search

ajabri / videowalk

Repository for "Space-Time Correspondence as a Contrastive Random Walk" (NeurIPS 2020)
http://ajabri.github.io/videowalk
MIT License
267 stars 38 forks source link

ResNet50 performance #30

Closed xvjiarui closed 3 years ago

xvjiarui commented 3 years ago

Hi Allan,

Have you tried ResNet50 with your approach? I recently tried by myself, but get very low performance for the starting epoches, ~30 J&F mean. Any ideas?

ajabri commented 3 years ago

Hi @xvjiarui ,

I haven't experimented much with ResNet50, but using different networks (especially deeper networks) will require changing certain hyper-parameters; e.g. you may need a more extreme temperature (lower, like 0.01), and you may want to try a learning rate warmup. Regularization is also probably quite important; consider using dropout in the activations of the network, or weight decay with the AdamW optimizer.