Closed matthew-frank closed 1 month ago
It's v1.5. @sgpyc could you update the appropriate documentation?
I create PR: https://github.com/mlcommons/training/pull/605
I believe PR https://github.com/mlcommons/training/pull/590 (earlier https://github.com/mlcommons/training/pull/516) also addresses this issue.
Closing as the benchmark is dropped from Training benchmark suite
https://github.com/mlcommons/training/blob/master/image_classification/README.md#1-problem says
While https://github.com/mlcommons/training/blob/master/image_classification/README.md#structure--loss says
Please clarify in the README.md whether this is actually a v1 or a v1.5 ResNet-50? Perhaps it would help to move the discussion from https://github.com/mlcommons/training/tree/master/image_classification/tensorflow/official/resnet into this top level document?
As I understand it, the difference between "v1" and "v1.5" is typically that in the bottleneck blocks with downsampling v1 puts the stride 2 convolution in the first 1x1 convolution of the bottleneck, while v1.5 puts the stride 2 into the 3x3 convolution of the bottleneck. If this really is a v1.5 network then I believe a better reference for the modification to the v1 network described in Kaiming He's 2015 paper is the blogpost http://torch.ch/blog/2016/02/04/resnets.html.