Christian Bartz
Christian Bartz
Seems the shapes produced by the network are not as they should be. Are you using your own data?
Then you should check the number of classes your dataset has. Did you adjust the network, to fit to your number of classes? How large is your validation set?
How many different characters do you want to recognize? 3700 images is not to much for validation. Actually it should work... I'm not sure why it doesn't. You can, however,...
Yes, FSNS is organized in such a way that one sample is actually comprised of 4 samples. The code snippet you refer to handles this case. If the flag `uses_original_data`...
Do you want to know how we deal with the 4 different views at the same time? This is actually quite easy: 1. we take the 4 views and split...
do you have problems with the blank label as the result of your prediction? If so, you can just strip this label from your predicted word.
Which part are you interested in? The part where we split the input image into 4 independent images, or the part where the features of all four images are fused...
We split the views [here](https://github.com/Bartzi/see/blob/master/chainer/models/fsns.py#L667) and reunite them [here](https://github.com/Bartzi/see/blob/master/chainer/models/fsns.py#L256)
I'm sorry but I do not speak chinese and I'm definitely not able to read it :sweat_smile: could you ask your question again in English?
Which version of `Chainer` and `Cupy` are you using?