Skip to content

Latest commit

 

History

History
103 lines (70 loc) · 3.15 KB

usage.md

File metadata and controls

103 lines (70 loc) · 3.15 KB
Raw

Example usage

This document assumes you have successfully built the code by following set up guide.

Run existing ONNX models

In this section, we run an existing ONNX model by the runtime of Chainer compiler.

To run ShuffleNet model in the repository, run

$ make -C build
$ ./build/tools/run_onnx --device cuda --test data/shufflenet --trace

The command above uses inputs and outputs in data/shufflenet/test_data_set_? to feed and verify the model.

VGG19 works, too:

$ wget https://s3.amazonaws.com/download.onnx/models/opset_8/vgg19.tar.gz
$ tar -xvzf vgg19.tar.gz
$ ./build/tools/run_onnx --device cuda --test vgg19 --trace

You can run more models defined in ONNX's tests:

$ ./scripts/runtests.py onnx_real -g

Generate a training graph from your Chainer model

First prepare a model which outputs a loss value as a single float. Here we use ch2o/tests/model/Resnet_with_loss.py as a sample.

You can generate an ONNX graph for inference by

$ PYTHONPATH=ch2o python3 ch2o/tests/model/Resnet_with_loss.py resnet50

Your ONNX file should be stored as resnet50/model.onnx. This ONNX file contains the computation graph for inference. Then, you can generate a training graph by

$ ./build/tools/run_onnx --onnx resnet50/model.onnx --out_onnx resnet50/backprop.onnx --backprop --compile_only

There would be a bunch of ways to analyze resnet50/backprop.onnx. For example, if you want a text dump, run

$ ./build/tools/dump resnet50/backprop.onnx

Or investigate it programmatically by Python:

>>> import onnx
>>> model = onnx.ModelProto()
>>> model.ParseFromString(open('resnet50/backprop.onnx', 'rb').read())
>>> [o.name for o in model.graph.output]

You can also use visualizers for ONNX such as netron.

Use chainer-compiler from Chainer

To use chainer-compiler from Chainer code, you first need to install Chainer from source code, for example:

$ CHAINER_BUILD_CHAINERX=1 CHAINERX_BUILD_CUDA=1 MAKEFLAGS=-j2 pip3 install --user third_party/chainer

You also need to rebuild chainer compiler with three flags:

$ cd build
$ cmake -DCHAINER_COMPILER_ENABLE_PYTHON=ON -DPYTHON_EXECUTABLE=/usr/bin/python3 -DCHAINERX_BUILD_PYTHON=ON ..
$ make
$ cd ..

Check if everything is fine by running pytest:

$ pytest python

You need to wrap your model by chainer_compiler.compile. E.g.,

model = YourModel()
model = chainer_compiler.compile(model, dump_onnx=args.dump_onnx)

See examples directory for more details. You can run the MNIST example by

$ python3 examples/mnist/train_mnist.py --compile --dump_onnx -d cuda

Use with SNPE

  • Set $SNPE_ROOT to the directory of extracted snpe-1.x.x.zip
  • Pass -DCHAINER_COMPILER_SNPE_INCLUDE_DIR=$SNPE_ROOT/include -DCHAINER_COMPILER_SNPE_LIBRARY_DIR=$SNPE_ROOT/lib/x86_64-linux-clang to cmake. Replace x86_64-linux-clang with target architecture for ChxVM runtime when needed