Skip to content

Latest commit

 

History

History
143 lines (90 loc) · 5.82 KB

readme.md

File metadata and controls

143 lines (90 loc) · 5.82 KB
Raw

Setup

Deep learning requires access to accelerated computation hardware. Specifically, NVIDIA GPUs, or Google TPUs.

If you have access to a computer that has an NVIDIA GPU and runs Linux, you're welcome to set it up for local user.

If you don't, you can compute using Google Colab.

If you'd like to use Google Cloud Platform, we have a tutorial for setting up an AI Platform Jupyter Lab instance for computing, which works quite well.

Colab

Watch As Video

Google Colab is a great way to get access to fast GPUs for free. All you need is a Google account (Berkeley accounts will work).

Go to https://colab.research.google.com, and create a New Notebook.

Connect your new notebook to a GPU runtime by doing Runtime > Change Runtime type > GPU.

Now, run !nvidia-smi in the first cell (press Shift+Enter to run a cell). You should see a table showing your precious GPU :)

Now, paste the following into a cell and run it:

# FSDL Spring 2021 Setup
!git clone https://github.com/full-stack-deep-learning/fsdl-text-recognizer-2021-labs
%cd fsdl-text-recognizer-2021-labs
!pip3 install boltons wandb pytorch_lightning==1.1.4 pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio==0.7.2 torchtext==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
%env PYTHONPATH=.:$PYTHONPATH

This will check out our lab repository, cd into it, install the only missing Python package (pytorch_lightning -- everything else, like PyTorch itself, comes pre-installed on Colab), and allows Python to find packages in the current working directory.

Now we can enter the lab1 directory and make sure things work:

Colab Pro

You may be interested in signing up for Colab Pro.

For $10/month, you get priority access to fast GPUs (e.g. V100 vs K80) and TPUs, a 24h rather than 12h runtime, and more RAM.

VSCode on Google Colab

(This is advanced and not necessary -- skip for now!).

It is possible to use the VSCode interface in Colab.

Open a Colab notebook (or continue in the one you may already have going), make sure you're connected to the runtime you want, and run this in a cell:

# Launch VSCode server
!curl -fsSL https://code-server.dev/install.sh | sh
!nohup code-server --port 9000 --auth none &

# Tunnel its IP using ngrok
!pip install pyngrok
from pyngrok import ngrok
# ngrok.set_auth_token("get from https://dashboard.ngrok.com/auth/your-authtoken, if you want to pay $10/month for a little bit better service")
url = ngrok.connect(9000)
print(url)

This is what you should see:

Clicking the ngrok link takes you to a web VSCode interface:

You can sign up for a paid version of ngrok ($10/month) in order to get https tunneling and a slightly nicer experience.

Local

Setting up a machine you can sit in front of or SSH into is easy.

1. Check out the repo

git clone https://github.com/full-stack-deep-learning/fsdl-text-recognizer-2021-labs.git
cd fsdl-text-recognizer-2021-labs

2. Set up the Python environment

We use conda for managing Python and CUDA versions, and pip-tools for managing Python package dependencies.

We add a Makefile for making setup dead-simple.

First: Install the Python + CUDA environment using Conda

Note: this step is not necessary if you are using the GCP AI Platform Notebooks instance.

Conda is an open-source package management system and environment management system that runs on Windows, macOS, and Linux.

To install conda, follow instructions at https://conda.io/projects/conda/en/latest/user-guide/install/linux.html

Note that you will likely need to close and re-open your terminal. Afterward, you should have ability to run the conda command in your terminal.

Run make conda-update to create an environment called fsdl-text-recognizer-2021, as defined in environment.yml. This environment will provide us with the right Python version as well as the CUDA and CUDNN libraries.

If you edit environment.yml, just run make conda-update again to get the latest changes.

Next, activate the conda environment.

conda activate fsdl-text-recognizer-2021

IMPORTANT: every time you work in this directory, make sure to start your session with conda activate fsdl-text-recognizer-2021.

Next: install Python packages

Next, install all necessary Python packages by running make pip-tools

Using pip-tools lets us do three nice things:

  1. Separate out dev from production dependencies (requirements-dev.in vs requirements.in).
  2. Have a lockfile of exact versions for all dependencies (the auto-generated requirements-dev.txt and requirements.txt).
  3. Allow us to easily deploy to targets that may not support the conda environment.

If you add, remove, or need to update versions of some requirements, edit the .in files, and simply run make pip-tools again.

Set PYTHONPATH

Last, run export PYTHONPATH=. before executing any commands later on, or you will get errors like ModuleNotFoundError: No module named 'text_recognizer'.

In order to not have to set PYTHONPATH in every terminal you open, just add that line as the last line of the ~/.bashrc file using a text editor of your choice (e.g. nano ~/.bashrc)

Summary

  • environment.yml specifies python and optionally cuda/cudnn
  • make conda-update creates/updates the conda env
  • conda activate fsdl-text-recognizer-2021 activates the conda env
  • requirements/prod.in and requirements/dev.in specify python package requirements
  • make pip-tools resolves and install all Python packages
  • add export PYTHONPATH=.:$PYTHONPATH to your ~/.bashrc and source ~/.bashrc