Issues

• Make AutoAugment work as non-final transformtion

  daniel-gallo/naraim#24 · Filter by author daniel-gallo opened on Jun 15, 2024

• Ablate the learning rate for attentive probe

  daniel-gallo/naraim#22 · Filter by author robertdvdk opened on May 29, 2024
  

• Check probe batch size 512 vs 1024

  daniel-gallo/naraim#21 · Filter by author robertdvdk opened on May 29, 2024
  

• Do attentive probing after 50k steps of pretraining, 100k, 150k, ..., 500k

  daniel-gallo/naraim#20 · Filter by author robertdvdk opened on May 29, 2024
  

• Implement gradient accumulation

  daniel-gallo/naraim#19 · Filter by author robertdvdk opened on May 29, 2024

• Check whether performance of the probe keeps improving even if the pre-training is overfitting

  daniel-gallo/naraim#18 · Filter by author robertdvdk opened on May 29, 2024
  

• Implement per-channel normalisation

  daniel-gallo/naraim#17 · Filter by author robertdvdk opened on May 29, 2024

• Implement visualization of error maps in the tensorboard logger

  daniel-gallo/naraim#16 · Filter by author robertdvdk opened on May 29, 2024
  

• Implement autoregressive generation of images

  daniel-gallo/naraim#14 · Filter by author robertdvdk opened on May 29, 2024

• Find out why we need to prepend 2 black patches in the visualisation

  daniel-gallo/naraim#13 · Filter by author robertdvdk opened on May 29, 2024

• Disable prefix sampling for evaluation

  daniel-gallo/naraim#12 · Filter by author daniel-gallo opened on May 28, 2024
  

• Native resolution training

  daniel-gallo/naraim#11 · Filter by author robertdvdk opened on May 28, 2024

• Data parallelism

  daniel-gallo/naraim#10 · Filter by author robertdvdk opened on May 28, 2024

• Check classification head

  daniel-gallo/naraim#9 · Filter by author robertdvdk opened on May 28, 2024

• Fix tests for ImageNet on CI/CD

  daniel-gallo/naraim#8 · Filter by author daniel-gallo opened on Apr 25, 2024