-
Notifications
You must be signed in to change notification settings - Fork 6
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Pretrained weights #4
Comments
Due to Intel's policy, we do not have an immediate plan to share the pre-trained weights. We ask you to train the model on your end since the whole training process takes less than a few hours. Thank you, |
Thanks for the response. Seems a bit silly, since they can be recreated. In any case, thanks again for making this repository available. 😁 |
Sorry to bother you again, but I'm having trouble recreating the training features in RESNET18-TSM-AUG. Using the active-speakers-context repository, I swap the existing model for your models_stage1_tsm.py model, use the pretrained weights and I execute STE_forward.py with number of frames=11. The resulting features are not however the same as yours. Did you use all default parameters? The only other change I made was to reshape the video_data so it would be compatible with your model. (The input video data shape is (1, 11, 3, 144, 144) and the audio is (1,1,13,40). Does that seem correct?) |
Hi there,
Thank you for this excellent work!!!
Would it be possible for you to share the pre-trained weights for the AVA active speaker model?
It would be much appreciated. :-)
The text was updated successfully, but these errors were encountered: