recognition_loss

[DaPenggg]

HI , I find that the recognition loss: loss = self.ctc_loss(pred[0].float(), pred[1].int(), gt[0].int(), gt[1].int()) but ,I think ,the correct is: loss = self.ctc_loss(pred[0].float(), gt[0].int(),pred[1].int(), gt[1].int())

[DaPenggg]

when ,i train the icdar2015 datasets, the loss is above 1000 basically.

[jiangxiluning]

@DaPenggg hi, the ctc implementation I used is https://github.com/t-vi/warp-ctc, you can check the signature. https://github.com/t-vi/warp-ctc/blob/6f3f1cb7871f682e118c49788f5e54468b59c953/pytorch_binding/warpctc/init.py#L90