Optimize SSL training, fix remaining training problems

HLCaptain / ssl-in-the-wild

Comparison of Self-Supervised Learning models.

0 stars 2 forks source link

Closed HLCaptain closed 10 months ago

HLCaptain commented 10 months ago

Replaced training data location
Fixed training problem with Tensorboard (not adding histogram at the end of epoch due to crash)
Optimized training. num_workers is now based on number of CPU cores in the OS. Mitigated data imput bottleneck, so training is 4x+ faster with 4090 and 16 core amd ryzed 7590x3d.