learn-fastai

This repository is a collection of my notes and Jupyter notebooks I used during learning fastai

• Deep Learning for Coders with Fastai and PyTorch (Amazon)
• Practical Deep Learning for Coders 2020 (Youtube)
• fastbook (github)

My environment is a LSF cluster of several hundred GPU nodes. I use Anaconda to manage Python packages in my environment. I am also using Jupyter notebooks for learning.

• fastai docs
• Jupyter Widets
• Addtional documentation on inference

Environment setup

Create a new Anaconda environment and install the required packages.

conda create --name learn-fastai python=3.7 -y
conda activate learn-fastai 
conda install -c fastai -c pytorch fastai -y
conda install -c conda-forge jupyterlab -y
conda install -c anaconda psutil 
conda clean --tarballs

To remove this env use the following command:

conda env remove -n learn-fastai

By default, Anaconda installs the packages in the environments in your home directory. If your home directory is limited in space, you may want to place your environments in a different location. Also note that the dataloaders of fastai, place the data (by default) in the <your home directory>/.fastai so you may also want to change that.

To check the size of your home directory:

du -sh ~

Optional

conda install -c fastai fastbook
conda install -c conda-forge sentencepiece 

Launching and using a Jupyter Notebook server

Activate the Anaconda environment conda activate learn-fastai then use the launch-jupyter-server.sh shell script. The script launches a LSF job which executes the jupyter-server-job.sh shell script on the node chosen by LSF. Once the job launches, use the LSF bpeek command to get the output of the job which has the jupyter server url. Use the browser on your local system to point to that url.