Utilities for converting PyTorch Hub models to TorchScript IR.
Generates an inlined IR file for each model, as well as a
submodule.tsir
file containing the graphs for each submodule in the
model. Optionally saves .pt
files with the scripted/traced models.
Note: The submodule.tsir
file has all occurances of
".___torch_mangle_[0-9]+"
removed so that otherwise duplicate graphs
can be filtered out.
# Optional: env setup.
python3 -m venv "${HOME?}/.venv/torchscript_ir"
source "${HOME?}/.venv/torchscript_ir/bin/activate"
# Install dependencies.
python3 -m pip install -r requirements.txt
# Generate IR for all TorchVision models under `tsir/torchvision/`
python3 generate_torchvision_ir.py
# Save scripted models to disk
python3 generate_torchvision_ir.py --models alexnet,vgg11 --save_models
// Included vscode settings:
"python.formatting.provider": "yapf",
// The IR is close enough to python that this is helpful for reading it.
"files.associations": {"*.tsir": "python"},