Skip to content

README.md

Latest commit

 

History

History
29 lines (16 loc) · 1.32 KB

README.md

File metadata and controls

29 lines (16 loc) · 1.32 KB

Architecture Search

Steps

  1. First, a suitable model for a given dataset is searched using DARTS architecture search.
    • The included code uses the 'CIFAR10' dataset.
  2. Then, the model is trained fully.

Executing this code may take a few hours, depending on your machine configuration.

Search model

$ python3 /app/Soteria/Architecture_Search/train_search.py

This will save a architecture pickle file nas_arch.pkl under directory /app/Soteria/Architecture_Search/trained_model.

If a pre-searched DARTS architecture exists: Place the architecture pickle file in this directory.

A pre-searched architecture pickle file can be found here. Place the trained architecture file directly in the directory /app/Soteria/Architecture_Search/trained_model.

The hyperparameters used to obtain this architecture are listed in file hyperparameters.csv.

Train searched model

$ python3 /app/Soteria/Architecture_Search/train.py

This will train the model fully.

The pre-trained model pickle for the architecture above can be found here. Place the trained model file directly in the directory /app/Soteria/Architecture_Search/trained_model.