This is the repository for the submission to SMCDC2021: Smoky Mountains Computational Sciences and Engineering Conference Data Challenge - Challenge 3 The code is developed based on the repository of Hierarchical Multi-Scale Attention for Semantic Segmentation

Installation

• The code is tested with pytorch 1.3 and python 3.6
• You need to install apex to run the code:

  > cd apex-master
  > pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

If the script above does not work please try

  > cd apex-master
  > pip install -v --disable-pip-version-check --no-cache-dir ./

• Install requirement

  > pip install -r requirements.txt

Run Inference

• Download pretrained weights from google drive and save it to folder weight

  > mkdir <weight>

• Create a directory where you save the testing images. In this case, we save image from cityscapes and SoftRainNoon (CARLA)

  > mkdir <imgs>
  > mkdir <imgs/Cityscapes>
  > mkdir <imgs/SoftRainNoon>

• Run the script below to get inference image. Remember to change the <eval_folder> and <result_dir>. We should keep the image from each domain into a seperated folder

> python -m torch.distributed.launch --nproc_per_node=1 train.py --dataset smc --cv 0 --apex --fp16 --bs_val 1 --eval folder --eval_folder ./imgs/SoftRainNoon/ --dump_all_images --n_scales 0.5,1.0,2.0 --snapshot ./weight/best_checkpoint_ep33.pth --arch ocrnet.HRNet_Mscale --result_dir ./imgs/SoftRainNoon_prediction/ --datasetBN
> python -m torch.distributed.launch --nproc_per_node=1 train.py --dataset smc --cv 0 --apex --fp16 --bs_val 1 --eval folder --eval_folder ./imgs/Cityscapes/ --dump_all_images --n_scales 0.5,1.0,2.0 --snapshot ./weight/best_checkpoint_ep33.pth --arch ocrnet.HRNet_Mscale --result_dir ./imgs/Cityscapes_prediction/ --datasetBN

Training

to be updated