Training much slower when adding the context (BiLSTM)

[pvthuy]

Hi,

Thanks for for your great implementation.

The motivation for adding the context, e.g., using BiLSTM, might be good for refining the semantic/instance masks. However, when I train the model with BiLSTM, the training time is much slower.

How many images in your train set? And how long did you train the model?
Do you have any suggestions for improving the train process? Thanks!