gan_training

GAN Training Tutorial Setup

This tutorial can be cloned via

git clone https://github.com/alexandonian/ganocracy.git

If you do not have git installed on your local machine, you can download the repo as a ZIP file with

wget https://github.com/alexandonian/ganocracy/archive/master.zip

Requirements:

• Python 3.6 or greater.
• PyTorch 1.1 (LTS). Detailed installation instructions can be found here.
• torchvision 0.3.0. Note: Version 0.3.0 was just announced on May 23, 2019, so existing installations may need updating.
• tqdm, numpy, scipy, pandas, and h5py
• moviepy (optional)

Step 1: Download Anaconda If your system does not already meet these requirements, we recommend downloading the Anaconda Distribution of Python 3. Anaconda comes with the package manager conda (installation instructions can be found here), which makes installing PyTorch and other dependencies much easier.

Step 2 (optional): Create conda environment After installing conda, you can create an environment for this project to manage dependencies:

# Create a conda environment called 'gantraining'
conda create --name gantraining

# To activate this environment, use:
conda activate gantraining

Step 3: Install PyTorch and dependencies:

Installing the latest version of PyTorch and torchvision can be done in a single line:

conda install pytorch torchvision -c pytorch

and the dependencies with:

conda install h5py pandas
pip install tqdm moviepy

Starting the Jupyter Notebook

Locally: You can start the tutorial notebook from the repository (ganocracy) you cloned or downloaded via:

cd gan_training
jupyter notebook

A browser window should automatically open, and you can run the notebook Training.ipynb

Remotely with SSH Port-Forwarding: If you are fortunate to have access to a headless remote server, ideally with several GPUs, it is possible to run the notebook on the server and still view it locally on your personal machine via ssh port-forwarding. Feel free to read more about ssh port-forwarding on your own, but it's not necessary for this tutorial.

Using port forwarding to view Juypter notebooks on a remote server requires two steps:

1. On the REMOTE machine: Start the jupyter notebook as normal. You may need to include the --no-browser option:

   cd tutorial
   jupyter notebook --no-browser
   

   Jupyter notebook will display a bunch of information, which should include something like:

   [I 18:20:18.770 NotebookApp] The Jupyter Notebook is running at: http://localhost:8888/?token=MY_TOKEN
   [I 18:20:18.770 NotebookApp] Use Control-C to stop this server and shut down all kernels (twice to skip confirmation).
   [C 18:20:18.771 NotebookApp]
   
       Copy/paste this URL into your browser when you connect for the first time,
       to login with a token:
           http://localhost:8888/?token=MY_TOKEN
   

   Make note of the port number and token. In this case, we have REMOTE_PORT=8888

2. On the LOCAL machine: Decide which port you want to use to view the notebook. Choose anything in the range 1024-49151 (unreserved "user" ports). Call the port forwarding command:

   LOCAL_PORT=8888
   REMOTE_PORT=8888  # From step 1.
   
   ssh -fNL $LOCAL_PORT:localhost:$REMOTE_PORT $REMOTE_IP_ADDR
   

   In your browser, navigate to localhost:$LOCAL_PORT. If you are prompted, enter your token from step 1. You should now be able to see your jupyter notebooks running on the remote server.

   Finally, run the notebook Training.ipynb