Question for train session

rgeirhos / texture-vs-shape

Pre-trained models, data, code & materials from the paper "ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness" (ICLR 2019 Oral)

https://openreview.net/forum?id=Bygh9j09KX

Other

785 stars 101 forks source link

Question for train session #28

Closed jacob-kang closed 2 years ago

jacob-kang commented 2 years ago

I have a question about your experiment. Did you train your model with Stylized-ImageNet from scratch (I mean without ImageNet pretraine)? or Did you train the model with SIN and with ImageNet pretrained?

If you trained your model with SIN and with ImageNet-pretrained weight, Was there a moment which is the texture-bias saturated?

jacob-kang commented 2 years ago

My mistake. I didn't see the read me in detail. You used pretrained model for model A and B. Thanks!