search

hytseng0509 / CrossDomainFewShot

Cross-Domain Few-Shot Classification via Learned Feature-Wise Transformation (ICLR 2020 spotlight)
331 stars 62 forks source link

about training problem. #22

Closed DavidKong96 closed 4 years ago

DavidKong96 commented 4 years ago

Thanks for your sharing. There is a question about should i train the pre-trained encoder from 0 ,or i can transfer the weights of resnet10/18/34 to train the model. Will that have any impact? Such as decreased accuracy, overfitting？ And is ResNet34 performs worse than resnet10 or 18? I just got in touch with this cross-domain not long，hope you can reply.It will help me a lot. Thank you so much!

hytseng0509 commented 4 years ago

You should train the pre-trained encoder using only the training categories in our setting. As the pre-trained model by torchvision may use the categories that serves as our testing categories to train the model, you should train the encoder from scratch. According to Table A5 in this paper, the model size does not have a significant impact on the few-shot classification performance. Hope that answers your questions.

DavidKong96 commented 4 years ago

ok，thanks. i will try it .It's very helpful