Feature Idea: Pre-Tuning Architectural Validation

[davidrsch]

When using tune to search over architectural parameters (e.g., num_dense), it's easy to create a grid with combinations that are architecturally impossible (e.g., a conv2d layer following a dense layer without reshaping). Currently, tune_grid() attempts to fit these models, fails with an error, and moves on. This is computationally wasteful and clutters the console with errors.

This should ben addressed by introducing a new set of functions that allow for a "pre-flight check" of the tuning grid.

How it would work:

1. compile_keras_grid(): A new function that iterates through a tuning grid and performs a "dry run" of model construction and compilation for each hyperparameter combination. It returns a tibble with the compiled model, a summary of the model's architecture, and any errors that occurred during compilation.

2. extract_valid_grid(): A helper function that takes the results of compile_keras_grid() and returns a new grid containing only the hyperparameter combinations that compiled successfully.

3. inform_errors(): A helper function that takes the results of compile_keras_grid() and prints a formatted summary of any compilation errors that occurred.

The current alternative is to let tune_grid() fail on invalid architectures, which is inefficient.

Acceptance Criteria

• Create and export the compile_keras_grid() function.
• Create and export the extract_valid_grid() and inform_errors() helper functions.
• Ensure compile_keras_grid() correctly identifies and reports invalid parameter sets.
• The extract_valid_grid() function must return a cleaned grid of only valid combinations.
• The inform_errors() function must return a report of the removed sets and the corresponding errors.
• Add unit tests for the new functionality.