variable_length_data.py

import theano
import theano.tensor as T
import numpy as np


def step(b, a, x, t, W):
    y = theano.dot(x[a:b], W)
    y = theano.printing.Print()(y)
    error = 0.5 * T.sqr(t[a:b] - y).sum()
    return b, error
    

if __name__ == '__main__':
    floatX = theano.config.floatX
    
    # input matrix (time steps * number of examples, features)
    x = T.matrix(name='x')
    # target matrix (time steps * number of examples, features)
    t = T.matrix(name='t')
    # number of time steps for each example
    i = T.vector(name='i', dtype='int64')
    
    W = theano.shared(np.random.uniform(-0.01, 0.01, size=(2, 2)).astype(floatX), name='W')
    
    i0 = T.constant(0, dtype='int64')
    (_, error), _ = theano.scan(step,
                                sequences=[T.cumsum(i)],
                                non_sequences=[x, t, W],
                                outputs_info=[i0, None])
    
    dW = theano.grad(error.mean(), W)
    
    f = theano.function([x, t, i], [error, dW])
    
    print f(np.random.uniform(size=(7, 2)).astype(floatX),
            np.random.uniform(size=(7, 2)).astype(floatX),
            np.asarray([2,3,2], dtype='int64'))