ConvNetJS implements Deep Learning models and learning algorithms as well as nice browser-based demos, all in Javascript.
For much more information, see the main page at convnetjs.com
- Convolutional Neural Network on MNIST digits
- Convolutional Neural Network on CIFAR-10
- Neural Network with 2 hidden layers on toy 2D data
- 1D regression
- Denoising Autoencoder
- Image Painting
- Comparison of Learning Methods
- MagicNet demo
A reasonably comprehensive Getting Started tutorial is also available on main page.
Also have a look at even more comprehensive Docs.
But here we go:
Import convnet.js into your document: <script src="build/convnet.js"></script>
We first have to create a network. If you have images, here's an example network:
var layer_defs = [];
layer_defs.push({type:'input', out_sx:32, out_sy:32, out_depth:3});
layer_defs.push({type:'conv', sx:5, filters:8, stride:1, pad:2, activation:'relu'});
layer_defs.push({type:'pool', sx:2, stride:2});
layer_defs.push({type:'fc', num_neurons:20, activation:'relu'});
layer_defs.push({type:'softmax', num_classes:10});
It takes 32x32x3 images (3 is for RGB), convolves with 8 5x5 filters with stride 1 (and uses 0 padding of size 2 on each side so that the output is exactly 32x32 spatially), uses Rectified Linear Unit activation function (i.e. it thresholds all values below zero to zero), then pools spatially, then there is a fully connected layer and finally a classifier with 10 classes.
If you don't have images but some 2-D data, for example, your main building block is a FullyConnected layer:
var layer_defs = [];
layer_defs.push({type:'input', out_sx:1, out_sy:1, out_depth:2});
layer_defs.push({type:'fc', num_neurons:20, activation:'relu'});
layer_defs.push({type:'fc', num_neurons:40, activation:'relu', drop_prob:0.5});
layer_defs.push({type:'softmax', num_classes:4});
Here we have a 2-layer neural network classifier for 4 classes working on 2-D points, where the second layer is also followed by dropout for regularization. The drop_prob must be in range (0,1).
To create a network out of layer_defs, use:
var net = new convnetjs.Net();
net.makeLayers(layer_defs);
To use a simple network on arbitrary vectors:
var some_input = new convnetjs.Vol(1,1,2); // a 2-dimensional point
var class_probabilities = net.forward(some_input); // forward props all layers in turn
To use a network on images, we must convert them to a Vol(), which is accepted by the networks. One option is to use var img_vol = convnetjs.img_to_vol(elt) which takes an img element from the DOM (which you can get using plan JS like for example getElementByID()) and reads in its pixels to a (WxHx4) Vol. (4 = RGBA). The other options are manual - to get access to pixel values you must first draw the image on a canvas and then query the canvas. See img_to_vol() for examples on how to do this. There is also some code that helps with data augmentation and cropping: See augment, both of these defined in src/convnet_vol_util.js.
To train the network we use the Trainer class:
var trainer = new convnetjs.SGDTrainer(net, {learning_rate:0.01, momentum:0.9, batch_size:16, l2_decay:0.001});
for(var i=0;i<my_dataset.length;i++) {
var x = new convnetjs.Vol(1,1,2,0.0); // a 1x1x2 volume initialized to 0's.
x.w[0] = my_data[i][0]; // Vol.w is just a list, it holds your data
x.w[1] = my_data[i][1];
trainer.train(x, my_labels[i]);
}
Once you train the network, simply use net.forward(x) for predictions.
There is also an implementation of L2 loss that can be used for regression to arbitrary values. Instead of softmax, as a last layer just use
layer_defs.push({type:'regression', num_neurons:3});
In example above we'd be regressing to a single output that you must specify in the backward call as a list:
trainer.train(x, [0.5, 1.2, -0.7]);
To run these locally it is recommended that you use Nodejs or you may run into cross-origin security issues and not being able to load images. Chrome will have this problem. Firefox will work fine but I found Chrome to run much faster and more consistently.
To setup a nodejs server and start training:
- install nodejs:
sudo apt-get install nodejs cdinto convnetjs directory- install the connect library for nodejs to serve static pages
npm install connect node nodejs_server.js- Access the demos. http://localhost:8080/demo/classify2d.html will just work out of the box, but mnist.html and cifar10.html will require that you download the datasets and parse them into images. (You can also use the ones on my webserver if you're clever enough to see how to change the paths but naturally I'd prefer if you didn't use too much of my bandwidth). The python scripts I used to parse the datasets are linked to from the demo pages and require numpy and scipy.
If you'd like to use your own images, also don't miss the utility function convnetjs.img_to_vol(document.getElementById('input_image')) which takes an image element in the DOM as input and returns a convnetjs.Vol() ready to be consumed by ConvNetJS. Instead of loading images individually, you can also batch them up like I do for MNIST/CIFAR, with every image as a row in a large image. For example, one MNIST batch is a 10,000x768 image that I load once and then I pluck out a row at a time and reshape it into 28x28 image to use in a net.
If you're not working on images but have some custom data, you probably want just a basic neural network with no convolutions and pooling etc. You likely want to use the FullyConnLayer layer and stack it once or twice. Make sure to follow the FullyConnLayers with ReLU layers to introduce nonlinearities (append activation:'relu' in the layer definition), and also Dropout layers (append drop_prob:0.5 in layer definition, or any amount of dropout you desire).
If you're intending to use the code you need not worry about this section. Simply use build/convnet.js (or the minified version for deployment). If you want to hack on the code, you have to worry about this :) The problem is that the library is >1000 lines of JS and had to be broken down to be modular (all files inside src/) and it is compiled to build/convnet.js and also a minified version build/convnet-min.js. This is done using an ant task that concatenates the .js files in src/ and then minifies the result using YUI Compressor. The relevant files are in compile/. Make sure you have ant installed (on Ubuntu simple sudo apt-get install it), cd into compile/ directory and run:
ant -lib yuicompressor-2.4.8.jar -f build.xml
This will concatenate all files in /src, compile to build/convnet.js and minify that with YUI compressor to an identical but much smaller version ready for deployment.
Every layer takes a 3D volume (dimensions of WIDTH x HEIGHT x DEPTH) and transforms it into a different 3D volume using some set of internal parameters. Some layers (such as pooling, dropout) have no parameters. Currently available layers are:
###Helpful with Images:
Convolutional Layer: convolves input volume with local filters of given size, at given stride. An optional amount of zero padding can also be added.Locally Connected Layer: same as Convolutional layer (so local connectivity only to a small region below) but does not share weights across spatial locations (so no convolution).Pooling Layer: max-pools neighboring activations in 3D volume, keeping depth the same but reducing width and height of volumeLocal Contrast Normalization Layer: Creates local competition among neurons along depth at specific location, for all locations.
###Helpful with arbitrary data:
Fully Connected Layer: a number of neurons connected densely to input volume. They each compute dot product with the input.Dropout Layer: implements dropout to control overfitting. Can be used after layers that have very large number of nodes for regularization. You don't have to add this explicitly, simply use drop_prob:0.5 (or other amount) in a layer def to automatically add a Dropout layer right after it.
###Non-linearities
ReluLayer: creates the ReLU (Rectified Linear Unit) activation function. You don't have to add this explicitly, simply use activation:'relu' in a layer def to follow that layer with ReLU.SigmoidLayer: can be used as nonlinearity instead of ReluLayer, computes the sigmoid function x->1/(1+e^(-x)) You don't have to add this explicitly, simply use activation:'sigmoid' in a layer def to follow that layer with the Sigmoid.MaxoutLayer: Computes max(x) where x is a group of elements in the input. By default, the group size is 2 but this can be changed by passing in a different group_size. Note that the size of the input should divide exactly into group_size. If you are using a convnet, the depth should divide nicely into group_size. If you're using a regular neural net, make sure the number of neurons in the layer before maxout has a multiple of group_size elements. MaxOut enjoys some nice properties computationally because it is linear (like ReLU) so it trains fast, but it isn't as unstable: for example, you don't run the danger of having ReLU units "die" if your learning rate is slightly too high. That's comforting.
###Objective layers
Softmax: this is a classifier layer that should currently be last. It computes probabilities via dense connections to the input volume and dot product with class weights.RegressionLayer: can be replaced with Softmax to do regression instead of classification.
If you're not dealing with images, the only layer that is of interest is the Fully Connected Layer, which you probably want to stack 1/2/3 times on top of your input. (Make sure you have non-linearities in between too). You also may consider using a Dropout layer in places where there are a lot of parameters to control overfitting (overfitting = your validation accuracy is much lower than the training accuracy).
If you're dealing with images, your networks should look similar to what you see in the demos.
The MagicNet class is great if you're a beginner with Neural Networks and just want to train something completely automatically that works well on your data. The MagicNet class takes a training set (as a list of convnetjs.Vols and training labels) and handles all aspects of the training completely automatically. Internally, it will generate multiple neural networks in turns, try out every one of them on your dataset using n-fold cross-validation, and finally it will create a powerful ensemble of the best neural networks it found to create the final predictor. The MagicNet class strives to be the ultimate black-box predictor: you don't have to worry about any learning rates, weight decays, non-linearities or anything.
Right now, only classification is supported with MagicNet. Regression is planned for later. See the MagicNet demo for more details and example code. See the ConvNetJS documentation for full details of the class.
The library also has a Reinforcement Learning demo that follows a Deep Q - Learning NIPS2013 Workshop Paper Playing Atari with Deep Reinforcement Learning. In short, a Neural Network is used to model the value function. The API is very simple to use. For example, lets train an agent that observes 3-dimensional states and is asked to do one of two actions. Lets reward the agent only for action 0 for sake of very simple example:
var brain = new deepqlearn.Brain(3, 2); // 3 inputs, 2 possible outputs (0,1)
var state = [Math.random(), Math.random(), Math.random()];
for(var k=0;k<10000;k++) {
var action = brain.forward(state); // returns index of chosen action
var reward = action === 0 ? 1.0 : 0.0;
brain.backward(reward); // <-- learning magic happens here
state[Math.floor(Math.random()*3)] += Math.random()*2-0.5;
}
brain.epsilon_test_time = 0.0; // don't make any more random choices
brain.learning = false;
// get an optimal action from the learned policy
var action = brain.forward(array_with_num_inputs_numbers);
Of course, there are many possible options you can set. have a look at the reinfocement learning demo rldemo.html inside demo/ folder.
- install:
npm install convnetjs var convnetjs = require("convnetjs");
MIT