search

arampacha / CLIP-rsicd

Apache License 2.0
211 stars 31 forks source link

Training logs #35

Closed INF800 closed 2 years ago

INF800 commented 2 years ago

Hi, I noticed that evaluation loss metric was too high compared to training loss in training notebook. Can you please share training logs for you best model(s) ?

arampacha commented 2 years ago

Hi. You can find links to wandb runs on modelcards. Here is the one with best performance modelcard : logs Here is a small report highlighting impact of augmentations on loss. As you can see we also have elav_loss > train_loss which is unsurprising given a relatively small dataset. Also notice that in many cases early-to-mid training checkpoints showed best performance. The later ones suffered more from overfitting. On the other hand it can also be concluded from our experiments this kind of fine-tuning/"domain adaptation" is highly beneficial for zero-shot classification performance. It might also worth a shot to freeze some of the layers during, but we haven't done those experiments in this project.