Oufattole/meds-torch

Implementation of a Masking Stage with Random Masking Options

Opened this issue · 1 comments

Implementation of a Masking Stage with Random Masking Options

Problem

The absence of a dedicated masking stage in our pipeline limits our ability to handle incomplete or noisy data effectively during model training.

Proposed Solution

Introduce a masking stage designed to randomly mask a specified percentage of the data or subsequences within the data:

  • Position: Place the masking stage after the input encoder and before the sequence model.
  • Functionality:
    • Support random masking, either a random percentage of the tokens are masked or a randomly sampled continuous subsequence is masked.
    • We should add to the batch a key indicating the labels that will be used by the Model stage to compute masked imputation loss.
  • Configurability: Allow users to set the percentage of data to mask.

The token loss will be the same as for forecasting:

Lemme know what you think @teyaberg