Experimental tools for converting PyTorch models to ONNX
Project description
PyTorch to ONNX Exporter
Experimental torch ONNX exporter.
[!WARNING] This is an experimental project and is not designed for production use. Use
torch.onnx.export
for these purposes.
Installation
pip install torch-onnx
Usage
import torch
import torch_onnx
from onnxscript import ir
import onnx
# Get an exported program with torch.export
exported = torch.export.export(...)
model = torch_onnx.exported_program_to_ir(exported)
proto = ir.to_proto(model)
# This will give you an ATen dialect graph (un-lowered ONNX graph with ATen ops)
onnx.save(proto, "model.onnx")
Design
{ExportedProgram, jit} -> {ONNX IR} -> {torchlib} -> {ONNX}
- Flat graph; Scope info as metadata, not functions
- Because existing tools are not good at handling them
- Eager optimization where appropriate
- Because exsiting tools are not good at optimizing
- Drop in replacement for torch.onnx.export
- Minimum migration effort
- Use ExportedProgram
- Rely on robustness of the torch.export implementation
- This does not solve dynamo limitations, but it avoids introducing additional breakage by running fx passes
- Build graph eagerly, in place
- Expose shape and dtype information to the op functions; build with IR
Why is this doable?
- We need to verify torch.export coverage on Huggingface Optimum https://github.com/huggingface/optimum/tree/main/optimum/exporters/onnx; and they are not patching torch.onnx itself.
- Path torch.onnx.export such that packages do not need to change a single line to use dynamo
- We have all operators implemented and portable
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
torch_onnx-0.0.5.tar.gz
(25.7 kB
view hashes)
Built Distribution
torch_onnx-0.0.5-py3-none-any.whl
(29.0 kB
view hashes)
Close
Hashes for torch_onnx-0.0.5-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 98256dd010264c6432a29befb2befffd98a811634d48466a6ebf0e78d40ef7a4 |
|
MD5 | 8fa13f6062228b46d57d7e6416f8222f |
|
BLAKE2b-256 | 44f597cd4ba13d0de862010e31ee73f24890dd66309b7e18ba13ec9b255a8f22 |