Replicating pytorch reported accuracy for qat mobilenet v3

I’m trying to replicate the accuracy pytorch reported for the quantization aware trained mobilenet_v3 model. I am running it with fbgemm backend, 128 batch size and ran it for 80 epochs. My best model got an accuracy of 71.596%. Pytorch reports an accuracy of 73.004% on the same. Has anyone been able to reach this accuracy? If so, can you please share your configurations or help me figure out why I’m stuck at this accuracy? I’m running my code on a single Tesla V100 GPU. It takes ~45 mins for one epoch to complete.

Thanks much!

Here is the recipe: vision/references/classification at main · pytorch/vision · GitHub