neuro-inc / cookbook Goto Github PK

If we assume that our customers will use them as starting points, then forking a repo is much more convenient than copy-pasting code. Also, having them in separate repositories makes community contributing easier.

Let's stick to 'ML Recipe' and 'Cookbook' terminology from now on for an individual project and a collection, respectively.

So, I suggest to do the following:

• Put each recipe in a separate repository called 'ml-recipe-XXX'.
• Rename this repository to 'cookbook'.
• Make other repositories submodules, tracking head of master branches.
• Update READMEs accordingly.

• Change GitHub links to https.
• Rearrage commands (make setup and make jupyter do not work together).
• Add a name of the Jupyter notebook to play with.

Update all Recipes to the latest version of the Cookiecutter template, which includes all of these:

• neuro-inc/cookiecutter-neuro-project#179
• neuro-inc/cookiecutter-neuro-project#182
• neuro-inc/cookiecutter-neuro-project#186
• neuro-inc/cookiecutter-neuro-project#211
• neuro-inc/cookiecutter-neuro-project#233
• neuro-inc/cookiecutter-neuro-project#234
• neuro-inc/cookiecutter-neuro-project#252

Also, every Recipe has to satisfy the following requirements:

• In README there is a Quick Start section with sign up && clone && setup && jupyter.
• If the data is downloaded, it is downloaded from S3 to Docker image local FS before each training/jupyter.
• If secrets are used (e.g. for w&b), they are passed securely (via files on storage, not env vars).
• Updated README should be on neu.ro.
• Update 'git' URIs to 'https'.
• Style: every inline code block should be in single back ticks; every other code block should be in "shell" mode.
• Style: Entries like

[ ! -f ../data/aclImdb.zip ] && wget http://data.neu.ro/aclImdb.zip -O ../data/aclImdb.zip
[ ! -d ../data/aclImbd ] && unzip ../data/aclImdb.zip -d ../data/

should be replaced by

wget -nc http://data.neu.ro/aclImdb.zip -O ../data/aclImdb.zip
unzip -n ../data/aclImdb.zip -d ../data/

Progress:

• object detection
• hier attention
• midi generator
• mountain car
• bone age

Would be nice to drag in the annotated transformer lesson from harvardnlp.
I think, the main work here is to bring the code up-to-date with the latest pytorch version

https://github.com/harvardnlp/annotated-transformer

Could not really google an example with sentiment analysis using some well-known framework

I suggest providing some wrapper over https://stanfordnlp.github.io/CoreNLP/sentiment.html

We should test the critical path nightly:

• make setup
• make jupyter
• make develop (without waiting until the process is finished)
• make kill-all

see neuro-inc/cookiecutter-neuro-project#70 (comment) for instructions

Apache 2.0
Copyright (c) 2019, Neuromation
See https://github.com/neuromation/ml-recipe-course-fast-ai-v3/blob/master/LICENSE