lower batch size ： 2 labeled + 2 unlabeled -----> bad AP

facebookresearch / unbiased-teacher

PyTorch code for ICLR 2021 paper Unbiased Teacher for Semi-Supervised Object Detection

https://arxiv.org/abs/2102.09480

MIT License

415 stars 83 forks source link

lower batch size ： 2 labeled + 2 unlabeled -----> bad AP #53

Open Saixiaoma opened 3 years ago

Saixiaoma commented 3 years ago

Thanks for the great work. I have been training the code with your instruction in README.

Due to device limitations, there is only one GPU and only 12g of memory. In order to train, I can only adjust the batchsize to 2, which leads to very poor final results, with an AP of only about 11%.

I would like to ask whether batchsize has such a big impact on the result.

I'm looking forward to your reply !!

hachreak commented 3 years ago

Did you try to scale the learning rate?

Saixiaoma commented 3 years ago

Yes. I conducted two more experiments afterwards. One was to repeat the experiment without modifying any parameters, and the other was to modify the learning rate 0.01 to 0.0025. The results of both times were much worse than the initial one.