layer - neural network inference from the command line

layer is a program for doing neural network inference the Unix way. Many modern neural network operations can be represented as sequential, unidirectional streams of data processed by pipelines of filters. The computations at each layer in these neural networks are equivalent to an invocation of the layer program, and multiple invocations can be chained together to represent the entirety of such networks.

For example, performing inference on a neural network with two fully-connected layers might look something like this:

cat input | layer full -w w.1 --input-shape=2 -f tanh | layer full -w w.2 --input-shape=3 -f sigmoid

layer applies the Unix philosophy to neural network inference. Each type of a neural network layer is a distinct subcommand. Simple text streams of delimited numeric values serve as the interface between different layers of a neural network. Each invocation of layer does one thing: it feeds the numeric input values forward through an instantiation of a neural network layer, then emits the resulting output numeric values.

Usage

Example: a convolutional neural network for CIFAR-10.

$ cat cifar10_x.csv \
| layer convolutional -w w0.csv -b b0.csv --input-shape=32,32,3 --filter-shape=3,3 --num-filters=32 -f relu \
| layer convolutional -w w1.csv -b b1.csv --input-shape=30,30,32 --filter-shape=3,3 --num-filters=32 -f relu \
| layer pooling --input-shape=28,28,32 --filter-shape=2,2 --stride=2 -f max

Example: a multi-layer perceptron for XOR.

$ # Fully connected layer with three neuronsecho"-2.35546875,-2.38671875,3.63671875,3.521484375,-2.255859375,-2.732421875"> layer1.weights
echo"0.7958984375,0.291259765625,1.099609375"> layer1.biases
$ # Fully connected layer with one neuronecho"-5.0625,-3.515625,-5.0625"> layer2.weights
echo"1.74609375"> layer2.biases
$ # Compute XOR for all possible binary inputsecho -e "0,0\n0,1\n1,0\n1,1" \
| layer full -w layer1.weights -b layer1.biases --input-shape=2 -f tanh \
| layer full -w layer2.weights -b layer2.biases --input-shape=3 -f sigmoid
0.00129012749948779
0.99147053740106
0.991243357927591
0.0111237568184365

Installation

Requirements: BLAS 3.6.0+

1. Download a release
2. Install BLAS 3.6.0+

• On Debian-based systems: apt-get install -y libblas3
• On RPM-based system: yum install -y blas
• On macOS 10.3+, BLAS is pre-installed as part of the Accelerate framework

3. Unzip the release and run [sudo] ./install.sh, or manually relocate the binaries to the path of your choice.

About

layer is currently implemented as a proof-of-concept and supports a limited number of neural network layer types. The types of layers are currently limited to feed-forward layers that can be modeled as sequential, unidirectional pipelines.

Input values, weights and biases for parameterized layers, and output values are all read and written in row-major order, based on the shape parameters specified for each layer.

layer is implemented in CHICKEN Scheme.

License

Copyright © 2018-2019