Feed-forward neural network implementation in C with SIMD instructions.
BTW: simd_neuralnet is a bit of a misnomer. It is not SIMD instructions apart from AVX, AVX2 and AVX512 instructions in some of the most heavy functions. Can anyone come up with a better name? It is a neural network library written in C. I'm not sure what else makes this library different from the other libraries out there.
It is very fast compared to other libraries like Keras and PyTorch - just give it a try and you will see.
This project started out as a study project for myself to improve my abilities to implement a feed forward neural network in C. A lot of this code will be based on code from some of my other projects. Hopefully this will be generally usable.
One of the ideas here is to make sure that this implementation is correct and gives the same results as other neural network implementations. We will therefore also implement a Python neural network in the verification. The weights and other parameters of other neural network libraries should therefore be interchangeable with the weights of these.
Say you train a neural network from Keras, and then you want to use it in your C coded application? Then you can simply store the weights from your Keras neural network and load them into this codes, and you whole neural network is totally independent of both Python and Keras (and a GPU).
Since this is pure ANSI C this library will be perfect for embedded devices. I'm also using this library in a online learning system where only one sample is fed into the trainer at a time in a reinforcement learning pattern. This library then works perfectly since there is no need to struggle with python bindings or slow memory transfers to GPU memory.
To be able to achieve the above, we need to set some limitations.
float32 precision only! The code will use SIMD instructions, so double precision will
slow things down, and float16 is not precise enough and has limited support.The following loss functions are implemented.
The following activation functions are implemented.
Most of these optimizers can also handle momentum and Nesterov momentum.
So, the idea is to keep this small and beautiful. Features, like:
... can be added successivly as needed. Please make pull requests.
The neural networks are stored as numpy arrays in the .npz in PKZIP manner. It is therefore simple to
interact between python based neural network frameworks and simd_neuralnet. Maybe you want to train your
neural networks with Keras and the deploy you trained neural network with simd_neuralnet? Not a problem at all.
The arrays that makes the weights and biases in the layers of the neural net, are simply stored as
.npy "files" in an PKZIP archive file which is the very same way numpy stores it's arrays in .npz files.
They are stores in input to output layer order, with weights first and bias following. To keep track of the
activation functions used, a simple array of strings (actually character arrays) are stored at the end of
the archive. Actually, the activations function string array can be stores anywhere in the chain of other
arrays, however the neuralnet_save() function will store it at the en of the zip archive.
A simple feed forward neural network with 3 inputs, 4 hidden units and 2 outputs, and tanh and sigmoid activations will then be stored as.
"tanh" and "sigmoid"Let me build that neural network in C code:
#include "neuralnet.h"
int main(){
neuralnet_t *nn = neuralnet_create( 2, INT_ARRAY( 3, 4, 2 ), STR_ARRAY( "tanh", "sigmoid"));
neuralnet_initialize( nn, "xavier", "xavier" );
/* I'm only showing this for being able to save, so I won't set a loss function */
neuralnet_save( nn, "my_first_neuralnet_%d_%d_%d.npz", 3,4,2 );
neuralnet_free( nn );
return 0;
}I am then able to open this file in python and numpy:
Python 3.8.1 (default, Jan 22 2020, 06:38:00)
[GCC 9.2.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import numpy as np
>>> weights = np.load("my_first_neuralnet_3_4_2.npz")
>>> params = [ weights[i] for i in weights.files]
>>> import pprint
>>> for array in params:
... pprint.pprint( array )
...
array([[ 0.6299053 , -0.19556482, 0.52419794, 0.5526036 ],
[ 0.7622228 , -0.56002605, -0.3051082 , 0.49666473],
[-0.41148126, 0.09993298, -0.04185252, 0.23862255]],
dtype=float32)
array([0., 0., 0., 0.], dtype=float32)
array([[-0.27043104, 0.02680182],
[ 0.9044595 , 0.8323902 ],
[ 0.27142346, 0.43459392],
[-0.71679485, 0.21393776]], dtype=float32)
array([0., 0.], dtype=float32)
array([b'tanh', b'sigmoid'], dtype='|S7')And of course it also works the other way, I can create an array in numpy and then stroe it back to be read in C. I only need to make sure I store the activation names in pure ascii bytes. (No unicode)
w1 = np.random.random( (3,4) ).astype(np.float32)
b1 = np.zeros( (4,), dtype=np.float32 )
w2 = np.random.random( (4,2) ).astype(np.float32)
b2 = np.zeros( (2,), dtype=np.float32 )
activations = np.array(["tanh", "sigmoid"]).astype('S')
np.savez( "numpy_saved_neuralnet.npz", w1, b1, w2, b2, activations )And that file can be used with neuralnet_load().