format all

.gitignore

@@ -15,4 +15,5 @@ encoding.egg-info/
 .*
 /encoding/lib/*/*.so
 /encoding/lib/*/*.o
-*.ninja
+*.ninja
+/tmp

Fix_Bugs_of_Pytorch_Encoding.md

@@ -0,0 +1,24 @@
+## Pytorch-Encoding
+
+After installing prerequisites [as this instruction](Install_Prerequisite_for_Pytorch_Encoding.md), we clone Pytorch-Encoding
+```bash
+# Tested with commit ce461da
+git clone https://github.com/zhanghang1989/PyTorch-Encoding.git
+```
+
+Then, according to [this issue](https://github.com/zhanghang1989/PyTorch-Encoding/issues/161), replace `#include <torch/extension.h>` with `#include <torch/serialize/tensor.h>` in all `encoding/lib/*/*.cpp` and `encoding/lib/*/*.cu` files. Also add `#include <torch/serialize/tensor.h>` to `encoding/lib/gpu/operator.h`.
+
+The Pytorch-Encoding doc requires `python setup.py install`. However, it's not necessary. You can just add the package path to `$PYTHONPATH` or `sys.path`.
+
+Run the examples
+```bash
+export CUDA_HOME=/mnt/data-1/data/houjing.huang/Software/cuda-9.0
+export LD_LIBRARY_PATH=${CUDA_HOME}/lib64:${LD_LIBRARY_PATH}
+
+python scripts/prepare_pcontext.py
+python test.py --dataset PContext --model-zoo Encnet_ResNet50_PContext --eval
+```
+
+## Misc
+
+- The [default ResNet](https://github.com/zhanghang1989/PyTorch-Encoding/blob/ce461dae8d088253dcd9818d2999d4049bce3493/encoding/models/resnet.py) in Pytorch-Encoding is different than pytorch [torchvision](https://github.com/pytorch/vision/blob/98ca260bc834ec94a8143e4b5cfe9516b0b951a2/torchvision/models/resnet.py). The former provides `deep_base`, `dilation`, `multi_grid` options.

Install_Prerequisite_for_Pytorch_Encoding.md

@@ -0,0 +1,122 @@
+We need
+- pytorch
+- cuda
+- ninja
+
+# CUDA 9.0
+
+You can install [CUDA9.0](https://developer.nvidia.com/cuda-90-download-archive), [cuDNN](https://developer.nvidia.com/cudnn), [NCCL](https://developer.nvidia.com/nccl) to local directory. 
+
+First install CUDA by running the downloaded executable.
+
+Then download and extract cuDNN.
+
+```
+cudnn-9.0-linux-x64-v7.4.2.24
+|-- include
+|   `-- cudnn.h
+|-- lib64
+|   |-- libcudnn.so -> libcudnn.so.7
+|   |-- libcudnn.so.7 -> libcudnn.so.7.4.2
+|   |-- libcudnn.so.7.4.2
+|   `-- libcudnn_static.a
+`-- NVIDIA_SLA_cuDNN_Support.txt
+```
+
+Move cuDNN files to CUDA directory by
+```bash
+cp cudnn-9.0-linux-x64-v7.4.2.24/include/* cuda-9.0/include/
+cp cudnn-9.0-linux-x64-v7.4.2.24/lib64/* cuda-9.0/lib64/
+```
+
+Then download and extract NCCL.
+
+```
+nccl_2.3.7-1+cuda9.0_x86_64
+|-- include
+|   `-- nccl.h
+|-- lib
+|   |-- libnccl.so -> libnccl.so.2
+|   |-- libnccl.so.2 -> libnccl.so.2.3.7
+|   |-- libnccl.so.2.3.7
+|   `-- libnccl_static.a
+`-- LICENSE.txt
+```
+
+Move NCCL files to CUDA directory by
+```bash
+cp nccl_2.3.7-1+cuda9.0_x86_64/include/* cuda-9.0/include/
+cp nccl_2.3.7-1+cuda9.0_x86_64/lib/* cuda-9.0/lib64/
+```
+
+# Pytorch 1.0
+
+It's required to install pytorch from source.
+
+Setup your python environment before installing pytorch. Anaconda is required in the following example. Change `WORKING_DIR`, `CUDA_HOME` to your paths and run the following commands.
+
+```bash
+WORKING_DIR=<your-directory-to-save-intermediate-results-of-installing-pytorch>
+TORCH_DIR_NAME=pytorch_v1.0.0
+mkdir -p ${WORKING_DIR}
+
+pip uninstall -y torch
+
+# Install basic dependencies
+conda install --yes numpy pyyaml mkl mkl-include setuptools cmake cffi typing
+conda install --yes -c mingfeima mkldnn
+# Add LAPACK support for the GPU
+conda install --yes -c pytorch magma-cuda90
+
+cd ${WORKING_DIR}
+rm -rf ${TORCH_DIR_NAME}
+# Tested with commit db5d313
+git clone --recursive --single-branch --branch v1.0.0  https://github.com/pytorch/pytorch.git ${TORCH_DIR_NAME}
+cd ${TORCH_DIR_NAME}
+
+rm -rf build
+rm -rf torch.egg-info
+export CUDA_HOME=<your-cuda-directory>
+export USE_SYSTEM_NCCL=1
+export NCCL_LIB_DIR=${CUDA_HOME}/lib64  # For CUDA version > 8.0, you have to download NCCL lib independently
+export NCCL_INCLUDE_DIR=${CUDA_HOME}/include
+export CMAKE_PREFIX_PATH="$(dirname $(which conda))/../" # [anaconda root directory]
+
+python setup.py install 2>&1 | tee ${WORKING_DIR}/install_pytorch.log
+cd ${WORKING_DIR}
+python test_data_parallel.py 2>&1 | tee test_pytorch_data_parallel.log
+```
+
+The contents of `test_data_parallel.py` in above commands is
+
+```python
+import torch
+import torch.nn as nn
+from torch.nn.parallel import DataParallel
+
+model = nn.Linear(10, 20).cuda()
+x = torch.ones(100, 10).float().cuda()
+model_w = DataParallel(model, device_ids=[0,1,2,3])
+x = model_w(x)
+# x = model(x)
+print(x.size())  # It should be (100, 20)
+```
+
+# ninja
+
+Download and extract ninja 1.8.2 from <https://github.com/ninja-build/ninja/releases/download/v1.8.2/ninja-linux.zip>. Add ninja to environment variable `PATH`.
+
+# Environment Variables
+
+After installing pytorch, cuda and ninja, modify and add following lines to your `.bashrc` file.
+
+```bash
+export anaconda_home=<your-anaconda-directory>
+export PATH=${anaconda_home}/bin:${PATH}
+export LD_LIBRARY_PATH=${anaconda_home}/lib:${LD_LIBRARY_PATH}
+
+export CUDA_HOME=<your-cuda-directory>
+export LD_LIBRARY_PATH=${CUDA_HOME}/lib64:${LD_LIBRARY_PATH}
+
+export PATH=<your-ninja-directory>:${PATH}
+```

README.md

@@ -1,35 +1,114 @@
-# PyTorch-Encoding
+# About
 
-created by [Hang Zhang](http://hangzh.com/)
+This is a fork from [PyTorch-Encoding](https://github.com/zhanghang1989/PyTorch-Encoding) (refer to it for the original README). In this repository, we use [DANet](https://github.com/junfu1115/DANet) to train a part segmentation model on [COCO Densepose Dataset](http://densepose.org/), for use in our person re-identification paper [EANet](https://github.com/huanghoujing/EANet).
 
-## [Documentation](http://hangzh.com/PyTorch-Encoding/)
+# Installation
 
-- Please visit the [**Docs**](http://hangzh.com/PyTorch-Encoding/) for detail instructions of installation and usage. 
+First install prerequsites [as this instruction](Install_Prerequisite_for_Pytorch_Encoding.md).
 
-- Please visit the [link](http://hangzh.com/PyTorch-Encoding/experiments/segmentation.html) to examples of semantic segmentation.
+Then clone this project
+```bash
+git clone https://github.com/huanghoujing/PyTorch-Encoding.git
+```
+
+# Dataset
 
-## Citations
+We transform COCO Densepose Dataset as described in paper [EANet](https://github.com/huanghoujing/EANet). For details, please refer to `scripts/coco_part`. **Note** that import and data paths in these scripts are not re-arranged and will cause error when executed. 
 
-**Context Encoding for Semantic Segmentation** [[arXiv]](https://arxiv.org/pdf/1803.08904.pdf)  
- [Hang Zhang](http://hangzh.com/), [Kristin Dana](http://eceweb1.rutgers.edu/vision/dana.html), [Jianping Shi](http://shijianping.me/), [Zhongyue Zhang](http://zhongyuezhang.com/), [Xiaogang Wang](http://www.ee.cuhk.edu.hk/~xgwang/), [Ambrish Tyagi](https://scholar.google.com/citations?user=GaSWCoUAAAAJ&hl=en), [Amit Agrawal](http://www.amitkagrawal.com/)
+The resulting segmentation dataset can be downloaded from [Baidu Cloud](https://pan.baidu.com/s/1Mm2gWO-Xg3wiyCd6SEAWaA#list/path=%2Fsharelink1629242940-281396814376268%2FEANet%2Fdataset%2Fcoco&parentPath=%2Fsharelink1629242940-281396814376268) or [Google Drive](https://drive.google.com/drive/folders/1gITlG2MfhJXUpfEPt6ohJCBgqigchabW).
+
+Prepare the dataset to have following structure
 ```
-@InProceedings{Zhang_2018_CVPR,
-author = {Zhang, Hang and Dana, Kristin and Shi, Jianping and Zhang, Zhongyue and Wang, Xiaogang and Tyagi, Ambrish and Agrawal, Amit},
-title = {Context Encoding for Semantic Segmentation},
-booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
-month = {June},
-year = {2018}
-}
+${project_dir}/dataset/coco_part
+    images
+    masks_7_parts
+    masks -> masks_7_parts  # This is a link created by `ln -s masks_7_parts masks`
+    train.txt
+    train_market1501_cuhk03_duke_style.txt
+    val.txt
 ```
 
-**Deep TEN: Texture Encoding Network** [[arXiv]](https://arxiv.org/pdf/1612.02844.pdf)  
-  [Hang Zhang](http://hangzh.com/), [Jia Xue](http://jiaxueweb.com/), [Kristin Dana](http://eceweb1.rutgers.edu/vision/dana.html)
+# Examples
+
+Download our trained model ([Baidu Cloud](https://pan.baidu.com/s/1Mm2gWO-Xg3wiyCd6SEAWaA#list/path=%2Fsharelink1629242940-281396814376268%2FEANet%2Fpart_segmentation_model&parentPath=%2Fsharelink1629242940-281396814376268) or [Google Drive](https://drive.google.com/drive/folders/1suBZk1WhpiS5PdB3GFySEEkC2FjzPQpL)) to `${project_dir}/exp/EANet_paper_ps_model/model_best.pth.tar`.
+
+## Infer and Visualize Your Images
+
+Specify your image directory `dir_of_im_to_vis` and an output directory `vis_save_dir` in the following command. It will select at most `max_num_vis` images to infer and save the visualization.
+
+```bash
+CUDA_VISIBLE_DEVICES=0 \
+python experiments/coco_part/train.py \
+--dir_of_im_to_vis YOUR_IMAGE_DIRECTORY \
+--vis_save_dir OUTPUT_DIRECTORY \
+--resume exp/EANet_paper_ps_model/model_best.pth.tar \
+--only-vis \
+--max_num_vis 128
 ```
-@InProceedings{Zhang_2017_CVPR,
-author = {Zhang, Hang and Xue, Jia and Dana, Kristin},
-title = {Deep TEN: Texture Encoding Network},
-booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
-month = {July},
-year = {2017}
-}
+
+`misc/example_visualization.png` is an example result.
+
+## Infer and Save Prediction
+
+Specify your image directory `dir_of_im_to_infer` and an output directory `infer_save_dir` in the following command. Prediction will be saved to `infer_save_dir`, with same sub paths as original images.
+
+```bash
+CUDA_VISIBLE_DEVICES=0 \
+python experiments/coco_part/train.py \
+--resume exp/EANet_paper_ps_model/model_best.pth.tar \
+--only-infer \
+--dir_of_im_to_infer YOUR_IMAGE_DIRECTORY \
+--infer_save_dir OUTPUT_DIRECTORY
+```
+
+For each image, the prediction is saved as a single-channel PNG image, with the same resolution as the input image. Each pixel value of the output image is its part label. Refer to [this link](https://github.com/huanghoujing/EANet#part-segmentation-label-format) for part index. Optionally, you can use script `experiments/coco_part/colorize_pred_mask.py` to colorize the predicted masks for visualization.
+
+## Validate on COCO Part Val Set
+
+The following command validates on val set, with single scale and flipping, but without cropping. You should get result `pixAcc: 0.9034, mIoU: 0.6670`.
+
+```bash
+CUDA_VISIBLE_DEVICES=0 \
+python experiments/coco_part/train.py \
+--resume exp/EANet_paper_ps_model/model_best.pth.tar \
+--only-val
+```
+
+
+## Training
+
+Since person images are much smaller compared with other segmentation tasks, we can use a single GPU while maintaining a large batch size. For example, when we set batch size to 16, the GPU usage is about 5600MB.
+
+```bash
+CUDA_VISIBLE_DEVICES=0 \
+python experiments/coco_part/train.py \
+--norm_layer bn \
+--train-split train \
+--batch-size 16 \
+--test-batch-size 16 \
+--exp_dir exp/train
 ```
+
+You can also try multi GPUs and synchronized BN by setting `norm_layer` to `sync_bn`
+
+```bash
+CUDA_VISIBLE_DEVICES=0,1 \
+python experiments/coco_part/train.py \
+--norm_layer sync_bn \
+--train-split train \
+--batch-size 16 \
+--test-batch-size 16 \
+--exp_dir exp/train
+```
+
+# Citation
+
+If you find our work useful, please kindly cite our paper:
+```
+@article{huang2018eanet,
+  title={EANet: Enhancing Alignment for Cross-Domain Person Re-identification},
+  author={Huang, Houjing and Yang, Wenjie and Chen, Xiaotang and Zhao, Xin and Huang, Kaiqi and Lin, Jinbin and Huang, Guan and Du, Dalong},
+  journal={arXiv preprint arXiv:1812.11369},
+  year={2018}
+}
+```

encoding/datasets/coco_part.py

@@ -1,6 +1,6 @@
 ###########################################################################
 # Created by: Houjing Huang
-# Copyright (c) 2018
+# Copyright (c) 2019
 ###########################################################################
 
 import os