GPU OOM during prediction

[mitchellxh]

Calling TreeOfLifeClassifier.predict keeps every batch’s probability tensors on GPU memory until completion, so VRAM usage grows with dataset size. I'm using >10,000 images; this exhausts my 32GB-RTX5000 at about 65% of dataset predicted:

torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 130.00 MiB...

Profiling shows that the create_probabilities_for_images helper keeps probability tensors on the device until the entire dataset finishes.

I fixed this by offloading each batch's probabilities to CPU with a tensor.detach().cpu() step before proceeding to the next. Submitted this as #144.

[mitchellxh]

Just want to add that this CPU-offloading fix assumes enough host RAM for the dataset. So this may be an acute solution rather than general (i.e. RAM $ <<< VRAM $).

Maybe consider writing them to disk is a general solution? Or just scaling run size to VRAM to avoid I/O.