return perturbRnn.loss( perturbRnn.feedforward(trainData.x)[0], trainData.y )
def Et(W):
pass #TODO
with Timer('bptt'):
dx, dh, dy = rnn.backprop_thru_time_chain(trainData.x, trainData.y)
with Timer('bptt double loop'):
dx2_t, dh2_t, dy2_t = rnn.backprop_thru_time_double_loop(trainData.x, trainData.y)
dx2 = sum(dx2_t)/T
dh2 = sum(dh2_t)/T
dy2 = sum(dy2_t)/T
with Timer('lagrange'):
dxl, dhl, dyl = rnn.backprop_thru_time_lagrange(trainData.x, trainData.y)
with Timer ('lagrange loop'):
dxll, dhll, dyll = rnn.backprop_thru_time_lagrange_loop(trainData.x, trainData.y)
with Timer('numerical'):
dn = numerical_gradient(E, weights_to_vector(*rnn.get_weights()), 1e-6)
dxn, dhn, dyn = vector_to_weights( dn, rnn.Nx, rnn.Nh, rnn.Ny )
#with Timer('double loop numerical'): #TODO: numerically test to check that each dE(t)/dW is correct
# dn = numerical_gradient(E, weights_to_vector(*rnn.get_weights()), 1e-6)
# dxn, dhn, dyn = vector_to_weights( dn, rnn.Nx, rnn.Nh, rnn.Ny )
# dxn2_t, dhn2_t, dyn2_t = numerical_gradient(Et, weights_to_vector(*rnn.get_weights()), 1e-6)
# dxn2 = sum(dxn2_t)/T
# dhn2 = sum(dhn2_t)/T
# dyn2 = sum(dyn2_t)/T
