Skip to content

Latest commit

 

History

History

resnet50

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
tests
tests
 
 
README.md
README.md
 
 

README.md

ResNet50 Demo

Introduction

ResNet50 is a deep convolutional neural network architecture with 50 layers, designed to enable training of very deep networks by using residual learning to mitigate the vanishing gradient problem.

Details

  • The entry point to the Metal ResNet model is ResNet in ttnn_functional_resnet50_new_conv_api.py.
  • The model picks up certain configs and weights from TorchVision pretrained model. We have used torchvision.models.ResNet50_Weights.IMAGENET1K_V1 version from TorchVision as our reference.
  • Our ImageProcessor on the other hand is based on microsoft/resnet-50 from huggingface.

Performance

T3000

End-to-End Performance

  • For end-to-end performance, run
  WH_ARCH_YAML=wormhole_b0_80_arch_eth_dispatch.yaml pytest models/demos/t3000/resnet50/tests/test_perf_e2e_resnet50.py::test_perf_trace_2cqs
  • This will generate a CSV with the timings and throughputs.
  • Expected end-to-end perf: For batch = 16 per device, or batch 128 in total, it is about 32,250 fps currently. This may vary machine to machine.