Getting started: Python and ML developers¶

This page turns a PyTorch model into a bundle that any Flutter app can load.

Install¶

pip install "modelport-cli[onnx,executorch,torchvision]"
modelport doctor

doctor lists which optional packages are installed and what to add for each feature.

Export¶

modelport export torchvision:mobilenet_v3_small --target onnx,executorch

This writes dist/mobilenet_v3_small/:

Sources can also be Hugging Face models (hf:facebook/deit-tiny-patch16-224) or your own code (file:my_model.py:build). See the CLI reference.

Make smaller variants¶

modelport quantize dist/mobilenet_v3_small --fp16 --int8

Each new variant is measured against PyTorch and gets a matching tolerance. See Smaller models.

Verify¶

modelport verify dist/mobilenet_v3_small

 variant                 ┃ output ┃ max diff ┃ cosine   ┃ top-1 ┃ result
━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━╇━━━━━━━━
 onnx-fp32               │ logits │ 3.34e-05 │ 1.000000 │ same  │ pass
 executorch-xnnpack-fp32 │ logits │ 3.34e-05 │ 1.000000 │ same  │ pass
 onnx-fp16               │ logits │ 1.05e-01 │ 0.999952 │ same  │ pass
 onnx-int8               │ logits │ 1.27e-01 │ 0.999933 │ same  │ pass

Publish¶

modelport publish dist/mobilenet_v3_small --github you/models --tag v1

hf auth login
modelport publish dist/mobilenet_v3_small --hf you/mobilenet_v3_small

Use it from Dart with types¶

modelport gen-dart dist/mobilenet_v3_small -o lib/models --location hf://you/mobilenet_v3_small

This writes a class whose inputs and outputs are named fields, so a typo in a tensor name is a compile error.