Deep learning laboratory

Includes:

• GPU support
• Python 3.7
• Jupyter + Tensorboard
• Pytorch + Ignite + TensorboardX
• Tensorflow + Keras
• Jupyter, numpy, scipy, pandas, skimage, matplotlib, tqdm
• Supervisord

How to use

Clone repository:

git clone https://github.com/vslutov/deeplearning.git

Run command:

usage: deeplearning.sh [-h] [-j JUPYTER_PORT] [-t TENSORBOARD_PORT]
                       [-p PASSWORD]
                       work_folder

Start deep learning laboratory

positional arguments:
  work_folder

optional arguments:
  -h, --help            show this help message and exit
  -j JUPYTER_PORT, --jupyter-port JUPYTER_PORT
                        The jupyter listen port [default: some free port]
  -t TENSORBOARD_PORT, --tensorboard-port TENSORBOARD_PORT
                        The jupyter listen port [default: some free port]
  -p PASSWORD, --password PASSWORD
                        The jupyter password [default: random string]

If you shut down jupyter server, container will shut down.

Tesnorboard is disabled by default, becouse it's too hungry for ram. You can enable tensorboard by command supervisorctl start tensorboard.

Multiple GPUs

You can use environment NVIDIA_VISIBLE_DEVICES to specify visible devices in container.

Requirements

1. Ubuntu, CentOS or RHEL.
2. Python 2.6+ or 3.2+, bash.
3. Docker
4. Nvidia-docker

How it works

Script deeplearning.sh:

If the container have been already started, run the shell in the container and exit.

1. Update the git repository (could be canceled with Ctrl-D).
2. Build the docker image. The first run may take a few minutes.
3. Generate unique jupyter password or take your -p option.
4. Create deeplearning_<name>_data volume. The name calculated from work_folder argument.
5. Run the docker container deeplearning_<name> from this image in background.
6. In the container work_folder mounts as /home/user/work and deeplearning_<name>_data as /home/user/data.
7. In the container starts supervisor, which launch a jupyter notebook and a tensorboard.
8. If you set the jupyter listening port, then the host port forwards to the jupyter notebook, else a random port sets.
9. The tensorboard runs with folder data_folder/runs as input. The listening port sets as like as the jupyter listening port.
10. Print out the container id, jupyter and tensorboard host ports and the jupyter password.
11. Run the shell in the container.