+    # 模型推理
+    with torch.no_grad():
+        output = model.model(input_tensor)
+
+        # 上采样到输入尺寸
+        upsample = torch.nn.Upsample(size=model.input_size, mode='bilinear', align_corners=True)
+        upsample_output = upsample(output[0][-1][0].unsqueeze(0))
+        upsample_output = upsample_output.squeeze()
+        upsample_output = upsample_output.permute(1, 2, 0)  # CHW -> HWC
+
+        # 变换回原始图像尺寸
+        logits_result = transform_logits(
+            upsample_output.data.cpu().numpy(),
+            person_center,
+            s,
+            w,
+            h,
+            input_size=model.input_size
+        )
+
+        # 获取分割结果
+        parsing_result = np.argmax(logits_result, axis=2)
+
+        # 转换为PIL图像并应用调色板
+        output_img = Image.fromarray(np.asarray(parsing_result, dtype=np.uint8))
+        output_img.putpalette(model.palette)
+
+        return output_img
+
+
+# 使用示例函数
+def parse_human_image(input_image, model_path, dataset='lip', device=None):
+    """
+    便捷函数：解析单张人体图像
+
+    Args:
+        input_image: 输入图片 (numpy数组、PIL Image或文件路径)
+        model_path: 预训练模型路径
+        dataset: 数据集类型 ('lip', 'atr', 'pascal')
+        device: 计算设备
+
+    Returns:
+        PIL Image对象，包含分割结果
+    """
+    # 创建模型实例
+    model = HumanParsingModel(model_path, dataset, device)
+
+    # 处理图像
+    result = process_single_image(model, input_image)
+
+    return result
+
+
+# 使用示例
+if __name__ == '__main__':
+    # 示例1：使用便捷函数
+    model_path = r"D:\work\PycharmProjects\PythonProject\checkpoints\exp-schp-201908261155-lip.pth"
+    input_image_path = r"D:\work\PycharmProjects\PythonProject\img1.jpg"
+
+    # result_image = parse_human_image(input_image_path, model_path, dataset='atr')
+    # result_image.save(r"D:\work\PycharmProjects\PythonProject\output_result.png")
+
+    # 示例2：使用类的方式（推荐用于批量处理）
+    model = HumanParsingModel(model_path, dataset='lip')
+
+    # 处理多张图片
+    image_paths = [r"D:\work\PycharmProjects\PythonProject\img1.jpg", r"D:\work\PycharmProjects\PythonProject\img2.jpg", r"D:\work\PycharmProjects\PythonProject\img3.jpg"]
+    for i, img_path in enumerate(image_paths):
+        result = process_single_image(model, img_path)
+        result.save(f"result_{i}.png")

app/models/schp/train.py

+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+
+"""
+@Author  :   Peike Li
+@Contact :   peike.li@yahoo.com
+@File    :   train.py
+@Time    :   8/4/19 3:36 PM
+@Desc    :
+@License :   This source code is licensed under the license found in the
+             LICENSE file in the root directory of this source tree.
+"""
+
+import os
+import json
+import timeit
+import argparse
+
+import torch
+import torch.optim as optim
+import torchvision.transforms as transforms
+import torch.backends.cudnn as cudnn
+from torch.utils import data
+
+import networks
+import utils.schp as schp
+from datasets.datasets import LIPDataSet
+from datasets.target_generation import generate_edge_tensor
+from utils.transforms import BGR2RGB_transform
+from utils.criterion import CriterionAll
+from utils.encoding import DataParallelModel, DataParallelCriterion
+from utils.warmup_scheduler import SGDRScheduler
+
+
+def get_arguments():
+    """Parse all the arguments provided from the CLI.
+    Returns:
+      A list of parsed arguments.
+    """
+    parser = argparse.ArgumentParser(description="Self Correction for Human Parsing")
+
+    # Network Structure
+    parser.add_argument("--arch", type=str, default='resnet101')
+    # Data Preference
+    parser.add_argument("--data-dir", type=str, default='./data/LIP')
+    parser.add_argument("--batch-size", type=int, default=16)
+    parser.add_argument("--input-size", type=str, default='473,473')
+    parser.add_argument("--num-classes", type=int, default=20)
+    parser.add_argument("--ignore-label", type=int, default=255)
+    parser.add_argument("--random-mirror", action="store_true")
+    parser.add_argument("--random-scale", action="store_true")
+    # Training Strategy
+    parser.add_argument("--learning-rate", type=float, default=7e-3)
+    parser.add_argument("--momentum", type=float, default=0.9)
+    parser.add_argument("--weight-decay", type=float, default=5e-4)
+    parser.add_argument("--gpu", type=str, default='0,1,2')
+    parser.add_argument("--start-epoch", type=int, default=0)
+    parser.add_argument("--epochs", type=int, default=150)
+    parser.add_argument("--eval-epochs", type=int, default=10)
+    parser.add_argument("--imagenet-pretrain", type=str, default='./pretrain_model/resnet101-imagenet.pth')
+    parser.add_argument("--log-dir", type=str, default='./log')
+    parser.add_argument("--model-restore", type=str, default='./log/checkpoint.pth.tar')
+    parser.add_argument("--schp-start", type=int, default=100, help='schp start epoch')
+    parser.add_argument("--cycle-epochs", type=int, default=10, help='schp cyclical epoch')
+    parser.add_argument("--schp-restore", type=str, default='./log/schp_checkpoint.pth.tar')
+    parser.add_argument("--lambda-s", type=float, default=1, help='segmentation loss weight')
+    parser.add_argument("--lambda-e", type=float, default=1, help='edge loss weight')
+    parser.add_argument("--lambda-c", type=float, default=0.1, help='segmentation-edge consistency loss weight')
+    return parser.parse_args()
+
+
+def main():
+    args = get_arguments()
+    print(args)
+
+    start_epoch = 0
+    cycle_n = 0
+
+    if not os.path.exists(args.log_dir):
+        os.makedirs(args.log_dir)
+    with open(os.path.join(args.log_dir, 'args.json'), 'w') as opt_file:
+        json.dump(vars(args), opt_file)
+
+    gpus = [int(i) for i in args.gpu.split(',')]
+    if not args.gpu == 'None':
+        os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu
+
+    input_size = list(map(int, args.input_size.split(',')))
+
+    cudnn.enabled = True
+    cudnn.benchmark = True
+
+    # Model Initialization
+    AugmentCE2P = networks.init_model(args.arch, num_classes=args.num_classes, pretrained=args.imagenet_pretrain)
+    model = DataParallelModel(AugmentCE2P)
+    model.cuda()
+
+    IMAGE_MEAN = AugmentCE2P.mean
+    IMAGE_STD = AugmentCE2P.std
+    INPUT_SPACE = AugmentCE2P.input_space
+    print('image mean: {}'.format(IMAGE_MEAN))
+    print('image std: {}'.format(IMAGE_STD))
+    print('input space:{}'.format(INPUT_SPACE))
+
+    restore_from = args.model_restore
+    if os.path.exists(restore_from):
+        print('Resume training from {}'.format(restore_from))
+        checkpoint = torch.load(restore_from)
+        model.load_state_dict(checkpoint['state_dict'])
+        start_epoch = checkpoint['epoch']
+
+    SCHP_AugmentCE2P = networks.init_model(args.arch, num_classes=args.num_classes, pretrained=args.imagenet_pretrain)
+    schp_model = DataParallelModel(SCHP_AugmentCE2P)
+    schp_model.cuda()
+
+    if os.path.exists(args.schp_restore):
+        print('Resuming schp checkpoint from {}'.format(args.schp_restore))
+        schp_checkpoint = torch.load(args.schp_restore)
+        schp_model_state_dict = schp_checkpoint['state_dict']
+        cycle_n = schp_checkpoint['cycle_n']
+        schp_model.load_state_dict(schp_model_state_dict)
+
+    # Loss Function
+    criterion = CriterionAll(lambda_1=args.lambda_s, lambda_2=args.lambda_e, lambda_3=args.lambda_c,
+                             num_classes=args.num_classes)
+    criterion = DataParallelCriterion(criterion)
+    criterion.cuda()
+
+    # Data Loader
+    if INPUT_SPACE == 'BGR':
+        print('BGR Transformation')
+        transform = transforms.Compose([
+            transforms.ToTensor(),
+            transforms.Normalize(mean=IMAGE_MEAN,
+                                 std=IMAGE_STD),
+        ])
+
+    elif INPUT_SPACE == 'RGB':
+        print('RGB Transformation')
+        transform = transforms.Compose([
+            transforms.ToTensor(),
+            BGR2RGB_transform(),
+            transforms.Normalize(mean=IMAGE_MEAN,
+                                 std=IMAGE_STD),
+        ])
+
+    train_dataset = LIPDataSet(args.data_dir, 'train', crop_size=input_size, transform=transform)
+    train_loader = data.DataLoader(train_dataset, batch_size=args.batch_size * len(gpus),
+                                   num_workers=16, shuffle=True, pin_memory=True, drop_last=True)
+    print('Total training samples: {}'.format(len(train_dataset)))
+
+    # Optimizer Initialization
+    optimizer = optim.SGD(model.parameters(), lr=args.learning_rate, momentum=args.momentum,
+                          weight_decay=args.weight_decay)
+
+    lr_scheduler = SGDRScheduler(optimizer, total_epoch=args.epochs,
+                                 eta_min=args.learning_rate / 100, warmup_epoch=10,
+                                 start_cyclical=args.schp_start, cyclical_base_lr=args.learning_rate / 2,
+                                 cyclical_epoch=args.cycle_epochs)
+
+    total_iters = args.epochs * len(train_loader)
+    start = timeit.default_timer()
+    for epoch in range(start_epoch, args.epochs):
+        lr_scheduler.step(epoch=epoch)
+        lr = lr_scheduler.get_lr()[0]
+
+        model.train()
+        for i_iter, batch in enumerate(train_loader):
+            i_iter += len(train_loader) * epoch
+
+            images, labels, _ = batch
+            labels = labels.cuda(non_blocking=True)
+
+            edges = generate_edge_tensor(labels)
+            labels = labels.type(torch.cuda.LongTensor)
+            edges = edges.type(torch.cuda.LongTensor)
+
+            preds = model(images)
+
+            # Online Self Correction Cycle with Label Refinement
+            if cycle_n >= 1:
+                with torch.no_grad():
+                    soft_preds = schp_model(images)
+                    soft_parsing = []
+                    soft_edge = []
+                    for soft_pred in soft_preds:
+                        soft_parsing.append(soft_pred[0][-1])
+                        soft_edge.append(soft_pred[1][-1])
+                    soft_preds = torch.cat(soft_parsing, dim=0)
+                    soft_edges = torch.cat(soft_edge, dim=0)
+            else:
+                soft_preds = None
+                soft_edges = None
+
+            loss = criterion(preds, [labels, edges, soft_preds, soft_edges], cycle_n)
+
+            optimizer.zero_grad()
+            loss.backward()
+            optimizer.step()
+
+            if i_iter % 100 == 0:
+                print('iter = {} of {} completed, lr = {}, loss = {}'.format(i_iter, total_iters, lr,
+                                                                             loss.data.cpu().numpy()))
+        if (epoch + 1) % (args.eval_epochs) == 0:
+            schp.save_schp_checkpoint({
+                'epoch': epoch + 1,
+                'state_dict': model.state_dict(),
+            }, False, args.log_dir, filename='checkpoint_{}.pth.tar'.format(epoch + 1))
+
+        # Self Correction Cycle with Model Aggregation
+        if (epoch + 1) >= args.schp_start and (epoch + 1 - args.schp_start) % args.cycle_epochs == 0:
+            print('Self-correction cycle number {}'.format(cycle_n))
+            schp.moving_average(schp_model, model, 1.0 / (cycle_n + 1))
+            cycle_n += 1
+            schp.bn_re_estimate(train_loader, schp_model)
+            schp.save_schp_checkpoint({
+                'state_dict': schp_model.state_dict(),
+                'cycle_n': cycle_n,
+            }, False, args.log_dir, filename='schp_{}_checkpoint.pth.tar'.format(cycle_n))
+
+        torch.cuda.empty_cache()
+        end = timeit.default_timer()
+        print('epoch = {} of {} completed using {} s'.format(epoch, args.epochs,
+                                                             (end - start) / (epoch - start_epoch + 1)))
+
+    end = timeit.default_timer()
+    print('Training Finished in {} seconds'.format(end - start))
+
+
+if __name__ == '__main__':
+    main()

app/models/schp/utils/consistency_loss.py

+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+
+"""
+@Author  :   Peike Li
+@Contact :   peike.li@yahoo.com
+@File    :   kl_loss.py
+@Time    :   7/23/19 4:02 PM
+@Desc    :   
+@License :   This source code is licensed under the license found in the 
+             LICENSE file in the root directory of this source tree.
+"""
+import torch
+import torch.nn.functional as F
+from torch import nn
+from datasets.target_generation import generate_edge_tensor
+
+
+class ConsistencyLoss(nn.Module):
+    def __init__(self, ignore_index=255):
+        super(ConsistencyLoss, self).__init__()
+        self.ignore_index=ignore_index
+
+    def forward(self, parsing, edge, label):
+        parsing_pre = torch.argmax(parsing, dim=1)
+        parsing_pre[label==self.ignore_index]=self.ignore_index
+        generated_edge = generate_edge_tensor(parsing_pre)
+        edge_pre = torch.argmax(edge, dim=1)
+        v_generate_edge = generated_edge[label!=255]
+        v_edge_pre = edge_pre[label!=255]
+        v_edge_pre = v_edge_pre.type(torch.cuda.FloatTensor)
+        positive_union = (v_generate_edge==1)&(v_edge_pre==1) # only the positive values count
+        return F.smooth_l1_loss(v_generate_edge[positive_union].squeeze(0), v_edge_pre[positive_union].squeeze(0))

app/models/schp/utils/criterion.py

+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+
+"""
+@Author  :   Peike Li
+@Contact :   peike.li@yahoo.com
+@File    :   criterion.py
+@Time    :   8/30/19 8:59 PM
+@Desc    :
+@License :   This source code is licensed under the license found in the
+             LICENSE file in the root directory of this source tree.
+"""
+
+import torch.nn as nn
+import torch
+import numpy as np
+from torch.nn import functional as F
+from .lovasz_softmax import LovaszSoftmax
+from .kl_loss import KLDivergenceLoss
+from .consistency_loss import ConsistencyLoss
+
+NUM_CLASSES = 20
+
+
+class CriterionAll(nn.Module):
+    def __init__(self, use_class_weight=False, ignore_index=255, lambda_1=1, lambda_2=1, lambda_3=1,
+                 num_classes=20):
+        super(CriterionAll, self).__init__()
+        self.ignore_index = ignore_index
+        self.use_class_weight = use_class_weight
+        self.criterion = torch.nn.CrossEntropyLoss(ignore_index=ignore_index)
+        self.lovasz = LovaszSoftmax(ignore_index=ignore_index)
+        self.kldiv = KLDivergenceLoss(ignore_index=ignore_index)
+        self.reg = ConsistencyLoss(ignore_index=ignore_index)
+        self.lamda_1 = lambda_1
+        self.lamda_2 = lambda_2
+        self.lamda_3 = lambda_3
+        self.num_classes = num_classes
+
+    def parsing_loss(self, preds, target, cycle_n=None):
+        """
+        Loss function definition.
+
+        Args:
+            preds: [[parsing result1, parsing result2],[edge result]]
+            target: [parsing label, egde label]
+            soft_preds: [[parsing result1, parsing result2],[edge result]]
+        Returns:
+            Calculated Loss.
+        """
+        h, w = target[0].size(1), target[0].size(2)
+
+        pos_num = torch.sum(target[1] == 1, dtype=torch.float)
+        neg_num = torch.sum(target[1] == 0, dtype=torch.float)
+
+        weight_pos = neg_num / (pos_num + neg_num)
+        weight_neg = pos_num / (pos_num + neg_num)
+        weights = torch.tensor([weight_neg, weight_pos])  # edge loss weight
+
+        loss = 0
+
+        # loss for segmentation
+        preds_parsing = preds[0]
+        for pred_parsing in preds_parsing:
+            scale_pred = F.interpolate(input=pred_parsing, size=(h, w),
+                                       mode='bilinear', align_corners=True)
+
+            loss += 0.5 * self.lamda_1 * self.lovasz(scale_pred, target[0])
+            if target[2] is None:
+                loss += 0.5 * self.lamda_1 * self.criterion(scale_pred, target[0])
+            else:
+                soft_scale_pred = F.interpolate(input=target[2], size=(h, w),
+                                                mode='bilinear', align_corners=True)
+                soft_scale_pred = moving_average(soft_scale_pred, to_one_hot(target[0], num_cls=self.num_classes),
+                                                 1.0 / (cycle_n + 1.0))
+                loss += 0.5 * self.lamda_1 * self.kldiv(scale_pred, soft_scale_pred, target[0])
+
+        # loss for edge
+        preds_edge = preds[1]
+        for pred_edge in preds_edge:
+            scale_pred = F.interpolate(input=pred_edge, size=(h, w),
+                                       mode='bilinear', align_corners=True)
+            if target[3] is None:
+                loss += self.lamda_2 * F.cross_entropy(scale_pred, target[1],
+                                                       weights.cuda(), ignore_index=self.ignore_index)
+            else:
+                soft_scale_edge = F.interpolate(input=target[3], size=(h, w),
+                                                mode='bilinear', align_corners=True)
+                soft_scale_edge = moving_average(soft_scale_edge, to_one_hot(target[1], num_cls=2),
+                                                 1.0 / (cycle_n + 1.0))
+                loss += self.lamda_2 * self.kldiv(scale_pred, soft_scale_edge, target[0])
+
+        # consistency regularization
+        preds_parsing = preds[0]
+        preds_edge = preds[1]
+        for pred_parsing in preds_parsing:
+            scale_pred = F.interpolate(input=pred_parsing, size=(h, w),
+                                       mode='bilinear', align_corners=True)
+            scale_edge = F.interpolate(input=preds_edge[0], size=(h, w),
+                                       mode='bilinear', align_corners=True)
+            loss += self.lamda_3 * self.reg(scale_pred, scale_edge, target[0])
+
+        return loss
+
+    def forward(self, preds, target, cycle_n=None):
+        loss = self.parsing_loss(preds, target, cycle_n)
+        return loss
+
+    def _generate_weights(self, masks, num_classes):
+        """
+        masks: torch.Tensor with shape [B, H, W]
+        """
+        masks_label = masks.data.cpu().numpy().astype(np.int64)
+        pixel_nums = []
+        tot_pixels = 0
+        for i in range(num_classes):
+            pixel_num_of_cls_i = np.sum(masks_label == i).astype(np.float)
+            pixel_nums.append(pixel_num_of_cls_i)
+            tot_pixels += pixel_num_of_cls_i
+        weights = []
+        for i in range(num_classes):
+            weights.append(
+                (tot_pixels - pixel_nums[i]) / tot_pixels / (num_classes - 1)
+            )
+        weights = np.array(weights, dtype=np.float)
+        # weights = torch.from_numpy(weights).float().to(masks.device)
+        return weights
+
+
+def moving_average(target1, target2, alpha=1.0):
+    target = 0
+    target += (1.0 - alpha) * target1
+    target += target2 * alpha
+    return target
+
+
+def to_one_hot(tensor, num_cls, dim=1, ignore_index=255):
+    b, h, w = tensor.shape
+    tensor[tensor == ignore_index] = 0
+    onehot_tensor = torch.zeros(b, num_cls, h, w).cuda()
+    onehot_tensor.scatter_(dim, tensor.unsqueeze(dim), 1)
+    return onehot_tensor

app/models/schp/utils/encoding.py

+##+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
+## Created by: Hang Zhang
+## ECE Department, Rutgers University
+## Email: zhang.hang@rutgers.edu
+## Copyright (c) 2017
+##
+## This source code is licensed under the MIT-style license found in the
+## LICENSE file in the root directory of this source tree
+##+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
+
+"""Encoding Data Parallel"""
+import threading
+import functools
+import torch
+from torch.autograd import Variable, Function
+import torch.cuda.comm as comm
+from torch.nn.parallel.data_parallel import DataParallel
+from torch.nn.parallel.parallel_apply import get_a_var
+from torch.nn.parallel._functions import ReduceAddCoalesced, Broadcast
+
+torch_ver = torch.__version__[:3]
+
+__all__ = ['allreduce', 'DataParallelModel', 'DataParallelCriterion', 'patch_replication_callback']
+
+def allreduce(*inputs):
+    """Cross GPU all reduce autograd operation for calculate mean and
+    variance in SyncBN.
+    """
+    return AllReduce.apply(*inputs)
+
+class AllReduce(Function):
+    @staticmethod
+    def forward(ctx, num_inputs, *inputs):
+        ctx.num_inputs = num_inputs
+        ctx.target_gpus = [inputs[i].get_device() for i in range(0, len(inputs), num_inputs)]
+        inputs = [inputs[i:i + num_inputs]
+                 for i in range(0, len(inputs), num_inputs)]
+        # sort before reduce sum
+        inputs = sorted(inputs, key=lambda i: i[0].get_device())
+        results = comm.reduce_add_coalesced(inputs, ctx.target_gpus[0])
+        outputs = comm.broadcast_coalesced(results, ctx.target_gpus)
+        return tuple([t for tensors in outputs for t in tensors])
+
+    @staticmethod
+    def backward(ctx, *inputs):
+        inputs = [i.data for i in inputs]
+        inputs = [inputs[i:i + ctx.num_inputs]
+                 for i in range(0, len(inputs), ctx.num_inputs)]
+        results = comm.reduce_add_coalesced(inputs, ctx.target_gpus[0])
+        outputs = comm.broadcast_coalesced(results, ctx.target_gpus)
+        return (None,) + tuple([Variable(t) for tensors in outputs for t in tensors])
+
+class Reduce(Function):
+    @staticmethod
+    def forward(ctx, *inputs):
+        ctx.target_gpus = [inputs[i].get_device() for i in range(len(inputs))]
+        inputs = sorted(inputs, key=lambda i: i.get_device())
+        return comm.reduce_add(inputs)
+
+    @staticmethod
+    def backward(ctx, gradOutput):
+        return Broadcast.apply(ctx.target_gpus, gradOutput)
+
+
+class DataParallelModel(DataParallel):
+    """Implements data parallelism at the module level.
+
+    This container parallelizes the application of the given module by
+    splitting the input across the specified devices by chunking in the
+    batch dimension.
+    In the forward pass, the module is replicated on each device,
+    and each replica handles a portion of the input. During the backwards pass, gradients from each replica are summed into the original module.
+    Note that the outputs are not gathered, please use compatible
+    :class:`encoding.parallel.DataParallelCriterion`.
+
+    The batch size should be larger than the number of GPUs used. It should
+    also be an integer multiple of the number of GPUs so that each chunk is
+    the same size (so that each GPU processes the same number of samples).
+
+    Args:
+        module: module to be parallelized
+        device_ids: CUDA devices (default: all devices)
+
+    Reference:
+        Hang Zhang, Kristin Dana, Jianping Shi, Zhongyue Zhang, Xiaogang Wang, Ambrish Tyagi,
+        Amit Agrawal. “Context Encoding for Semantic Segmentation.
+        *The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2018*
+
+    Example::
+
+        >>> net = encoding.nn.DataParallelModel(model, device_ids=[0, 1, 2])
+        >>> y = net(x)
+    """
+    def gather(self, outputs, output_device):
+        return outputs
+
+    def replicate(self, module, device_ids):
+        modules = super(DataParallelModel, self).replicate(module, device_ids)
+        return modules
+
+
+class DataParallelCriterion(DataParallel):
+    """
+    Calculate loss in multiple-GPUs, which balance the memory usage for
+    Semantic Segmentation.
+
+    The targets are splitted across the specified devices by chunking in
+    the batch dimension. Please use together with :class:`encoding.parallel.DataParallelModel`.
+
+    Reference:
+        Hang Zhang, Kristin Dana, Jianping Shi, Zhongyue Zhang, Xiaogang Wang, Ambrish Tyagi,
+        Amit Agrawal. “Context Encoding for Semantic Segmentation.
+        *The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2018*
+
+    Example::
+
+        >>> net = encoding.nn.DataParallelModel(model, device_ids=[0, 1, 2])
+        >>> criterion = encoding.nn.DataParallelCriterion(criterion, device_ids=[0, 1, 2])
+        >>> y = net(x)
+        >>> loss = criterion(y, target)
+    """
+    def forward(self, inputs, *targets, **kwargs):
+        # input should be already scatterd
+        # scattering the targets instead
+        if not self.device_ids:
+            return self.module(inputs, *targets, **kwargs)
+        targets, kwargs = self.scatter(targets, kwargs, self.device_ids)
+        if len(self.device_ids) == 1:
+            return self.module(inputs, *targets[0], **kwargs[0])
+        replicas = self.replicate(self.module, self.device_ids[:len(inputs)])
+        outputs = _criterion_parallel_apply(replicas, inputs, targets, kwargs)
+        return Reduce.apply(*outputs) / len(outputs)
+
+
+def _criterion_parallel_apply(modules, inputs, targets, kwargs_tup=None, devices=None):
+    assert len(modules) == len(inputs)
+    assert len(targets) == len(inputs)
+    if kwargs_tup:
+        assert len(modules) == len(kwargs_tup)
+    else:
+        kwargs_tup = ({},) * len(modules)
+    if devices is not None:
+        assert len(modules) == len(devices)
+    else:
+        devices = [None] * len(modules)
+
+    lock = threading.Lock()
+    results = {}
+    if torch_ver != "0.3":
+        grad_enabled = torch.is_grad_enabled()
+
+    def _worker(i, module, input, target, kwargs, device=None):
+        if torch_ver != "0.3":
+            torch.set_grad_enabled(grad_enabled)
+        if device is None:
+            device = get_a_var(input).get_device()
+        try:
+            if not isinstance(input, tuple):
+                input = (input,)
+            with torch.cuda.device(device):
+                output = module(*(input + target), **kwargs)
+            with lock:
+                results[i] = output
+        except Exception as e:
+            with lock:
+                results[i] = e
+
+    if len(modules) > 1:
+        threads = [threading.Thread(target=_worker,
+                                    args=(i, module, input, target,
+                                          kwargs, device),)
+                   for i, (module, input, target, kwargs, device) in
+                   enumerate(zip(modules, inputs, targets, kwargs_tup, devices))]
+
+        for thread in threads:
+            thread.start()
+        for thread in threads:
+            thread.join()
+    else:
+        _worker(0, modules[0], inputs[0], kwargs_tup[0], devices[0])
+
+    outputs = []
+    for i in range(len(inputs)):
+        output = results[i]
+        if isinstance(output, Exception):
+            raise output
+        outputs.append(output)
+    return outputs

app/models/schp/utils/kl_loss.py

+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+
+"""
+@Author  :   Peike Li
+@Contact :   peike.li@yahoo.com
+@File    :   kl_loss.py
+@Time    :   7/23/19 4:02 PM
+@Desc    :   
+@License :   This source code is licensed under the license found in the 
+             LICENSE file in the root directory of this source tree.
+"""
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+
+def flatten_probas(input, target, labels, ignore=255):
+    """
+    Flattens predictions in the batch.
+    """
+    B, C, H, W = input.size()
+    input = input.permute(0, 2, 3, 1).contiguous().view(-1, C)  # B * H * W, C = P, C
+    target = target.permute(0, 2, 3, 1).contiguous().view(-1, C)  # B * H * W, C = P, C
+    labels = labels.view(-1)
+    if ignore is None:
+        return input, target
+    valid = (labels != ignore)
+    vinput = input[valid.nonzero().squeeze()]
+    vtarget = target[valid.nonzero().squeeze()]
+    return vinput, vtarget
+
+
+class KLDivergenceLoss(nn.Module):
+    def __init__(self, ignore_index=255, T=1):
+        super(KLDivergenceLoss, self).__init__()
+        self.ignore_index=ignore_index
+        self.T = T
+
+    def forward(self, input, target, label):
+        log_input_prob = F.log_softmax(input / self.T, dim=1)
+        target_porb = F.softmax(target / self.T, dim=1)
+        loss = F.kl_div(*flatten_probas(log_input_prob, target_porb, label, ignore=self.ignore_index))
+        return self.T*self.T*loss # balanced

app/models/schp/utils/miou.py

+import cv2
+import os
+import numpy as np
+
+from collections import OrderedDict
+from PIL import Image as PILImage
+from utils.transforms import transform_parsing
+
+LABELS = ['Background', 'Hat', 'Hair', 'Glove', 'Sunglasses', 'Upper-clothes', 'Dress', 'Coat', \
+          'Socks', 'Pants', 'Jumpsuits', 'Scarf', 'Skirt', 'Face', 'Left-arm', 'Right-arm', 'Left-leg',
+          'Right-leg', 'Left-shoe', 'Right-shoe']
+
+
+# LABELS = ['Background', 'Head', 'Torso', 'Upper Arms', 'Lower Arms', 'Upper Legs', 'Lower Legs']
+
+def get_palette(num_cls):
+    """ Returns the color map for visualizing the segmentation mask.
+    Args:
+        num_cls: Number of classes
+    Returns:
+        The color map
+    """
+
+    n = num_cls
+    palette = [0] * (n * 3)
+    for j in range(0, n):
+        lab = j
+        palette[j * 3 + 0] = 0
+        palette[j * 3 + 1] = 0
+        palette[j * 3 + 2] = 0
+        i = 0
+        while lab:
+            palette[j * 3 + 0] |= (((lab >> 0) & 1) << (7 - i))
+            palette[j * 3 + 1] |= (((lab >> 1) & 1) << (7 - i))
+            palette[j * 3 + 2] |= (((lab >> 2) & 1) << (7 - i))
+            i += 1
+            lab >>= 3
+    return palette
+
+
+def get_confusion_matrix(gt_label, pred_label, num_classes):
+    """
+    Calcute the confusion matrix by given label and pred
+    :param gt_label: the ground truth label
+    :param pred_label: the pred label
+    :param num_classes: the nunber of class
+    :return: the confusion matrix
+    """
+    index = (gt_label * num_classes + pred_label).astype('int32')
+    label_count = np.bincount(index)
+    confusion_matrix = np.zeros((num_classes, num_classes))
+
+    for i_label in range(num_classes):
+        for i_pred_label in range(num_classes):
+            cur_index = i_label * num_classes + i_pred_label
+            if cur_index < len(label_count):
+                confusion_matrix[i_label, i_pred_label] = label_count[cur_index]
+
+    return confusion_matrix
+
+
+def compute_mean_ioU(preds, scales, centers, num_classes, datadir, input_size=[473, 473], dataset='val'):
+    val_file = os.path.join(datadir, dataset + '_id.txt')
+    val_id = [i_id.strip() for i_id in open(val_file)]
+
+    confusion_matrix = np.zeros((num_classes, num_classes))
+
+    for i, pred_out in enumerate(preds):
+        im_name = val_id[i]
+        gt_path = os.path.join(datadir, dataset + '_segmentations', im_name + '.png')
+        gt = np.array(PILImage.open(gt_path))
+        h, w = gt.shape
+        s = scales[i]
+        c = centers[i]
+        pred = transform_parsing(pred_out, c, s, w, h, input_size)
+
+        gt = np.asarray(gt, dtype=np.int32)
+        pred = np.asarray(pred, dtype=np.int32)
+
+        ignore_index = gt != 255
+
+        gt = gt[ignore_index]
+        pred = pred[ignore_index]
+
+        confusion_matrix += get_confusion_matrix(gt, pred, num_classes)
+
+    pos = confusion_matrix.sum(1)
+    res = confusion_matrix.sum(0)
+    tp = np.diag(confusion_matrix)
+
+    pixel_accuracy = (tp.sum() / pos.sum()) * 100
+    mean_accuracy = ((tp / np.maximum(1.0, pos)).mean()) * 100
+    IoU_array = (tp / np.maximum(1.0, pos + res - tp))
+    IoU_array = IoU_array * 100
+    mean_IoU = IoU_array.mean()
+    print('Pixel accuracy: %f \n' % pixel_accuracy)
+    print('Mean accuracy: %f \n' % mean_accuracy)
+    print('Mean IU: %f \n' % mean_IoU)
+    name_value = []
+
+    for i, (label, iou) in enumerate(zip(LABELS, IoU_array)):
+        name_value.append((label, iou))
+
+    name_value.append(('Pixel accuracy', pixel_accuracy))
+    name_value.append(('Mean accuracy', mean_accuracy))
+    name_value.append(('Mean IU', mean_IoU))
+    name_value = OrderedDict(name_value)
+    return name_value
+
+
+def compute_mean_ioU_file(preds_dir, num_classes, datadir, dataset='val'):
+    list_path = os.path.join(datadir, dataset + '_id.txt')
+    val_id = [i_id.strip() for i_id in open(list_path)]
+
+    confusion_matrix = np.zeros((num_classes, num_classes))
+
+    for i, im_name in enumerate(val_id):
+        gt_path = os.path.join(datadir, 'segmentations', im_name + '.png')
+        gt = cv2.imread(gt_path, cv2.IMREAD_GRAYSCALE)
+
+        pred_path = os.path.join(preds_dir, im_name + '.png')
+        pred = np.asarray(PILImage.open(pred_path))
+
+        gt = np.asarray(gt, dtype=np.int32)
+        pred = np.asarray(pred, dtype=np.int32)
+
+        ignore_index = gt != 255
+
+        gt = gt[ignore_index]
+        pred = pred[ignore_index]
+
+        confusion_matrix += get_confusion_matrix(gt, pred, num_classes)
+
+    pos = confusion_matrix.sum(1)
+    res = confusion_matrix.sum(0)
+    tp = np.diag(confusion_matrix)
+
+    pixel_accuracy = (tp.sum() / pos.sum()) * 100
+    mean_accuracy = ((tp / np.maximum(1.0, pos)).mean()) * 100
+    IoU_array = (tp / np.maximum(1.0, pos + res - tp))
+    IoU_array = IoU_array * 100
+    mean_IoU = IoU_array.mean()
+    print('Pixel accuracy: %f \n' % pixel_accuracy)
+    print('Mean accuracy: %f \n' % mean_accuracy)
+    print('Mean IU: %f \n' % mean_IoU)
+    name_value = []
+
+    for i, (label, iou) in enumerate(zip(LABELS, IoU_array)):
+        name_value.append((label, iou))
+
+    name_value.append(('Pixel accuracy', pixel_accuracy))
+    name_value.append(('Mean accuracy', mean_accuracy))
+    name_value.append(('Mean IU', mean_IoU))
+    name_value = OrderedDict(name_value)
+    return name_value

app/services/feature_extractor.py

+from pathlib import Path
+
 import numpy as np
 import torch
 import clip
@@ -6,6 +8,10 @@ from PIL import Image
 from io import BytesIO
 import logging
 
+from app.models.schp.mask import extract_color_region_simple
+from app.models.schp.simple_extractor import HumanParsingModel, process_single_image, get_color_by_label
+
+
 class FeatureExtractor:
     __logger = logging.getLogger(__name__)
 
@@ -14,6 +20,8 @@ class FeatureExtractor:
         device = "cpu"
         self.model, self.preprocess = self.init_model(device, model_name)
         self.device = device
+        model_path = Path(__file__).parent.absolute() / "../models/schp/checkpoints/exp-schp-201908261155-lip.pth"
+        self.schp_model = HumanParsingModel(model_path, dataset='lip')
 
     @staticmethod
     def init_model(device="xpu" if torch.xpu.is_available() else "cpu", model_name="ViT-B/32"):
@@ -52,7 +60,7 @@ class FeatureExtractor:
 
         return new_img
 
-    def extract_from_url(self, image_url):
+    def extract_from_url(self, image_url, part):
         """
         从URL加载图像并提取特征向量
 
@@ -71,7 +79,7 @@ class FeatureExtractor:
             # 将图片数据转换为 PIL Image 对象
             image = Image.open(BytesIO(response.content)).convert("RGB")
 
-            return self.extract_from_image(image)
+            return self.extract_from_image(image, part)
 
         except requests.RequestException as e:
             self.__logger.error(f"Network error when downloading image from {image_url}: {e}")
@@ -80,7 +88,7 @@ class FeatureExtractor:
             self.__logger.error(f"Error extracting features from URL {image_url}: {e}")
             return None
 
-    def extract_from_image(self, img):
+    def extract_from_image(self, img, part):
         """
         从PIL图像对象提取特征向量
 
@@ -100,6 +108,12 @@ class FeatureExtractor:
         # model_name = "ViT-L/14@336px"
         # model, preprocess = self.init_model(device, model_name)
 
+        color = get_color_by_label(part)
+
+        if color:
+            mask_img = process_single_image(self.schp_model, img)
+            img = extract_color_region_simple(img, mask_img, color)
+
         try:
             # 调整图像大小并添加填充
             image = self.resize_with_padding(img)