I would like to ask a few questions regarding the paper.
When you plugged the CRF layer to the pretrained FCN-8s and trained end-to-end, did the (softmax) loss actually go down? How many epochs did you train the model to see that it eventually went down? I tried training end-to-end ('lr=1e-13, momentum=0.99' as described in the paper) with the pretrained FCN-8s weights extracted from your model but the loss does not seem to go down :(
Could you please share the (Keras) pretrained FCN-8s weights that you used? Also, did you use any data augmentation?
chciw
commented
2 years ago
Hi Sadeep,
Thanks for making this code available!
I would like to ask a few questions regarding the paper.
When you plugged the CRF layer to the pretrained FCN-8s and trained end-to-end, did the (softmax) loss actually go down? How many epochs did you train the model to see that it eventually went down? I tried training end-to-end ('lr=1e-13, momentum=0.99' as described in the paper) with the pretrained FCN-8s weights extracted from your model but the loss does not seem to go down :(
Could you please share the (Keras) pretrained FCN-8s weights that you used? Also, did you use any data augmentation?
Thank you very much in advance! Best regards.