zero loss

[ShangwuYao]

I am getting zero loss in calculating the CTCLoss, I don't know if it is supposed to happen. logits, labels, input_length, labels_length could be found in https://github.com/ShangwuYao/support-issues/tree/master/warpctc

[Jackson-Kang]

Same error, too..

Did you solved the problem?

[guilk]

@ShangwuYao Did you solve this problem? My training loss does not decrease and all predictions are _blank characters.

[ShangwuYao]

No, not yet

[Jackson-Kang]

Try to run the example written in README.md. Is loss still 0?

[ShangwuYao]

No the example works fine.

[FAhmedB]

This happened to me when the number of features is less than the number of labels

[simpleoier]

It seems that I also met the same case with @FAhmedB , when the seq-length of features is less than the seq-length of labels.

[MarStarck]

same problem. but if don't use cuda the loss is normal. so I think it may be environment problem. I'm using cuda8.0, pytorch 0.4.1

[MarStarck]

I installed cuda 9.0 with cudnn 7, pytorch 0.4.0 and remake and reinstall it and it is not 0 anymore!

[jordanott]

@ShangwuYao Did you solve this problem? My training loss does not decrease and all predictions are _blank characters.

My training loss fluctuates but all predictions are blank characters. Any suggestions?

[pxkc]

I've installed pytorch0.4.0 instead of the previous 0.2.0. And reinstalled the warpctc, and it worked like a charm.

[wuliebucha]

I have the same problem，conda environment(cudatoolkit-9.0、cudnn-7.1.2、pytorch-0.4.0、warpctc-pytorch-0.1、python-3.6.2 ) . I doubt it caused by environment, then i tried many version of toolkit, such python( 3.5、3.6) 、cuda(8、9)、cudnn（6、7）、pytorch（0.4、0.41、1.0）and so on, but still can't resolve it, any one have ideas？

[niuwagege]

Maybe you can try this. It works for me!

[kokoyy]

Maybe you can try this. It works for me!

Thanks, it works for me.