This pull request introduces several improvements and bug fixes to the DatasetTorch class and its usage, focusing on batch size handling, error checking, and test consistency. The main changes enforce that the batch size must evenly divide the number of samples per file, update the logic for calculating batches per file, and update tests and configuration to use consistent, valid batch sizes.

Core logic and validation improvements:

• Added validation in DatasetTorch to raise a ValueError if batch_size does not evenly divide the number of samples per file, ensuring data consistency and preventing subtle bugs during batching.
• Simplified the calculation of the number of batches per file and the logic for generating batch indices in __getitem__, improving code clarity and reliability.

Test updates and coverage:

• Updated all relevant tests in tests/test_torch_mlp.py to use batch sizes that evenly divide the sample count, and added a new test to verify that an error is raised if this condition is not met. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10]

Configuration and constants alignment:

• Updated batch size values in test configuration files and constants to reflect the new divisibility requirement, ensuring tests and sample configs are valid. [1] [2] [3] [4]

Minor code cleanups:

• Replaced usage of .keys() with direct in checks for dictionary membership throughout the code for improved readability and Pythonic style. [1] [2] [3]