6.1. Overview

Converter is a tool which automatically convert DNN graph representation into executable binary working in multi device environments throughout C source codes. Also, it can import blueoil’s config and exported test data in order to control runtime binary behavior automatic test to guarantee the output must return the same result than that of the input.

converter contains following 3 separable steps.

  1. Import blueoil’s DNN graph and load graph as converter graph

  2. Optimize converter graph for the target run-time device

  3. Generate codes which will be compiled to the target device


6.2. Current acceptable graph representation

  • Tensorflow protocol buffer


6.3. Installation

6.3.1. Setup.py installation

To install converter run the command:

$ python setup.py install
[long output, could take some minutes depending on your environment]

If you fail in installing python packages like numpy, tensorflow, etc., try upgrading setuptools first, then install converter again.

$ pip install --upgrade setuptools

Now you should be able to successfully run:

$ python blueoil/converter/generate_project.py --help
Usage: generate_project.py [OPTIONS]

Options:
...

6.4. Getting Started

generate_project.py is a script which automatically runs all of above steps in overview. It imports blueoil’s TensorFlow and generate all cpp source headers and other control files like Makefile. You can quickly carry out this conversion with a classification sample: examples/classification/blueoil_quantize_cifar10/minimal_graph_with_shape.pb.

CAUTION This Getting Started is intended only working on CPU not FPGA. If you need to run the generated binary on FPGA, please follow Custom project generation & Optimizations section.

>> python blueoil/converter/generate_project.py -i examples/classification/blueoil_quantize_cifar10/minimal_graph_with_shape.pb -o tmp/ -p classification
>> cd tmp/classification.prj
>> make lib_x86 -j 4
>> python utils/run_test.py  -i <path to input image>/raw_image.png -l <path to generated library>/libconverter_x86.so -e <path to expected output>/xxx_output\:0.npy

Here the npy file xxx_output\:0.npy starts with a certain number, followed by “_output”.

The description of each step in the script are available below, which means you can run separately each of them.

6.5. Custom project generation & Optimizations

There are several ways to generate a custom project. Some of them must be applied when you want to run the program on FPGA. Others are mainly for optimizing the execution like acceleration and memory usage reduction.

6.5.1. Hard Quantize

This optimization is necessary to execute the graph on FPGA because this convert the float quantized weight while training, into integer one without breaking any consistency of the result. This is available by taking -hq flag to generate_project.py program which you can also see in Getting Started section.

6.5.1.1. example

>> python blueoil/converter/generate_project.py -i examples/classification/blueoil_quantize_cifar10/minimal_graph_with_shape.pb -o tmp/ -p classification_hq -hq