[Share]How to run demo and training in a docker container

[liupgd]

After miserable experience on running training on this repo. I think I need to write it down to share to others. My situation is: I failed run the model on my own ubuntu server, as I didn't install CUDA10.0. But, running in the docker is also not so easy.

Run the demo

• Build the docker

  docker build -t new-comod-gan .
  docker run -itd -v /your_work_dir:/work  -v /your_data_dir:/data --name comod -p 7200-7220:7200-7220 --gpus all new-comod-gan /bin/bash

• Install extra packages in your container

  pip install opencv-python 
  pip install tqdm
  pip install scikit-learn

• Prepare your test image dataset

  python dataset_tools/create_from_images.py --val-image-dir ./your_test_images_dir --tfrecord-dir ./tfrecords

• Add cuda to path in your container, you'd better add this to .bashrc

  export PATH="/usr/local/nvidia/bin:/usr/local/cuda/bin:$PATH"

• Now you can run the demo with pretrained model

  python run_demo.py -c ./pretrained/co-mod-gan-places2-050000.pkl -d ./tfrecords

• But...you didn't get the GUI from your remote container? Maybe VcXsrv can help you if you are using windows locally. My solution is:
  • Install VcXsrv in windows
  • Use vscode to access my container, and install Remote X11 plugin in vscode.
  • In vscode settings->Remote, search Host Option, and set your remote Host'IP with your remote server's IP(Not the container's IP)
  • Start VcXsrv first in your local Windows, and then run your remote demo. GUI will be displayed if there is no error.

Run training

• Preparing your own training dataset Here, I prepared some images in ./imgs/png_samples/ for training test.

  python dataset_tools/create_from_images.py --train-image-dir ./imgs/png_samples/ --val-image-dir ./imgs/png_samples/ --tfrecord-dir ./train_dataset --resolution 512 --num-channels 3

  Note:

  1. Only 3 channels can be used. If you're using png files, do not set --num-channels to 4, you'll get error in training.
  2. --val-image-dir should be specified, or you'll have error in training.

• Run your training

  python run_training.py --data-dir=./  --dataset=train_dataset --metrics=ids10k --mirror-augment True --num-gpus=4

  For researchers in China, you may need a VPN. The training process will download an inception model file. You can:

  export https_proxy="https://your_vpn_ip&port"