This is not an issue but I am still curious.
Before feeding those features into the lstm, the CNN will have to extract the features by using a for loop, and this might slow down the process of training. Actually, when I am using your code to train on my own dataset, which contains more than 100000 images, it takes more than half an hour to train the resnet+LSTM for one epoch on 3 TITANX GPU(batch size = 48, num_workers=32). Do you have any idea how to improve the speed?
This is not an issue but I am still curious. Before feeding those features into the lstm, the CNN will have to extract the features by using a for loop, and this might slow down the process of training. Actually, when I am using your code to train on my own dataset, which contains more than 100000 images, it takes more than half an hour to train the resnet+LSTM for one epoch on 3 TITANX GPU(batch size = 48, num_workers=32). Do you have any idea how to improve the speed?