about using coc_medium
lai-pf opened this issue · comments
A few changes seem to be required in the code
In models/context_cluster.py change
Cluster class: self.f --> self.fc1 (rename)
self.proj --> self.fc2 (rename)
self.v --> self.fc_v (rename)
Mlp class: self.fc1 --> nn.Conv2d(in_features, hidden_features, kernel_size=1) (change linear to conv2d)
self.fc2 --> nn.Conv2d(hidden_features, out_features, kernel_size=1) (change linear to conv2d)
also remove all permutes in forward()
In validate.py change
change line 155 to
load_checkpoint(model, args.checkpoint, args.use_ema, strict=False)
Although I'm not the author of this paper, these changes seemed to do the trick for me
@lai-pf @myeyeget Sry for the late reply.
We slightly re-organized the codes for fast training and inference. @myeyeget 's solution seems good.
You can also try the submission link here: https://anonymous.4open.science/r/ContextCluster/README.md
We will update the checkpoints soon. Thanks for your issue.
@ma-xu The submission link has expired. Could you please update it so that I can solve the mismatch problem?
@ErlebnisW @myeyeget @lai-pf
Thanks for your patience. The updated checkpoints are updated, with slightly improvements. Mainly two updates: 1) fix a bug mask.sum to sim.sum; 2) conv1x1 to Linear implementation.
Closed no further discussions. Feel free to reopen it if any further questions.