butd-pytorch-docker

Dockerfile for MILVLG/bottom-up-attention.pytorch, a PyTorch reimplementation of the bottom-up-attention project based on Caffe

Quick start

Pre-start actions

In order use GPUs inside containers, we'll be using NVIDIA Container Toolkit.

Make sure you have installed the NVIDIA driver and Docker 19.03 for your Linux distribution Note that you do not need to install the CUDA toolkit on the host, but the driver needs to be installed.

For more information, please visit NVIDIA Container Toolkit official repository.

Pull or Build the image

To image, you can directly pull from Dockerhub[recommended] or build from the local environment.

To Pull

docker pull denton35/butd-pytorch-docker

will pull the latest image from dockerhub. Use sudo if needed.

To Build

Pull the Dockerfile: git clone https://github.com/HyperDenton/butd-pytorch-docker

Goto the repo directory: cd butd-pytorch-docker

Build the image: docker build .

Try the image

Start the image and enter the image bash:

sudo docker run --gpus all --rm -it -v <absolute-path-to-repo>/bottom-up-attention.pytorch:/workspace/bottom-up-attention.pytorch denton35/butd-pytorch-docker

in which

--gpus all to enable NVIDIA Container Toolkit;

--rm to remove the image after use;

-it to enter the image bash;

-v <local-path>:<image-path> to access host system's directory.

For more usage of Docker, please visit Docker Reference Page.

Feature extraction

Set up

If it's the first time you run the script or you made changes to the source code, please run under repo directory inside the image:

python setup.py build develop

Demo

python3 extract_features.py --mode caffe --config-file configs/bua-caffe/extract-bua-caffe-r101.yaml --image-dir datasets/demo/ --out-dir output/ --resume

this will extract every image in <datasets/demo/> and put the features to <output/> directory.

FAQ

`docker pull` takes too long

Consider using docker accelerator:

sudo mkdir -p /etc/docker
sudo tee /etc/docker/daemon.json <<-'EOF'
{
  "registry-mirrors": ["https://****.****.****.com"]
}
EOF
sudo systemctl daemon-reload
sudo systemctl restart docker

where https://****.****.****.com is your address of your accelerator.

To get address accelerator, you might need to register to the docker accelerator service provider, e.g. Aliyun

HyperDenton/butd-pytorch-docker