Aidge Compression Module

A module for compressing Convolutional Neural Networks (CNNs).

This module decomposes convolutional layers into multiple smaller convolution operations, reducing the total number of computations and improving inference speed.

It provides a C++/Python interface for compression.

Installation

Work in progress — PyPI packages are not yet available.

From GitLab Package Registry (recommended)

The wheels are published to the Aidge GitLab Package Registry on every push to main (pre-release .dev0) and on tag commits (stable release).

Requires a GitLab personal access token with read_api scope (set as GITLAB_TOKEN):

export GITLAB_TOKEN=<your_personal_access_token>

# Install latest stable release
pip install aidge_compression \
  --index-url https://gitlab.eclipse.org/api/v4/projects/eclipse%2Faidge%2Faidge_compression/packages/pypi/simple

# Install latest pre-release (dev version)
pip install --pre aidge_compression \
  --index-url https://gitlab.eclipse.org/api/v4/projects/eclipse%2Faidge%2Faidge_compression/packages/pypi/simple

# Install a specific version
pip install aidge_compression==0.0.5 \
  --index-url https://gitlab.eclipse.org/api/v4/projects/eclipse%2Faidge%2Faidge_compression/packages/pypi/simple

The dependencies aidge_core and aidge_backend_cpu are also published to the monorepo GitLab Package Registry. Add the monorepo index as well so pip can resolve them:

pip install aidge_compression \
  --index-url https://gitlab.eclipse.org/api/v4/projects/eclipse%2Faidge%2Faidge_compression/packages/pypi/simple \
  --extra-index-url https://gitlab.eclipse.org/api/v4/projects/eclipse%2Faidge%2Faidge/packages/pypi/simple

From PyPI (planned)

pip install aidge_compression

From source

Requires aidge_core and aidge_backend_cpu to be installed first (from GitLab Package Registry, PyPI, or built from source):

pip install aidge_core aidge_backend_cpu
pip install .

Requirements

System

• Python >= 3.10
• C++17 compiler (GCC, Clang, MSVC)
• CMake >= 3.22
• OpenMP support

Python dependencies (installed automatically)

Package	Version
numpy	>= 1.21.6
aidge_core	any

C++ dependencies (fetched automatically at build time)

Library	Version	Purpose
Eigen	3.4.0	Linear algebra
nlohmann/json	3.11.3	JSON parsing
pybind11	>= 3.0	Python bindings

Aidge ecosystem

Package	Role
aidge_core	Core graph, operators, tensor infrastructure
aidge_backend_cpu	CPU backend (required for compression evaluation)

Building from source

# With Python bindings (default)
pip install .

# C++ only, without Python
cmake -B build -DPYBIND=OFF
cmake --build build

# With tests
cmake -B build -DTEST=ON
cmake --build build
ctest --test-dir build

CMake options:

Option	Default	Description
PYBIND	ON	Build Python bindings
WERROR	ON	Treat warnings as errors
TEST	OFF	Build C++ unit tests
COVERAGE	OFF	Enable code coverage (GCC only)

Usage

Example usage with resnet18, where the first convolution is ignored, and manual ranks are assigned to specific layers (-1 ignores the layer, [0.,1.] represents a weakening factor):

import aidge_core
import aidge_onnx
import aidge_compression

model = aidge_onnx.load_onnx("resnet18.onnx")

ignores = {"conv1_Conv"}
ranks = {
    "layer1_layer1_1_conv1_Conv": -1.,
    "layer2_layer2_0_conv1_Conv": -1.,
}

aidge_compression.Compressor(ignores, ranks, 0.8).compress(model)

aidge_onnx.export_onnx(
    model,
    "resnet18_0.8.onnx",
    inputs_dims={
        list(model.get_input_nodes(aidge_core.InputCategory.Data))[0].name(): [
            [1, 3, 224, 224]
        ]
    },
    outputs_dims={list(model.get_output_nodes())[0].name(): [[1, 100]]},
    opset=18,
)

A more complete set of examples to use this module is provided in the examples/ folder, including a tutorial notebook.

Compression methods

This module compresses CNNs through tensor decomposition, replacing a single convolutional tensor with multiple smaller tensors, reducing computational complexity.

{width=50%}

The process starts with rank selection for each layer's decomposition:

• Without a training dataset: The module uses VBMF to estimate the rank for each layer.

• With a training dataset: It leverages inference on the dataset to assess the impact of decomposition on each layer.

• With a partial dataset: The method can still use the available data, but it's preferable to ensure the dataset represents various cases the network will encounter.

Following rank selection, each layer undergoes Tucker-2 decomposition or WeightSVD:

{width=50%}

• WeightSVD: Applied to 2D tensors, replacing a single tensor with two smaller ones.

• Tucker-2: Applied to 4D tensors, replacing a 4D tensor with two 2D tensors and a smaller 4D tensor.

TODO list

• Tucker-2 implementation.
• VBMF rank automatic selection.
• Weight SVD for linear tensors.
• Validation dataset aided rank selection.
• CP decomposition
• Automatic hybrid selection between CP / Tucker
• Robust network decomposition