gistfile1.py

from functools import partial
from theano.ifelse import ifelse
import numpy as np
import theano
import theano.tensor as T

def maxx(e):
    return ifelse(T.gt(e, 0.0), e, 0.0)

def s(x, y, U):
    return T.dot(T.dot((x).reshape((1,-1)), U.T), T.dot(U, (y)))

def getTrain(D=100, W=100):
    U_O = theano.shared(np.random.randn(D, W))
    U_R = theano.shared(np.random.randn(D, W))
    m_o1 = T.vector('m_o1')
    m_o2 = T.vector('m_o2')
    x = T.vector('x')
    y = T.vector('y')
    r = T.vector('r')
    gamma = T.scalar('gamma')
    f1 = T.vector('f1')
    f2 = T.vector('f2')
    L = T.matrix('L') # list of messages
    V = T.matrix('V') # vocab
    sO = partial(s, U=U_O)
    sR = partial(s, U=U_R)
    cost1, u1 = theano.scan(lambda f_bar: maxx(gamma - sO(x, f1) + sO(x, f_bar)), L)
    cost2, u2 = theano.scan(lambda f_bar: gamma - sO(x, f2) - sO(m_o1, f2) + sO(x, f_bar) + sO(m_o1, f_bar), L)
    cost3, u3 = theano.scan(lambda r_bar: gamma - sR(x, r) - sR(m_o1, r) - sR(m_o2, r) + sR(x, r_bar) + sR(m_o1, r_bar) + sR(m_o2, r_bar), V)
    cost = cost1.sum() + cost2.sum() + cost3.sum()
    g_uo, g_ur = T.grad(cost, [U_O, U_R])

    train = theano.function(
        inputs=[m_o1, m_o2, x, gamma, f1, f2, L, V, r],
        outputs=[cost],
        updates=[(U_O, U_O-0.1*g_uo), (U_R, U_R-0.1*g_ur)])
    return train

getTrain()