Sequence dimension

[cellistigs]

The original code has a sequence dimension in addition to the initial batch dimension. Apparently, training previously used entire trials as held in/held out, with all the (subsampled) frames from a single trial being included along the sequence dimension for a total batch size of 8 (trials)x 62 (frames per trial).

At the moment, we just take uniform random samples from our subsampled frames. Let's test if this is necessary/makes a difference in terms of training behavior.

[cellistigs]

The actual original protocol is as follows:

• Each trial is around 4 seconds total.
• 20x subsampling gives 62 frames per trial.
• The actual frame choice is a (fixed seed) random process: the entire frame sequence is first binned into 62 chunks. A single frame from each chunk is then chosen as the representative.