I’m making a genre classifier using the GTZAN dataset, where I’m generating overlapping patches of Melspectrograms for each input. The final output of the model is the sum (or average) output for each patch.
Jupyter notebook with outputs
The problem is that after the second batch or so the network is not deviating from 1 class prediction (even with ‘average’ combination for the final prediction). The output from the model is also weird, because the predictions are identical for each input of the batch. Same with no activation on the output.
The network is 2 conv2D layers with maxpooling and a fully connected one with 10 outputs.
Optimizer: Adam with weight decay (5e-4) and lr (0.01)
Is the network built wrong? All the help would be much appreciated!
EDIT: Fixed the issue, remember to activate your fully connected layers kids