qualia-codegen-core 2.0.0

Generate portable C inference code from Keras or PyTorch model

Copyright 2021 (c) Pierre-Emmanuel Novac [e-mail] Université Côte d'Azur, CNRS, LEAT. All rights reserved.

Qualia-CodeGen-Core

Converts a pre-trained Keras .h5 or PyTorch model to C code for inference.

Generated C code uses channels_last data format.

Supported layers

• Activation: ReLU (combined to previous Conv1D, Dense, MaxPooling1D, AveragePooling1D AddV2), Softmax
• Conv1D: optional bias, valid padding only
• Dense: optional bias
• MaxPooling1D: valid padding only
• AveragePooling1D: valid padding only
• Flatten: implies reordering next layer's kernel for data format conversion
• ZeroPadding1D: combined with next Conv1D
• AddV2

Dependencies

python >= 3.9

Python:

jinja2
numpy

Keras

Python:

tensorflow >= 2.6.0
keras >= 2.6.0 

PyTorch

Python:

torch >= 1.8.0

Installation

pip install -e .

Usage

Generate C code from Keras .h5

qualia_codegen <model.h5> <output directory>

Use in your C code

Include the model: (can also be built as a separate object)

#include "model.h"

Allocate inputs and outputs arrays with correct dimensions. Remember that inputs must have channels_last data format.

Call it in your C code:

cnn(inputs, outputs);

Add the source file model.c to your build system. It includes all the other source files for layers, don't add these to the build system.

Examples

See the src/qualia_codegen_core/examples/Linux directory for a demo console application to evaluate model accuracy.

src/qualia_codegen_core/examples/qualia_codegen-NucleoL476RG contains an STM32CubeIDE project for the Nucleo-L476RG board that's currently broken due to some recent changes

Documentation

Nothing much…

Source tree

src/qualia_codegen_core/Allocator.py: manages activation buffer allocation. Tries to group all buffers into one, except when they cannot be overwritten (dependencies).

src/qualia_codegen_core/Converter.py: the actual conversion code, parses a Keras model and use the template file associated to each layer to generate C code. When weights have to be written, they are optionally quantized to fixed-point by setting the appropriate parameters of Converter constructor (see its definition)

src/qualia_codegen_core/Validator.py: work in progress, should contain functions to check if a model can be successfully converted. For now only check activation function.

src/qualia_codegen_core/assets/: contains the templates to generate C inference code

src/qualia_codegen_core/assets/layers/: contains the implementation of the various supported layers

src/qualia_codegen_core/assets/layers/weights: contains the support for the trainable layers weights