Training loss and val loss decrease normally, but accuracy is always 0

[meijieru]

Training loss and val loss decrease normally, but accuracy is always 0

[abhiML]

did you run it with adadelta?

[yisongbetter]

When I use the pretrained model @meijieru offers, the problem gone, so maybe I have not get enough data to train my model from scratch.

[yisongbetter]

I have solve the problem,adadelta or adam is ok @abhiML. I use 800 pictures to train crnn from scratch, which is stupid , when I use the retrained model, the problem disappear.
Here is what I got:

[abhiML]

Well I was using adam on the Synth dataset. It showed me infinite loss. But when I used adadelta it ran perfect. I did not use the pretrained model.

[yisongbetter]

@abhiML It should be the learning rate problem that bothers you, because adadelta requires no manual setting of a learning rate, so when using some momentum methods like adam, you should carefully set the initial learning rate, 1e-3 works well in my situation.

[meijieru]

Since it has been solved, I closed this issue temporally. Feel free to re-open it when necessary.

[SenWang-NEU]

I also encountered the same situation, because my data set was self-built. Is there any other way besides loading the pre-trained model?