[WIP] Gradual Unfreezing to mitigate catastrophic forgetting

[ethanreidel]

Adds the ability to gradually unfreeze or thaw specific layers within a pre-trained model's architecture. Aims to mitigate catastrophic forgetting/improve transfer learning capabilities. Currently works for ECD architecture.

User passes in two things: thaw_epochs (list of integers) and layers_to_thaw (2D array of layer strings)

thaw_epochs: -1 -2 layers_to_thaw:

• ["features.0", "features.1"] (thaws these layers (weights+biases) at epoch 1)
• ["features.2", "features.3"] (epoch 2) (keep in mind "features.0" will thaw all layers with the prefix "features.0" e.g. "features.0.1/2/3")

TODO/potential issues:

• potentially change config syntax
• users currently need to know the exact strings in architecture for thawing which is inconvenient
• unittest iffy

test: [tests/ludwig/modules/test_gradual_unfreezing.py]

Any and all feedback is greatly appreciated. 👍

[github-actions[bot]]

Unit Test Results

       6 files  ±       0         6 suites  ±0   52m 7s :stopwatch: + 22m 2s 2 990 tests  -        3  2 966 :heavy_check_mark:  -      15  23 :zzz: +11  1 :x: +1  8 970 runs  +5 941  8 898 :heavy_check_mark: +5 893  69 :zzz: +45  3 :x: +3 

For more details on these failures, see this check.

Results for commit d2ba5cb0. ± Comparison against base commit 606c732a.

[ethanreidel]

@skanjila @saad-palapa