I want to use 3DCNN for images classification, i.e. classification with a sequence of images as inputs to 3DCNN models, then output the corresponding feature maps, during the training, how to define the loss function and accuracy? I implemented a version, but the loss function did not decrease significantly, resulting in no improvement in accuracy, and the training was consistent from beginning to end.
I feel that there may be something wrong with the loss function and accuracy calculation, because the learning rate I define is steadily declining. So do you have any good suggestions to help me find the problem?
