mlpfile

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.
• Perform a step of gradient descent (in place, for one datapoint, no momentum).
• C++ interface uses Eigen types.
• Generate fast allocation-free C or C++/Eigen code (faster) for the forward pass.
• Binary file I/O (no C++ dependency on Protobuf, etc.)

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

Installation

To use the Python export and/or bindings, install the pip package:

pip install mlpfile

If you only need to load and evaluate networks in C++, the easiest way is to either 1) copy the files from mlpfile/cpp into your project, or 2) include this repo as a submodule.

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. TorchScript is surprisingly fast for the manually-computed Jacobian, but is still slow for the forward pass. You can test on your own hardware by running benchmark.py.

$ python benchmark.py

┌─────────────────┐
│ Model structure │
└─────────────────┘
mlpfile::Model with 5 Layers, 40 -> 10
Linear: 40 -> 100
ReLU
Linear: 100 -> 100
ReLU
Linear: 100 -> 10

┌─────────┐
│ Forward │
└─────────┘
        torch:   15.72 usec
  torchscript:    6.97 usec
         onnx:    5.98 usec
         ours:    1.91 usec
    codegen_c:   10.10 usec
codegen_eigen:    1.11 usec

┌──────────┐
│ Jacobian │
└──────────┘
    torch-autodiff:   88.19 usec
      torch-manual:   40.82 usec
torchscript-manual:   16.81 usec
              onnx:   42.00 usec
              ours:   11.97 usec

┌────────────┐
│ OGD-update │
└────────────┘
torch:  129.38 usec
 ours:   10.17 usec

Motivation

The performance 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. Compiling from source is very slow for ONNX-runtime; I have not tried TorchScript yet.

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. It turns out that PyTorch's torch.func.jacrev generates a computational graph that can't be serialized with TorchScript or 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.

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:
    2 - linear
    3 - relu

header:
    number of layers (uint32)
    input dimension (uint32)

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

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