Fix spurious empty final batch when ntrain is a multiple of batch_size

[rezanematpour]

Problem

Across all training loop functions in Fin-GAN-online.py (e.g., TrainLoopForGAN, and the various TrainLoopMain* variants), batch count is computed as:

nbatches = ntrain//batch_size+1

This unconditionally assumes there is always a remainder batch to add. When ntrain is an exact multiple of batch_size, this creates one extra, spurious batch whose size is computed as:

curr_batch_size = totlen - i*batch_size

which evaluates to 0 for that final batch. This produces empty-tensor batches (e.g., condition/real of shape (0, ...), zero-batch LSTM hidden states h_0/c_0), which can break the forward pass and downstream indexing such as disc_fake_pred[0][0][0] used for logging.

Example

With ntrain=1000, batch_size=100:

• nbatches = 1000//100 + 1 = 11
• Batches i=0..9 correctly cover all 1000 rows (size 100 each)
• Batch i=10 gets curr_batch_size = 1000 - 1000 = 0 — an empty batch that shouldn't exist

Fix

Only add the extra batch when a remainder actually exists:

nbatches = ntrain//batch_size + (1 if ntrain % batch_size != 0 else 0)

This is a minimal, behavior-preserving change:

• No effect when ntrain % batch_size != 0 (the common case) — same batch count and sizes as before.
• Fixes the edge case when ntrain % batch_size == 0 — no more empty final batch.

Applied consistently to all occurrences of this pattern in Fin-GAN-online.py.

Testing

Verified the corrected formula against both cases:

• ntrain=1050, batch_size=100 → 11 batches, sizes [100]*10 + [50] (unchanged from original behavior)
• ntrain=1000, batch_size=100 → 10 batches, sizes [100]*10, no empty batch (previously would have produced an 11th, empty batch)

No other logic was changed.