kimhc6028 / relational-networks

Pytorch implementation of "A simple neural network module for relational reasoning" (Relational Networks)
https://arxiv.org/pdf/1706.01427.pdf
BSD 3-Clause "New" or "Revised" License
812 stars 160 forks source link

Train Relational Networks for 10 epochs #1

Closed xunhuang1995 closed 7 years ago

xunhuang1995 commented 7 years ago

I have trained RN for 10 epochs. The final test set accuracy is 73% for relational question and 72% for non-relational question. It seems that there is no significant improvement for relational questions.

kimhc6028 commented 7 years ago

CNN + MLP (Not RN) scored approximately 44% on relational questions. I guess relational networks does improves relational questions, but I might forgot or mistook some settings. I am working on it currently.

justinbuzzni commented 7 years ago

I have trained RN for 20 epochs. The last test set accuracy is 84% for relation question and 93% for non-relational questions.

kimhc6028 commented 7 years ago

@justinbuzzni Would you please PR the trained model after 20 epochs? This should be saved at /model/epoch_20.pth

kimhc6028 commented 7 years ago

I chaned the architecture, and the result is written in the README. If any other things come out, please reopen this issue.