How to get network convergent?

[haithanhp]

Hi @SimJeg ,

Thanks for your great work. I have the following concerns:

1. How many epochs to get the model convergent on camvid data?
2. In the paper, you use exponential decay of -0.995 for learning rate (1e-3). Is that mean you compute lr = 1e-3 * e^(-0.995 * epoch_id) for each epoch?
3. I want to reproduce your results in pytorch. I test on 9 images with visualized grounth truth and predictions below. But after 50 epochs, I did not see better.

Could you share with the tricks to train model? Best, Hai