Add PyTorch TorchScript backend for vision benchmarks

[amd-arsuresh]

Summary

Add a new pytorch-torchscript backend for the vision classification and detection benchmarks that loads TorchScript models exported via torch.jit.trace or torch.jit.script.

Problem

The existing resnet50-pytorch profile in SUPPORTED_PROFILES incorrectly sets "backend": "tensorflow", making it non-functional. The pytorch-native backend uses torch.load() and then tries to access ONNX graph attributes (model.graph.initializer, model.graph.input), which fails on TorchScript models; and the pytorch backend uses caffe2/ONNX, requiring an ONNX model file rather than a native PyTorch format.

There is no backend that can load standard TorchScript .pt files (the most common way to serialize PyTorch models for deployment).

Changes

• New file: vision/classification_and_detection/python/backend_pytorch_torchscript.py
  • BackendPytorchTorchScript class that uses torch.jit.load() to load TorchScript models
  • Supports both single-tensor and tuple/list outputs
  • Automatic CPU/CUDA device selection
• Modified: vision/classification_and_detection/python/main.py
  • Register pytorch-torchscript in get_backend()
  • Fix resnet50-pytorch profile: change backend from tensorflow to pytorch-torchscript

Usage

# Export a TorchScript model
python -c "
import torch, torchvision
model = torchvision.models.resnet50(pretrained=True).eval()
traced = torch.jit.trace(model, torch.randn(1, 3, 224, 224))
traced.save('resnet50_torchscript.pt')
"

# Run with the new backend
cd vision/classification_and_detection
./run_local.sh pytorch-torchscript resnet50 cpu --scenario Offline

Testing

Tested with resnet50 TorchScript model on CPU - loads and runs inference correctly.

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅