Training accuracy always 0? Need help please

1

I have done a lot of searching around trying to find answers to this issue that I face. I have tried reddit, SO and just searching on google but to no avail. I am having issues with my results showing the following:

https://pastebin.com/CbjY80Qe

This is my code that I am using right now:
https://pastebin.com/9QbpV4h8

I have tried changing the learning rate, batch size and I have also tried multiplying the train_acc and train_loss by 100 after dividing by the number of images. None of this works and still produces the unusual results.

Please help me figure out what I am doing wrong !!!

Thank you very much,

Aeryes.

2

Could you check the type of (prediction == labels.data)? If it’s a ByteTensor, could cast it to float before summing it?
Also could you divide by a float number, i.e. 4242. instead of 4242?

2 Likes
3

maybe make sure the types match in torch.sum(prediction == labels.data) and then when you normalize, cast it as float:

acc = torch.sum(prediction == labels.data).float() / number_of_examples * 100
3 Likes
4

Thank you so much. I did this and it seems to have done the job. I am now getting better readings !!!

5

This helps.
sum() worked before. but doesn’t work now. use torch.sum(), any differences between sum() and torch.sum()

6

There might be some other issue then. sometensor.sum() and torch.sum(sometensor) should yield the same results. Maybe you are calling .sum() on a non-pytorch tensor (e.g., a numpy array)?