Home

Awesome

<div align = center> <br> <img src = "docs/_static/imgs/logo.svg" width="80%"> <br> <br> <!-- [![DeepCodebase](docs/_static/imgs/logo.svg)](https://github.com/hughplay/DeepCodebase) --> <!-- https://ileriayo.github.io/markdown-badges/ -->

Docker Anaconda

A template for deep learning projects.

[Use this template] [Document]

</div> <br>

Pre-requirements

<!-- <br> -->

Quick Start

TLDR; We have prepare the commands in the Makefile for you:

make init                           # run in host, build a docker image for api server and launch the container
make in                             # run in host, ctrl + d to exit the container

# experiment
make wandb_login                    # run in container, login to wandb
make train                          # run in container, train the model
make table                          # run in container, generate the latex table

# commit
make copy_git                       # run in host, copy git config to the container home
make precommit_install              # run in container, prepare pre-commit for tidy code
git add . && git commit -m "xxx"    # run in container, commit code, it will run pre-commit check

# profiling
python train.py experiment=mnist_lenet debug=profiler  # run in container, using pytorch profiler to profile the code
make nsightrun cmd="python train.py experiment=mnist_lenet debug=nsight"  # run in container, launch a nsight system profiler

Below is the detailed instructions.

1. Create new project from template

Click [Use this template] to create a new project from this template.

2. Clone the new project to local

You can find the repository url in the new project page.

git clone <repo url>

3. Prepare docker environment

Build the docker image and launch the container with:

python docker.py startd --build

4. Enter the docker container

Enter the container with:

python docker.py

The docker container is our development environment.

5. Train a model

Launch training with:

python train.py experiment=mnist_lenet

or launch multiple training at once:

python batchrun.py scripts/train/train_baselines.sh --quotas 1

6. View results on wandb

The training results will be automatically uploaded to wandb. You can view the results on wandb.ai.

7. Generate Latex table

After training multiple models, you can generate a latex table with:

python scripts/gen_latex_table.py --table baseline

This will generate a table like this:

\usepackage{booktabs}

\begin{table}[t]
\centering
\small
\caption{Performance on the MNIST dataset. }
\label{tab:baseline}
\begin{tabular}{lr}
\toprule
Model & Accuracy \\
\midrule
DNN & 0.974566 \\
LeNet & \textbf{0.983000} \\
\bottomrule
\end{tabular}
\end{table}

8. Custom your own project

Modify the code to fit your own deep learning project. Check our document for more details.

9. Commit you modifications

We strongly recommend you to use pre-commit to check your code before commiting. Before you first commit, you need to install the pre-commit hook with:

pre-commit install

Then, you can commit your modifications with normal git commands.

git add .
git commit -m "commit message"

It will show you the errors if your code does not pass the pre-commit checks. After you fix the errors, you need to git add the modified files and commit again.

Cite us

Give a star and cite us if you find this project useful.

DeepCodebase

[![DeepCodebase](https://img.shields.io/badge/Deep-Codebase-2d50a5.svg?style=flat-square)](https://github.com/hughplay/DeepCodebase)

This is a project based on DeepCodebase template.

*This is a project based on [DeepCodebase](https://github.com/hughplay/DeepCodebase) template.*

Projects based on DeepCodebase

License

DeepCodebase is released under the MIT license.