Closed man0007 closed 4 years ago
From the plot it seems like the loss is smaller for the default LR compared to LR you suggested.
Yeah that's true my assumption is that if I would increase the Epocs it may converge as equal to the lr=0.0001, by this we can achieve a stability and reduced loss.
HI,
I was trying to tweak the learning_rate and dropout parameters for the handwriting_line_recognition.py model.
Since there is no much change in the loss for changing the dropout parameters (20%, 35%, 50%) i'm just fixing the default one.
But for the learning rate change from 0.0001 to 0.00001 there is a huge increase in the stability of the model as plotted below. (training loss is equivalent to the test loss)
plotted graph image: https://prnt.sc/rv6lzm
graph_label notations:
lr-e5 => learning_rate = 0.00001 lr-e4 => learning_rate = 0.0001
-> Bottom two lines are the train and test loss calculation for the 0.0001 learning_rate parameters and all above lines are plotted for 0.00001. We could see the bottom two lines are not stable where as the other lines are very stable (training loss is equivalent to the test loss)
Since the lr 0.00001 is better than 0.0001, can we fix 0.00001 as default or do we face any other problem if we use this new lr rate?
Please advice.
Thanks, Anand.