search

ryanjcooper / EMNIST

A project designed to explore CNN and the effectiveness of RCNN on classifying the EMNIST dataset.
MIT License
95 stars 37 forks source link

[NON-ISSUE] Cross validating training result #5

Open louisgv opened 7 years ago

louisgv commented 7 years ago

Hello,

I would like to cross validate my loss and accuracy with other. Below is my result:

image

ryanjcooper commented 7 years ago

Hi @louisgv, your results seem fairly consistent with my training/test accuracy (~86%). I will be working on a deep-model extension to this project this weekend (hopefully), and I will report my findings.

It seems that creators of this data set reported 69.71% ± 1.47% (see table 3 from here) accuracy using an OPIUM-based model, so this method beats that already. I have no benchmark for state-of-the-art (I would assume that >95% could be achieved using DCNNs).

I will leave this thread open for people to report any model progress or accuracy improvements.

Thanks for your interest!

EDIT: I also noticed that they reported a peak accuracy of 77.57% ± 0.08% on the by_class set using a deep OPIUM-based classifier, should be noted that they claim that they believe higher accuracy can be achieved.

jamesyzhb commented 6 years ago

why image

jamesyzhb commented 6 years ago

thanks

ryanjcooper commented 6 years ago

Hello @jamesyzhb, please open a new issue on the repository, but I believe it is an issue with a version mismatch with this code and Keras. What version of Keras and TensorFlow/Theano are you using?

jamesyzhb commented 6 years ago

Is a FLask version issue, thanks for the reminder