0

I've trained a model for heart sound classification with transfer learning (MobileNet) on Physionet dataset, and it works fine.

However, when I train it on my own dataset, it seems that it can not learn anything: more specifically, the loss is not decreasing and the accuracy is not going up. I've checked my labels and they seem to be correct. What other things should I check?

nbro
  • 39,006
  • 12
  • 98
  • 176
Sepehr Golestanian
  • 31
  • 5
  • Can you clarify what you mean by "it seems that it cannot learn anything". How did you measure the learning rate? – Jean-Marc Volle Jul 13 '21 at 06:18
  • @Jean-MarcVolle it's loss is not decreasing and accuracy is not going up, unfortunately I don't have access to learning curves right now, otherwise I would have shared them. – Sepehr Golestanian Jul 13 '21 at 06:47
  • Go through these dot points and try each of them to see if they help: https://ai.stackexchange.com/a/26987/26726 – Recessive Jul 13 '21 at 07:28
  • @Recessive I think those points help if it was overfitting, the problem is it not overfitting, it is not learning at all! Accuracy and loss on the training data, itself, does not change! BTW, thanks. – Sepehr Golestanian Jul 13 '21 at 12:11
  • @SepehrGolestanian I've had similar issues with training and have had success trying those techniques. Honestly the biggest mistake most people make is not normalizing their data. Without any code or more details though, there is no way for us to help because it could literally be anything that's causing the problem. – Recessive Jul 14 '21 at 00:32
  • I always normalize my data, but I don't know what is the best way to normalize sound spectrum images. – Sepehr Golestanian Jul 14 '21 at 06:01
  • Could you please give more info about your dataset? What are the inputs and outputs? What is its size? Moreover, how exactly are you training the MobileNet with transfer learning on your dataset? Are you freezing the initial layers and only letting the last layers be learnable, i.e. fine-tuning MobileNet on your dataset? Please, edit your post to include these details directly in it. – nbro Jul 15 '21 at 12:53
  • @nbro Thanks for your comment. I will update the question later. – Sepehr Golestanian Jul 17 '21 at 14:17

0 Answers0