Training model on japanese dataset

[Preethse]

I created my own dataset of 14M images and started training on it. Weirdly the results are like this after first epoch

Really would appreciate feedbacks to improve my training.

[Chenxx017]

Hi, can this model be used to identify languages, such as Chinese, Japanese, etc.

[lerndeep]

https://github.com/bharatsubedi/PARseq_torch check in this repo you just update the character set of language you need you will able to train network.

[baudm]

@Preethse try early stopping. Your training is collapsing after a while. You could also try tuning the learning rates. The training dynamics drastically changes if you change the training data, so the default hyperpameters might not work well for you.

Also, for experimentation, use a bigger model, like the base Transformer configuration (d=768).

[baudm]

@Chenxx017 Sorry, but no. That's outside the scope of this project.