search

csyanbin / TPN-pytorch

Pytorch Code for ICLR19 paper: Learning to Propagate Labels: Transductive Propagation Network for Few-shot Learning.
175 stars 29 forks source link

Accruracy of TPN-semi on miniImagenet #3

Closed bfshi closed 4 years ago

bfshi commented 4 years ago

Hi, congrats on the work! I got some problem when running TPN in semi-supervision setting on miniImagenet. I split each class into 40% : 60% for labeled and unlabeled data, and train with extra 5 unlabeled data for each support class (same as the setting in https://arxiv.org/abs/1803.00676). However, I only got accuracy of ~51.3, which is 1.5 lower than your paper. Do you know why this may happen?

csyanbin commented 4 years ago

For few-shot learning, the test accuracy has some variance, depending on different test episodes selected. One suggestion is to early-stop the training using the validation set, e.g., evaluate the model on checkpoint that has the best validation accuracy.

bfshi commented 4 years ago

THanks. I've tried testing the model with best val result. However, the situation still exists. Btw, there is a accuracy gap of 2%~5% between val and test result. Is that normal? THanks for helping.

csyanbin commented 4 years ago

The gap between val and test result is normal. Can you leave your email? I can send the my tf code for semi-supervised exp.

bfshi commented 4 years ago

Ok thanks! my email is 1159516387bs at gmail dot com

csyanbin commented 4 years ago

Sent.

bfshi commented 4 years ago

Thank you kindly!

Hugo101 commented 2 years ago

Hello! I am sorry to bother the closed issue. Can I have the TPN-semi code as well? I am not sure how to modify the TPN code to TPN-semi. Thank you!

csyanbin commented 2 years ago

Hi @Hugo101, can you leave your email?

Hugo101 commented 2 years ago

Hi @csyanbin , my email is lichangbin101@gmail.com Thank you so much!

abbod commented 1 year ago

Hello @csyanbin
Please give me the code of TPN-semi my email: abdoh86@gmail.com thanks

Tobu76 commented 1 year ago

Hello @csyanbin, I am very interested in your work on TPN-Semi and would like to learn more about it. Would it be possible for you to share the source code with me? I would greatly appreciate it. My email is jingmingxu97@gmail.com

Best regards.