Closed teemiik closed 5 years ago
@weinman I can make you this scenario, because I still need it, if the neural network shows a good result, I'll put it here, but I need the training checkpoint files for test.
@weinman I just don't have a chance to train her for the test yet.
Using the tfrecord files included in the repo, you should be able to run one (or even zero) rounds of training, even on a CPU to get model.ckpt-00000 to work from.
@weinman but in this case, I do not know in practice how well the network works, you have even examples of work I do not see here
I'm simply pointing out everything is already provided to produce a simple training checkpoint, enough to write a script that produces a stored model (even if the model is terrible). I will gladly use such a script to produce and link to a well-trained mjsynth-based model.
@weinman Well good, is your model trained for English only? is it possible to support multiple languages?
@teemilk Can you read about mjsynth before you ask any more questions.
The model checkpoint (trained using MapTextSynth) from the ICDAR 2019 paper is posted here: https://weinman.cs.grinnell.edu/data/cnn_lstm_ctc_ocr_checkpoint_icdar19.tar.bz2
PR welcome for a script that will save the model (not as a training checkpoint).