I was training and dense net model on emotion recognition on the sewa dataset. Therefore, at the end I have 2 outputs. One for arousal and the other for valence (These dimensions for emotions). So I am doing regression at the end.
My model contains batch normalization and dropout layers. after examining carefully the training and the testing loss; I found a considerable difference between them. Below are recorded losses on the training and testing dataset.
And I got almost the same behavior on the arousal dimension (2nd output). Hence, it is obvious that there is a difference in the recorded loss, and maybe I am overfitting.
Now, I tried combining the training and the testing dataset to test the recorded loss. Thus, now my I trained the model on both the training and the validation dataset. But, I got the following graphs for the loss:
Therefore, here in this case, I was training for one epoch on one hand, while recording the loss, and then feeding the whole data again (without running the train_step, i.e., optimizing the network) in the second epoch.
Second, shouldn't I have a loss graph on the validation dataset that is decreasing almost all the time?(Shouldn't this be the case?)
Do the batch normalization and the dropout layers cause this behavior of the network?
Any help is much appreciated!!