@dxo/resnet

Developer preview — API unstable. External Neural / weight-asset package (not inside dxo-framework publish).

create() returns @dxo/vision ResNet backbones. Runtime loading via @dxo/vision loadWeights accepts DXO-key safetensors only (same keys as new ResNet({ depth: 18 }).state()).

Torch / .pth → DXO

Do not expect loadWeights to rename torchvision keys. Convert offline with the external Python tool:

# Convert a trusted torchvision .pth directly to DXO-native safetensors:
python tools/convert_torchvision_resnet18.py --input weights/resnet18/model.pth --output weights/resnet18/model.safetensors --report weights/resnet18/conversion-report.json
# Then load the converted asset:
# loadWeights(backbone, { path: 'model.safetensors', scope: 'backbone' })

The mapping is in tools/dxo_resnet_mapping.py. It drops the classification fc.* head and emits one-based DXO canonical names. Python is used only by this external weight-asset toolchain; DXO runtime packages do not read .pth.

node --experimental-strip-types scripts/convert-map.test.ts

Link against local vision

pnpm link --global ../dxo-framework/projects/runtimes/dxo-vision
# or dependencies already use file:../dxo-framework/...