Does bonnetal classification module support multilabel classification?

PRBonn / bonnetal

Bonnet and then some! Deep Learning Framework for various Image Recognition Tasks. Photogrammetry and Robotics Lab, University of Bonn

MIT License

233 stars 60 forks source link

Does bonnetal classification module support multilabel classification? #7

Closed newforrestgump001 closed 5 years ago

newforrestgump001 commented 5 years ago

Does bonnetal classification module support multilabel classification? Thanks a lot! Jack

tano297 commented 5 years ago

Hi, As it is, it doesn't. BUT, it should be straightforward to implement. Changes that would be necessary:

During training: changing Softmax and CrossEntropy with Sigmoid and BinaryCrossEntropy
During inference: changing Softmax for Sigmoid, and replacing the Argmax operation (which gets only the highest class), for a threshold in the confidence of your choice, and then a selection of all the classes that correspond with this criteria.

If you're interested in making this happen neatly, I can give you pointers, and maybe you can submit a PR.

Best,

newforrestgump001 commented 5 years ago

Thanks a lot for your information. My filelist.txt is like this: 4590.jpg 1 1 0 0 7331.jpg 1 0 1 0 4941.jpg 0 1 1 0 1030.jpg 0 1 1 0 3094.jpg 1 1 1 0 8528.jpg 1 0 0 0 1314.jpg 1 0 0 0 5335.jpg 0 1 1 0 4097.jpg 0 0 0 1 7059.jpg 1 0 0 1 2858.jpg 1 1 1 1 2972.jpg 0 0 1 0 2070.jpg 0 0 1 0 3927.jpg 1 0 0 1 3106.jpg 1 0 0 0 6214.jpg 1 1 0 0 5578.jpg 0 1 0 0 3766.jpg 1 1 1 1 6954.jpg 1 1 1 1 6997.jpg 1 1 1 1 8901.jpg 0 1 0 0 632.jpg 1 0 1 0 1799.jpg 0 0 1 0 5157.jpg 1 0 1 0 2456.jpg 0 1 1 0 7542.jpg 1 1 1 1 9093.jpg 0 1 1 0 2956.jpg 0 1 0 0 680.jpg 1 0 1 0 1441.jpg 1 1 1 1 Could you give me some pointers about how to modify the framework on the main points? Thanks a lot!

newforrestgump001 commented 5 years ago

@tano297 Could you kindly give me some pointers for the modification of the classification module to support multi-label classification, I'm new to bonnetal framework. Thanks a lot!

tano297 commented 5 years ago

criterion needs to be modified to allow for multiple labels. See pytorch BCEWithLogitsLoss
user.py tok's needs to be replaced with a threshold and a torch.nonzero in order to get all the positive labels.
Your parser needs to output the labels in your multilabel format, rather than as a single number, in order to make the BCEwithlogitsloss work properly

newforrestgump001 commented 5 years ago

@tano297 Thanks for your help, I will have a try. Jack

tano297 commented 5 years ago

I'll close this for now, let me know if you need more pointers

newforrestgump001 commented 5 years ago

@tano297 Thanks, you're very nice! Jack