mlpfile 0.0.1

Multilayer perceptron file format and evaluation.

A simple file format and associated tools to save/load multilayer perceptrons (aka fully-connected neural networks).

Features:

• Create the files in Python from a torch.nn.Sequential.
• Load the files in C++, or in Python via bindings.
• Evaluate the network and/or its Jacobian on an input.
• C++ interface uses Eigen types.
• Binary file I/O (no C++ dependency on Protobuf, etc.)

API docs: https://jpreiss.github.io/mlpfile/api.html

Installation

TODO: The pip package is almost ready.

Example code

Python:

model_torch = <train a torch.nn.Sequential somehow>
mlpfile.torch.write(model_torch, "net.mlp")

model_ours = mlpfile.Model.load("net.mlp")
x = <appropriate input>
y = model.forward(x)

C++:

mlpfile::Model model = mlpfile::Model::load("net.mlp");
Eigen::VectorXf x = <appropriate input>;
Eigen::VectorXf y = model.forward(x);

Performance

mlpfile is faster than popular alternatives for small networks on the CPU. This is a very small example, but such small networks can appear in time-sensitive realtime applications.

Test hardware is a 2021 MacBook Pro with Apple M1 Pro CPU.

mlpfile is over 3x faster than ONNX on both forward pass and Jacobian in this test. You can test on your own hardware by running benchmark.py.

$ python benchmark.py

mlpfile::Model with 6 Layers
Input: 40
Linear: 40 -> 100
ReLU
Linear: 100 -> 100
ReLU
Linear: 100 -> 10

***********
* Forward *
***********
torch:   23.79 usec
 onnx:    6.67 usec
 ours:    1.95 usec

************
* Jacobian *
************
torch-autodiff:  246.24 usec
  torch-manual:   52.66 usec
   onnx-python:   42.85 usec
          ours:   11.61 usec

Motivation

The performace shown above is a major motivation, but besides that:

The typical choices for NN deployment from PyTorch to C++ (of which I am aware) are TorchScript and the ONNX format. Both are heavyweight and complicated because they are designed to handle general computation graphs like ResNets, Transformers, etc. Their Python packages are easy to use via pip, but their C++ packages aren't a part of standard package managers, and compiling from source (at least for ONNX-runtime) is very slow.

Intel and NVidia's ONNX loaders might be better, but they are not cross-platform.

ONNX-runtime also doesn't make it easy to extract the model weights from the file. This means we can't (easily) use their file format and loader but compute the neural network function ourselves for maximum speed.

Also, we want to evaluate the NN's Jacobian in our research application. To do this with ONNX, we must represent the MLP Jacobian's computational graph in the file instead of the MLP itself. It turns out that PyTorch's torch.func.jacrev generates a computational graph that can't be serialized with PyTorch's own ONNX exporter. Therefore, we must write the symbolically-derived Jacobian by hand in PyTorch. So that unwanted complexity must exist somewhere, whether it is C++ or Python.

It's possible that TorchScript is better than ONNX in some of these issues, but I was tired of searching for libraries and wanted to ensure top speed anyway.

File format

It is a binary file format. All numerical types are little-endian, but the code currently assumes it's running on a little-endian machine.

The file format is not stable!

layer types enum:
    1 - input
    2 - linear
    3 - relu

header:
    number of layers (uint32)

    for each layer:
        enum layer type (uint32)
        if input:
            input dim (uint32)
        if linear:
            output dim (uint32)
        if relu:
            nothing

data:
    for each layer:
        if linear:
            weight (float32[], row-major)
            bias (float32[])
        otherwise:
            nothing