Training with latent cache

In your code, the .npz file is used to preserve the sampled results, but I think a better way is to preserve the mean and standard deviation, and then resample each time to ensure randomness.

Start by locating where the .npz file is written and read during training, then trace how sampled results are produced. Determine whether mean and standard deviation can be preserved without changing the expected training behavior. Done means an agreed design and implementation that resamples each time while retaining the required results and passing the relevant training checks.

Written by the indexing model from the issue text.