Ho do I make inference for textrec after training?

[stasbel]

@Bartzi So I made my own dataset, do the training and then trying to make an inference for new image. How do I do that? It turns out that my log file lack of certain attributes if I trying to use text_recognition_demo.py:

Traceback (most recent call last): File "see/chainer/text_recognition_demo.py", line 150, in target_shape = Size._make(log_data['target_size']) KeyError: 'target_size'

[Bartzi]

That's odd... the log file should be created after the first 100 iterations and contain the necessary data. Is your dataset large enough for the training to still be in epoch 0 at iteration 100? This could be the reason why this information is not logged.

[stasbel]

Hi, thx you for the response! I use exact same data set you provided at https://bartzi.de/research/see. Training stage goes fine and creates train log afterwards at specified dir, but, unlike log at provided dataset, it lacks certain attributes and I don't know why.

[Bartzi]

Hmm, I did not provide a dataset for text recognition, as there is not enough space to host that data there. which dataset are you using? And did you check that you are still in epoch 0 after 100 iterations?

[stasbel]

I talk about text_recognition_model.zip at https://bartzi.de/research/see. No, after 100 iterations I already got 10 epochs, that's may be the problem. Gonna retrain it soon.

[Bartzi]

Ok, you can change this line to: if stats_cpu['iteration'] == args.log_interval: and it should work.

[FelixSchwarz]

That helped me too - would you mind updating the repo or should I create a pull request?

[Bartzi]

I'd be happy merging your PR =)