takashiishida / comp Goto Github PK

View Code? Open in Web Editor NEW

45.0 45.0 16.0 3.29 MB

[NeurIPS 2017] [ICML 2019] Code for complementary-label learning

License: MIT License

Python 100.00%

deep-learning machine-learning

comp's People

Contributors

Stargazers

Watchers

Forkers

liudefu stjordanis cao98 yoshum murphysong lifanghe gaoyi439 quekaixx theainerd murez maadsaifuddin53 hongxin001 pigcv sunshiding asionw yuz128pitt

comp's Issues

non_negative_loss: ccp[k] instead of ccp?

Hi authors:
Thanks for sharing your code. I have a question regarding your implementation of non_negative_loss. Based on the formulation (14) in your paper, should it be torch.mul(ccp[k], ...) instead of torch.mul(ccp, ...)? Refer to https://github.com/takashiishida/comp/blob/master/utils_algo.py#L23

Thanks in advance.

forward loss: normalize the score vector before applying the transition matrix Q

The forward loss of Yu et al. (2018) assumes that the classifier g(x) outputs a probability vector with length K. In your implementation, f, the output of linear_model or mlp_model, is an unnormalized score vector. However, the forward_loss function directly applies the transition matrix Q to f, leading to a disparity between the algorithm and the implementation.
https://github.com/takashiishida/comp/blob/master/utils_algo.py#L37

Recommend Projects