About the re-implement performance about pvse HOT 4 CLOSED

I have also experienced the same issue but I think I found the solution. Turning on the --img-finetune and --txt-finetune will fix it.
Also, don't forget to turn off legacy mode when you evaluate the newly trained model.

from pvse.

For prospective users:

Equivantly, you can just replace self.legacy argument in EncoderText to self.training that is already built in nn.Module for simplicity: change if self.legacy: to if self.training:

I have successfully reproduced the results in the main paper.
Many thanks to insight from @kdwonn

from pvse.

Hi there! Not really sure what might be causing your issue. It's been a while since I worked on this so my memories are fading. I'd suggest trying with a simpler model (e.g., disable MIL by setting --num_embeds 1, disable attention components, etc.), check the convergence/performance, and slowly add the options back into the model while monitoring the progress.

from pvse.

Closing due to inactivity

from pvse.