Different backbone used by MaskRCNN

[ShengYang1]

Hi,

MaskRCNN seems to have different backbone used by reference implementation (pytorch) and Google's Training v0.6 submission (TensorFlow). PyTorch's reference implementation is using ResNet50 v1, which has stride = 2 for 1x1 conv and 1 for 3x3 conv. Google's submission is using Resnet v1.5, which has stride = 1 for 1x1 conv and 2 for 3x3 conv.

So I want to clarify whether this is actually allowed (backbone could be different in closed division), or there are something I missed.

• Tensorflow: https://github.com/mlperf/training_results_v0.6/blob/master/Google/benchmarks/mask/implementations/tpu-v3-32-mask/mask_rcnn/resnet.py#L215-L229
• PyTorch: https://github.com/mlperf/training/blob/master/object_detection/pytorch/maskrcnn_benchmark/modeling/backbone/resnet.py#L262-L284 https://github.com/mlperf/training/blob/master/object_detection/pytorch/maskrcnn_benchmark/config/defaults.py#L264

[peladodigital]

In an effort to clean up the git repo so we can maintain it better going forward, the MLPerf Training working group is closing out issues older than 2 years, since much has changed in the benchmark suite. If you think this issue is still relevant, please feel free to reopen. Even better, please come to the working group meeting to discuss your issue