train loss is always 0 - Githubissues

MhLiao / TextBoxes_plusplus

TextBoxes++: A Single-Shot Oriented Scene Text Detector

Other

953 stars 279 forks source link

train loss is always 0 #75

Closed TLedTS closed 6 years ago

TLedTS commented 6 years ago

As the title said, when i train the net with ICDAR2015 dataset, the loss is always 0. Could anyone help me with this problem? Thanks a lot. Btw, i didn't compile the crnn cause it always went wrong. I just create data under the guide this page said and run the train order.

TLedTS commented 6 years ago

And this is my xml file. @MhLiao Could you please give me a favor?

MhLiao commented 6 years ago

@TLedTS Difficult should set to 0, otherwise, it would be ignored when calculating the loss.

TLedTS commented 6 years ago

@MhLiao OK, because I don't care about the content, I thought it will be okay if I set all of difficult to 1 as your example shown.
Now, i fixed it. Thanks a lot! :)