facebookresearch / jepa

PyTorch code and models for V-JEPA self-supervised learning from video.
Other
2.61k stars 248 forks source link

Why no Colab notebook? #2

Open cperry-goog opened 6 months ago

cperry-goog commented 6 months ago

Just curious.

mikerabbat commented 6 months ago

We certainly welcome contributions if you would like to add one :-)

thomasf1 commented 6 months ago

I´ve put together a colab :)

However, the eval script vith16_k400_16x8x3.yaml seems to hardly utilize the GPU. It occasionally spikes to 100% but mostly is idle.

optimization:
  attend_across_segments: true
  num_epochs: 20
  resolution: 224
  batch_size: 16*

PS: Started the training with

!python -m evals.main --fname /content/jepa/xxx2.yaml --devices cuda:0