RecurrentJS is a Javascript library that implements:
- Deep Recurrent Neural Networks (RNN)
- Long Short-Term Memory networks (LSTM)
- In fact, the library is more general because it has functionality to construct arbitrary expression graphs over which the library can perform automatic differentiation similar to what you may find in Theano for Python, or in Torch etc. Currently, the code uses this very general functionality to implement RNN/LSTM, but one can build arbitrary Neural Networks and do automatic backprop.
An online demo that memorizes character sequences can be found below. Sentences are input data and the networks are trained to predict the next character in a sentence. Thus, they learn English from scratch character by character and eventually after some training generate entirely new sentences that sometimes make some sense :)
Character Sequence Memorization Demo
The core of the library is a Graph structure which maintains the links between matrices and how they are related through transformations. Another important building block is the Mat class which represents a 2-dimensional N x D matrix, its values in field .w and its derivates in field .dw. Here is how you would implement a simple Neural Network layer:
var W = new R.RandMat(10, 4); // weights Mat
var x = new R.RandMat(4, 1); // random input Mat
var b = new R.RandMat(10, 1); // bias vector
// matrix multiply followed by bias offset. h is a Mat
var G = new R.Graph();
var h = G.add(G.mul(W, x), b);
// the Graph structure keeps track of the connectivities between Mats
// we can now set the loss on h
h.dw[0] = 1.0; // say we want the first value to be lower
// propagate all gradients backwards through the graph
// starting with h, all the way down to W,x,b
// i.e. this sets .dw field for W,x,b with the gradients
G.backward();
// do a parameter update on W,b:
var s = new R.Solver(); // the Solver uses RMSProp
// update W and b, use learning rate of 0.01,
// regularization strength of 0.0001 and clip gradient magnitudes at 5.0
var model = {'W':W, 'b':b};
s.step(model, 0.01, 0.0001, 5.0)To construct and train an LSTM for example, you would proceed as follows:
// takes as input Mat of 10x1, contains 2 hidden layers of
// 20 neurons each, and outputs a Mat of size 2x1
var hidden_sizes = [20, 20];
var lstm_model = R.initLSTM(10, hidden_sizes, 2);
var x1 = new R.RandMat(10, 1); // example input #1
var x2 = new R.RandMat(10, 1); // example input #2
var x3 = new R.RandMat(10, 1); // example input #3
// pass 3 examples through the LSTM
var G = new R.Graph();
var out1 = R.forwardLSTM(G, lstm_model, hidden_sizes, x1, {});
var out2 = R.forwardLSTM(G, lstm_model, hidden_sizes, x2, out1);
var out3 = R.forwardLSTM(G, lstm_model, hidden_sizes, x3, out2);
// the field .o contains the output Mats:
// e.g. x1.o is a 2x1 Mat
// for example lets assume we have binary classification problem
// so the output of the LSTM are the log probabilities of the
// two classes. Lets first get the probabilities:
var prob1 = R.softmax(out1.o);
var target1 = 0; // suppose first input has class 0
cost += -Math.log(probs.w[ix_target]); // softmax cost function
// cross-entropy loss for softmax is simply the probabilities:
out1.dw = prob1.w;
// but the correct class gets an extra -1:
out1.dw[ix_target] -= 1;
// in real application you'd probably have a desired class
// for every input, so you'd iteratively se the .dw loss on each
// one. In the example provided demo we are, for example,
// predicting the index of the next letter in an input sentence.
// update the LSTM parameters
G.backward();
var s = new R.Solver();
// perform RMSprop update with
// step size of 0.01
// L2 regularization of 0.00001
// and clipping the gradients at 5.0 elementwise
s.step(lstm_model, 0.01, 0.00001, 5.0);You'll notice that the Softmax and so on isn't folded very neatly into the library yet and you have to understand backpropagation. I'll fix this soon.
This code works fine, but it's a bit rough around the edges - you have to understand Neural Nets well if you want to use it and it isn't beautifully modularized. I thought I would still make the code available now and work on polishing it further later, since I hope that even in this state it can be useful to others who may want to browse around and get their feet wet with training these models or learning about them.
MIT