Christian Bartz

Oh okay, I suggest not mixing the `easy` and `centered` datasets, because they are visually very different! Use either `easy` or `centered` but not both for the same model. Your...

You tried to use the FSNS model for your image? Yeah that won't work -.- deep learning systems are unfortunately not good enough for such kind of transfer (at least,...

The FSNS model is able to recognize latin characters, but the smple you are using, is totally different from everything the model has seen before...

you forgot to add metainformation to the label_file. This is described [here](https://github.com/Bartzi/see#preparations).

Yes, the input size is defined by the variable `image_size` [here](https://github.com/Bartzi/see/blob/master/chainer/train_svhn.py#L54) all images will be resized to this size. They will also be normalized automatically, so there shouldn't be anything...

Yes, you'll need to create the `curriculum.json` on your own, as indicated in the `README`. You should then point to the location of the file `curriculum.json` instead of using the...

The shape of the arrays does not fit, as we can see from the error. What was the command line call you used to start the training? It is likely...

No, there should be no restriction of punctuations. It might be difficult for the network to learn to predict the punctuation, but this might be circumvented with extra training data...

Hmm, I don't think it would work well... you see one of the problems with the approach is that it is currently not really possible to work with images that...

As I said, I don't think it will work out of the box, as the proposed system has problems with many textlines in large position variations. The guidance obtained by...