nnstats

Package for calculating statistics for Neural Networks.

Installation

  npm install --save nnstats

Usage

Basic usage

The following show basic usage for analyzing CNN in Tensorflow (padding takes either SAME or VALID).

'use strict';


let nnStats = require('./');
let analyzer = nnStats.analyzer;
let utils = nnStats.utils;

let model = [{
    'type': 'conv',
    'filter': [5, 5, 32],
    'strides': [1, 1],
    'padding': 'SAME',
    'datasize': 4
},{
    'type': 'pool',
    'filter': [2, 2],
    'strides': [2, 2],
    'padding': 'SAME',
    'datasize': 4
},{
    'type': 'conv',
    'filter': [5, 5, 64],
    'strides': [2, 2],
    'padding': 'SAME',
    'datasize': 4
},{
    'type': 'pool',
    'filter': [2, 2],
    'strides': [2, 2],
    'padding': 'SAME',
    'datasize': 4
},{
    'type': 'fc',
    'hidden': 2048
},{
    'type': 'fc',
    'hidden': 1000
}]


let input = [28, 28, 1]
let options = {
    tensorflow: true
}

let stats = analyzer.analyzeNetwork(model, input, options);

utils.report(stats);

Model layout

Currently, here are 3 supported layers type: convolution layer (conv), pooling layer (pool), and fully-connected layer ('fc'). In all cases, type field is required in layer object to identity layer type.

Fields in conv layer:

filter: 3D array kernal size (height, width, outChannel).
strides: 2D array strides size (height, width).
padding: 2D array padding size (height, width).
datasize (optional): number of byte of one value (for float32, this should be 4 because 32 / 8 = 4). This value is used to calculate how much memory needed for the network.

Fields in pool layer:

filter: 2D array kernal size (height, width).
strides: 2D array strides size (height, width).
padding: 2D array padding size (height, width).
datasize (optional): number of byte of one value (for float32, this should be 4 because 32 / 8 = 4). This value is used to calculate how much memory needed for the network.

Fields in fc layer:

hidden: number of neurons in hiddent layers.

Created with