Training set of 96 images, batch size of 12 with gradient accumulation of 8 for 110 epochs (880 steps). Using ADAMW with constant LR of 0.00075.
FP16 precision.
Bit dissatisfied with the final result, but it's reasonably passable. Working on another version that I may post.
Training set of 96 images, batch size of 12 with gradient accumulation of 8 for 110 epochs (880 steps). Using ADAMW with constant LR of 0.00075.
FP16 precision.
Bit dissatisfied with the final result, but it's reasonably passable. Working on another version that I may post.