dnnets 0.2.0

Running multilayer perceptrons on the CPU

dnnets

This is a lightweight and dependency-free library to run multilayer perceptrons on the CPU. Given a model and an input, it will compute the output of a forward pass. Nothing more. The computation of a single pass runs in a single thread. Generally, the library tries to make a good trade-off between simplicity and performance. A model can be provided as a JSON file (format: see below), DNNF file (format: see dnnf.md) or created dynamically at runtime. A model can also be exported to both JSON and DNNF.

Note: in this library, activation functions are also modeled as layers.

Installation

To build and install the shared library along with the header file and a pkg-config file to /usr/local (or another location of your choice), use the (default) install target:

zig build install -Doptimize=ReleaseFast
zig build install -Doptimize=ReleaseFast -p /usr/local

To automaticallly build a Python wheel with the provided build script, you need to have Hatch installed. For bundling the Python bindings together with the shared library into a wheel, use build-wheel-with-lib:

zig build build-wheel-with-lib -Doptimize=ReleaseFast

If you only want to include the bindings without the shared library, use build-wheel:

zig build build-wheel

In both cases, you will find the generated wheel in zig-out/dist.

If you'd like to use your own frontend for building the Python wheel, you can use the prepare-wheel-with-lib and prepare-wheel targets instead:

zig build prepare-wheel-with-lib -Doptimize=ReleaseFast
# or
zig build prepare-wheel

You can then navigate to zig-out/wheel-builds/<your target> and invoke the frontend of your choice there. Note that hatchling is still used as the backend. <your-target> is either any if the wheel does not contain the shared library or a combination of <CPU architecture>-<OS>-<ABI> otherwise.

If you need to cross compile/build the Python wheel, you can use build-all-wheels which builds the wheels for all available targets:

zig build build-all-wheels -Doptimize=ReleaseFast

Using from Zig

This package exposes a module called dnnets that you can use in your own Zig code. Simply add this package to the dependencies in your build.zig.zon. In your build file, you can add the module as an import to your own module with something similar to this:

const dnnets_pkg = b.dependency("dnnets", .{});
const dnnets_mod = dnnets_pkg.module("dnnets");
exe_mod.addImport("dnnets", dnnets_mod);

The API is available as an online documentation at https://hannesbraun.net/share/doc/dnnets/zig

This is a small example of its usage:

var model = try dnnets.json.load(allocator, "sample_model.json");
defer model.deinit();
var input = [_]f32{ -3.5987e-01, -4.6701e-01, 1.6226e+00, 5.3634e-02 };
const output = model.getOutputBuffer(); // The output buffer will be reused.
model.forwardPass(&input);

C interface

dnnets has a C-compatible ABI. Documentation for the C interface is available in the header at include/dnnets.h.

If you don't like browsing the header file directly, there's also an online version of this documentation available: https://hannesbraun.net/share/doc/dnnets/c

This is the example provided above adjusted for C (without error handling):

dnnets_init(); // Initialize library

struct dnnets_model *model;
dnnets_load_json(&model, "sample_model.json");

const float input[] = { -3.5987e-01, -4.6701e-01, 1.6226e+00, 5.3634e-02 };
const float* out_buf = dnnets_get_output_buffer(model);
const unsigned int out_len = dnnets_get_output_len(model);
dnnets_forward_pass(model, input, 4);

dnnets_free_model(model);

dnnets_deinit(); // Deinitialize library

Python bindings

dnnets provides Python bindings. You can install the package (including the native library) via the PyPI:

pip install dnnets

You can find a documentation of the Python API here: https://hannesbraun.net/share/doc/dnnets/python

The Python bindings can optionally be used together with NumPy. This feature can be enabled using dnnets.enable_numpy(). Then, all weights, biases and inputs need to be NumPy arrays. Outputs will then also be NumPy arrays.

Here's a quick example:

import dnnets

model = dnnets.load_json("sample_model.json")
input = [-3.5987e-01, -4.6701e-01, 1.6226e+00, 5.3634e-02]
output = model.forward_pass(input)

JSON format

Models can be stored in a JSON format. This is simply an object. The object has two attributes: input_size and layers. input_size is the size of the input that the model accepts. layers is an array of Layers. Each layer has a type attribute that describes the layer type and a size attribute describing the size of the layer (the number of outgoing values). A layer type can be one of the following:

• linear
• layer_norm
• elu
• leaky_relu
• relu
• relu6
• sigmoid
• tanh
• clip

You can also have a view at sample_model.json to see how this looks as a whole.

Linear

This is an example for a linear layer where the previous layer has 2 outputs and this layer has 3 outputs.

{
    "type": "linear",
    "size": 3,
    "weight": [
        [0.1, 0.2],
        [0.4, 0.4],
        [0.5, 0.6]
    ],
    "bias": [0.1, 0.2, 0.3]
}

Layer Normalization

This is an example for a layer normalization layer.

{
    "type": "layer_norm",
    "size": 3,
    "weight": [0.1, 0.2, 0.3],
    "bias": [0.1, 0.2, 0.3],
    "eps": 0.00001
}

eps is an optional attribute with a default value of 1e-5.

ELU

This is an example for an ELU layer.

{
    "type": "elu",
    "size": 3,
    "alpha": 0.9
}

alpha is an optional attribute with a default value of 1.0.

LeakyReLU

This is an example for an ELU layer.

{
    "type": "leaky_relu",
    "size": 3,
    "negative_slope": 0.1
}

negative_slope is an optional attribute with a default value of 0.01.

ReLU

This is the representation of a ReLU layer.

{
    "type": "relu",
    "size": 3
}

ReLU6

This is the representation of a ReLU6 layer.

{
    "type": "relu6",
    "size": 3
}

Sigmoid

This is the representation of a sigmoid layer.

{
    "type": "sigmoid",
    "size": 3
}

Tanh

This is the representation of a Tanh layer.

{
    "type": "tanh",
    "size": 3
}

Clip

This is an example for a clipping layer.

{
    "type": "clip",
    "size": 3,
    "min": -3.0,
    "max": 3.0
}

License

This library is licensed under the terms of the GNU Lesser General Public License (version 3 only) as published by the Free Software Foundation. For more information, see LICENSE.