0425v1

Semantic segmentation on Cityscapes/README.md

@@ -1,58 +1,14 @@
-# Pytorch-segmentation-toolbox Pytorch-1.1 [DOC](https://weiyc.github.io/assets/pdf/toolbox.pdf)
-Pytorch code for semantic segmentation. This is a minimal code to run PSPnet and Deeplabv3 on Cityscape dataset.
-Shortly afterwards, the code will be reviewed and reorganized for convenience.
+# Semantic Segmentation on Cityscapes
 
-### Highlights of Our Implementations
-- Synchronous BN
-- Fewness of Training Time
-- Better Reproduced Performance
+Our code is mainly based on
+[pytorch-segmentation-toolbox](https://github.com/speedinghzl/pytorch-segmentation-toolbox).
+Please refer to their docs.
 
-### Requirements && Install
-Python 3.7
 
-4 x 12g GPUs (e.g. TITAN XP)
+## Run
 
-```bash
-# Install **Pytorch-1.1**
-$ conda install pytorch torchvision cudatoolkit=9.0 -c pytorch
+Train Deeplab-V3 on Cityscapes
 
-# Install **Apex**
-$ git clone https://github.com/NVIDIA/apex
-$ cd apex
-$ pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./
-
-# Install **Inplace-ABN**
-$ git clone https://github.com/mapillary/inplace_abn.git
-$ cd inplace_abn
-$ python setup.py install
 ```
-
-### Dataset and pretrained model
-
-Plesae download cityscapes dataset and unzip the dataset into `YOUR_CS_PATH`.
-
-Please download MIT imagenet pretrained [resnet101-imagenet.pth](http://sceneparsing.csail.mit.edu/model/pretrained_resnet/resnet101-imagenet.pth), and put it into `dataset` folder.
-
-### Training and Evaluation
-```bash
-./run_local.sh YOUR_CS_PATH [pspnet|deeplabv3] 40000 769,769 0
-``` 
-
-### Benefits
-Some recent projects have already benefited from our implementations. For example, [CCNet: Criss-Cross Attention for semantic segmentation](https://github.com/speedinghzl/CCNet) and [Object  Context  Network(OCNet)](https://github.com/PkuRainBow/OCNet) currently  achieve  the  state-of-the-art  resultson  Cityscapes  and  ADE20K. In  addition, Our code also make great contributions to [Context Embedding with EdgePerceiving (CE2P)](https://github.com/liutinglt/CE2P), which won the 1st places in all human parsing tracks in the 2nd LIP Challange. 
-
-### Citing
-
-If you find this code useful in your research, please consider citing:
-
-    @misc{huang2018torchseg,
-      author = {Huang, Zilong and Wei, Yunchao and Wang, Xinggang, and Liu, Wenyu},
-      title = {A PyTorch Semantic Segmentation Toolbox},
-      howpublished = {\url{https://github.com/speedinghzl/pytorch-segmentation-toolbox}},
-      year = {2018}
-    }
-
-### Thanks to the Third Party Libs
-[inplace_abn](https://github.com/mapillary/inplace_abn) - 
-[Pytorch-Deeplab](https://github.com/speedinghzl/Pytorch-Deeplab) - 
-[PyTorch-Encoding](https://github.com/zhanghang1989/PyTorch-Encoding)
+./run_local.sh YOUR_CS_PATH [deeplabv3|deeplabv3_isda] 40000 769,769 0
+```

Semantic segmentation on Cityscapes/run_local.sh

@@ -14,6 +14,6 @@ INPUT_SIZE=$4
 OHEM=$5
 GPU_IDS=0,1,2,3
 
-CUDA_VISIBLE_DEVICES=${GPU_IDS} python -m torch.distributed.launch --nproc_per_node=4 train_isda.py --data-dir ${CS_PATH} --model ${MODEL} --random-mirror --random-scale --restore-from ./dataset/resnet101-imagenet.pth --input-size ${INPUT_SIZE} --gpu ${GPU_IDS} --learning-rate ${LR}  --weight-decay ${WD} --batch-size ${BS} --num-steps ${STEPS} --ohem ${OHEM}
+CUDA_VISIBLE_DEVICES=${GPU_IDS} python -m torch.distributed.launch --nproc_per_node=4 train_isda.py --data-dir ${CS_PATH} --model ${MODEL} --random-mirror --random-scale --restore-from ./dataset/resnet101-imagenet.pth --input-size ${INPUT_SIZE} --gpu ${GPU_IDS} --learning-rate ${LR}  --weight-decay ${WD} --batch-size ${BS} --num-steps ${STEPS} --ohem ${OHEM} --lambda_0 7.5
 
 CUDA_VISIBLE_DEVICES=${GPU_IDS} python -m torch.distributed.launch --nproc_per_node=4 evaluate.py --data-dir ${CS_PATH} --model ${MODEL} --input-size ${INPUT_SIZE} --batch-size 4 --restore-from isda_results/CS_scenes_${STEPS}.pth --gpu 4