Skip to content

README.md

Latest commit

 

History

History
27 lines (20 loc) · 1.08 KB

README.md

File metadata and controls

27 lines (20 loc) · 1.08 KB
Raw

Vendored MODNet source

The files under src/ are vendored from the official ZHKKKe/MODNet repository (src/, Apache License 2.0) so that the ONNX export pipeline in onnx_tools/ can run without cloning the upstream project.

Only the backbone package required by onnx_tools/modnet_onnx.py is included:

src/models/backbones/
├── __init__.py       # SUPPORTED_BACKBONES
├── mobilenetv2.py    # MobileNetV2 implementation
└── wrapper.py        # MobileNetV2Backbone (enc_channels, forward, load_pretrained_ckpt)

The full model definition (src/models/modnet.py) and training code are not vendored — onnx_tools/modnet_onnx.py provides an ONNX-friendly replacement of the model.

Changes from upstream

Kept minimal so the code stays easy to diff against upstream:

  • Removed unused imports (json, functools.reduce).
  • torch.load(...) in the pretrained-backbone loaders now passes map_location='cpu' and weights_only=True (compatible with the PyTorch 2.6+ default and CPU-only machines).