You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Do not max-pool the last convolutional layer. Do what is necessary to allow attention to work with the result
Make a toy model using just a small subset of the easiest ICFHR data. Just get attention to be able to learn this data easily. IAM dataset could also be something to try
Train attention and train attention CTC look pretty much the same for the attention part.
I could start training on the small subset of data using the CTC-attention model, to see how far attention can go on that data. Then I need to debug just the attention alone, probably picking it apart piece by piece in a jupyter notebook to see what it is doing.