The loss is not decreasing

1

I am training a neural nework the architecture is pretty simple ? conv layers with bach norm and 3 linear layers with relu , the loss is never decreasing but the accuracy is getting good (train loss and train accuracy are fine but test loss is not decreasing and test accuracy is not in the propoer shape .
test_accuracy

test_loss

the training loss and accuracy are :

train_accuracy
train_loss

2

@OBouldjedri without looking at further details, it would be difficult to ascertain the exact reason. Some pointers

  • Ensure that the distribution of labels in train and test data do not vary a lot. This shows up as train accuracy high but test accuracy low

  • Overfitting. Can you overlay the test as well as the train accuracy in the chart

1 Like
3

I uploaded what the train accuracy and loss looks like
.