search

keplr-io / quiver

Interactive convnet features visualization for Keras
https://keplr-io.github.io/quiver/
MIT License
1.75k stars 223 forks source link

Output predictions are incorrect #36

Closed harveyslash closed 7 years ago

harveyslash commented 7 years ago

I use quiver like this:

classList =['one','two','three'] server.launch(mmodel,classes=classList)

when testing images from the dashboard, the classes outputs are wrong. Most of the time it displays the same class predictions for images from different classes

jakebian commented 7 years ago

ping @kretes do you have time to take a look at this issue with code you introduced?

kretes commented 7 years ago

@harveyslash could you describe the issue in detail?

e.g. attach a screenshot of quiver showing prediction for some image + attach a code and an output showing what is the proper prediction for given image from your model

jakebian commented 7 years ago

I was getting this too on a VGG model, the default function loading images (@kretes you added this) in that case gives something that doesn't make much sense. Restored the old default loading function for now.

More generally we should allow passing this preprocessing function as an argument because for example, @kretes the function you added were to fit some other model you were playing with. I'll add that.

This issue should be solved. https://github.com/keplr-io/quiver/commit/2fe605e8520c49e15a75ef5f3d749fb23e53197b