parameter stem_multiplier and module stem
xjhcassy opened this issue · comments
Hi, @chenxin061 ,Thanks for sharing the work. I have some questions. Could you please explain the meaning of the parameter "stem_multiplier" and the benefit of module stem. Thank you very much.
BTW，why did you only restrict the number of skip-connect in the normal cell?
The stem convolution is widely used in networks for image classification, including ResNet, MobileNet. etc. It converts the input image into raw feature maps for further processing (convolution).
The reason for restricting the number of skip-connect operations is clearly stated in our paper. Please refer to our paper for more information.
@chenxin061 ，Thanks for your reply. I have another question about the code. It;s in the operations.py. Why did you change the Zero operations? Are there any differences? Thanks.
@xjhcassy Functionally there is no difference between there two implementations.
Hi, @chenxin061, Have you experimented with different values of the parameter "stem_multiplier", I tried stem_multiplier=1, 2, and 3. And I find that the result is 1>2>3. What do you think of this situation? Thanks
@xjhcassy Glad to hear that you tried new settings and got good results.
The ''stem_multiplier'' is related to the output channel of the first convolution in the network. If the improvement you got was consistent and was validated by running multiple times to get rid of the randomness on CIFAR, it would be an interesting phenomenon and worth further exploration.
We followed the setting of DARTS and did not modify the network configuration for a fair comparison.