PyTorch support

[lutzroeder]

PyTorch and TorchScript support multiple serialization formats. Based on which format is used, a .pth or .pt file might contain only tensor data, a partial module hierarchy without graph connections or a TorchScript serialized graph.

• torch.save(model.state_dict(), file) - the app will load the tensors, grouping names by modules
• torch.save(model, file) - the app will show the module hierarchy without node connections
• torch.jit.script(model).save(file) - the app will try to extract and render the TorchScript graph, see #842.
• torch.utils.tensorboard.SummaryWriter - the app will open the graph from the events file, see #1024.
• torch.export - the app loads and renders the exported .pt2 graph, see #1211

PyTorch support is experimental. The recommended approach is to save the model to ONNX.

pytorch/pytorch#31877 for discussion of PyTorch container formats.

Examples:
alexnet.pkl.pth.zip
alexnet.zip.pth.zip
densenet161.zip.pth.zip
inception_v3.pkl.pth.zip
inception_v3.zip.pth.zip
resnet101.pkl.pth.zip
r3d_18.zip.pth.zip
shufflenet_v2_x1_0.zip.pth.zip
squeezenet1_1.pkl.pth.zip
valid-bert-base-uncased.zip
ENet.pth.zip
pytorch_invalid_file.zip
tutorial_models.zip via pytorch-tutorial
labels.pth.zip
mnist_bfloat16.pt.zip
fruit_veg_model.pkl.zip
yolov5n.tensor.pth.zip
from_numpy.pth.zip
data.pkl.zip
yolov5n-0.5.pt.zip
torch_fx_sample.pt.zip
tensors.data.pkl.zip
densenet.data.pkl.zip
fast.ai.data.pkl.zip
yolov5n.tensor.data.pkl.zip
vicuna-7b.zip
bad-hands-5.pt.zip
tensors.pkl.zip
cup_wild_vit_l_1img.ckpt.zip
hrnet_posenet_FP32.pth.zip
InternVideo2-stage2_1b-224p-f4.pt.zip
cloudpickle.pth.zip

Tools:
pytorch_strip_weights.zip