Skip to content

Latest commit

 

History

History
56 lines (41 loc) · 2.02 KB

README.md

File metadata and controls

56 lines (41 loc) · 2.02 KB
Raw

Tract examples: ONNX MobileNet v2 optimized for batch

This project demounstrate how to use symbolic dimensions in tract while performing network optimisation. It builds upon the onnx-mobilenet-v2 example, but here we will optimise the network without assuming it will always run on one single image.

Assets

Model is the same as in the quick start example:

wget https://github.com/onnx/models/raw/master/vision/classification/mobilenet/model/mobilenetv2-7.onnx

We will use Grace Hopper portrait again, and a cat image.

Loading and optimising the model

    let batch = Symbol::new('N');
    let model = tract_onnx::onnx()
        .model_for_path("mobilenetv2-7.onnx")?
        .with_input_fact(
            0,
            InferenceFact::dt_shape(
                DatumType::F32,
                &[batch.to_dim(), 3usize.into(), 224usize.into(), 224usize.into()],
                ),
                )?
        .with_output_fact(0, InferenceFact::default())?
        .into_optimized()?
        .into_runnable()?;

The differences in model preparation is the most tricky part: we need to introduce a "variable" N to be use in shape computation. It represent the batch size, it's the N from NCHW. We could actually use any letter or charater, tract does not really know anything specific about batch size.

&[batch.to_dim(), 3usize.into(), 224usize.into(), 224usize.into()].

The shape definition, while a bit complex syntax-wise, just says "N,3,224,224".

One other trick is... this network actually encodes the batch has N in two places in the protobuf file: on the input (which we just override with our N-based input fact) but also in its output. So we need to "erase" that output shape information:

.with_output_fact(0, InferenceFact::default())?

We could also define it as N,1001, but we just let tract shape inference do the work.

Loading images and running

The rest is pretty straightforward. We make a tensor with two images, but we could also use the same optimised network with one image or any number...