I use QAM for training, the number of iterations is 75K, but the loss only drops to about 0.5, is this normal?And how much does the final loss value of the original code drop? about qdmn HOT 12 CLOSED

Does the loss here refer to the total loss? Sorry, I can't remember the exact value. But 0.5 seems to be within the normal range. Is the 75k here the last stage? If so, maybe you can test the corresponding results after training.

from qdmn.

OK, thanks a lot. It's a total loss. I try the effect of the model

from qdmn.

Does the loss here refer to the total loss? Sorry, I can't remember the exact value. But 0.5 seems to be within the normal range. Is the 75k here the last stage? If so, maybe you can test the corresponding results after training.

In fact, there is another question, why is the value of score always 0 in the first stage of training process?Is this normal?

from qdmn.

This is not a normal phenomenon, you can check your code

from qdmn.

This is not a normal phenomenon, you can check your code

But when I use your code to run the first stage, the score is also 0, and the number of iterations is 10K, which is very strange

from qdmn.

This is not a normal phenomenon, you can check your code

The FC layer in QAM in the paper uses the sigmoid function. Why is Relu used in the code? Is it better?

from qdmn.

I run the first training stage with my code. Now the iteration is 7k and the score loss is about 0.7. There is no situation where score loss is always 0. The current code was re-implemented by me and there may be some inconsistencies, but I don't think this will cause a big difference.

from qdmn.

from qdmn.

In fact, the score mentioned above refers to the score obtained through QAM, not the score loss. When I debugged, I found that it was always 0 at the back. The loss I run is also not 0

from qdmn.

Sure, there is something wrong here, when using relu, the score is always 0. What you said above is correct, I change the activation function to sigmoid and the output score is normal now. This is something I didn't notice when I re-implemented the code, thanks for discovering this problem.

from qdmn.

I have fixed this issue. Since the current code was re-implemented after my internship ended, there may be some inconsistencies. Thanks again for finding this issue!

from qdmn.

ok,ok,thank you!

from qdmn.