resnet 32 or resnet 34

[CVHvn]

In this code, you use Resnet34. so the result in paper is resnet34 or resnet32? can you provide more accuracy of resnet34?

[ynu-yangpeng]

Thank you for your question. I have reviewed the project code and found that the code for the network structure was not fully uploaded. I have now uploaded the complete code, which includes the implementation of Resnet-32 and Resnet-34.

Both Resnet-32 and Resnet-34 are from "Deep Residual Learning for Image Recognition." Since CIFAR has smaller image sizes compared to ImageNet, modifications have been made to the original Resnet to better suit CIFAR.

To the best of my knowledge, There are two main modification methods:

The first version is to remove the maxpool layer in the first convolutional module of the original Resnet-34 and use a 3x3 convolution with a padding value of 1.

The second version is specifically designed for CIFAR-10, with Resnet-(32, 44, 56, 110, 1202), but since the parameter volume of Resnet-32 is small (only 0.46M), its performance is relatively poor on the CIFAR dataset. We tested that setting the initial dimension to 64 in Resnet-32 improves its performance.

After our testing, there is no significant difference in performance between the two modified versions on CIFAR.

Our paper reports the results of Resnet-32, but it is worth noting that the input dimension of Resnet-32 has been modified to 64, which can be found in the code file. We have also added the accuracy of Resnet-34. The accuracy of Resnet-34 on CIFAR-100 is:

IF=100: ALL-ACC: 55.48% IF=50: ALL-ACC: 61.09% IF=10: ALL-ACC: 70.81%

Thank you again for your inquiry, which has made our code more rigorous!

[CVHvn]

Thank you for helpful response. I will close this issue!