This updates the optimize_plain.py script into a simpler standalone reproduction that we can use to run hyper-parameter sweeps. I updated the parameters inside the script to match that of the data that you gave Charles. Using n_trials=1 generates around 100+ unique configurations which is close to the number of trials inside the .csv that you gave to him.
Changes
Added a simple requirements.txt file
Added a deduplication step that uses the hash of the config dataset to ensure we don't run duplicate experiments.
Converted the data class to a frozen data class so we can hash it - realised that we had some configurations that were being generated multiple times
Added a download dataset method so that users can download our dataset from quora and immediately cache it in a volume of choice.
:rocket: This description was created by Ellipsis for commit 9c40ffadf4e8cea15aad9957f01b9453b6ad36fc
Summary:
This PR simplifies the optimize_plain.py script for hyper-parameter sweeps, adds a deduplication step, converts the data class to a frozen data class, and adds a method to download the dataset from Quora.
Key points:
Simplified optimize_plain.py into a standalone reproduction for hyper-parameter sweeps.
Updated parameters to match provided data.
Added deduplication step using hash of config dataset.
Converted data class to a frozen data class for hashing.
This updates the
optimize_plain.pyscript into a simpler standalone reproduction that we can use to run hyper-parameter sweeps. I updated the parameters inside the script to match that of the data that you gave Charles. Usingn_trials=1generates around 100+ unique configurations which is close to the number of trials inside the .csv that you gave to him.Changes
Summary:
This PR simplifies the
optimize_plain.pyscript for hyper-parameter sweeps, adds a deduplication step, converts the data class to a frozen data class, and adds a method to download the dataset from Quora.Key points:
optimize_plain.pyinto a standalone reproduction for hyper-parameter sweeps.requirements.txtwith new dependencies.Generated with :heart: by ellipsis.dev