def test_pooling_rnn(): l = SupervisedRecurrentNetwork(2, [3], 1, hidden_transfers=['sigmoid'], pooling='mean', loss='ncac') l = SupervisedRecurrentNetwork(2, 3, 1, 'sigmoid', 'identity', 'ncac', 'mean') f = l.function(['inpt', 'target'], 'loss', mode='FAST_COMPILE') d_loss_wrt_pars = T.grad(l.exprs['loss'], l.parameters.flat) fprime = l.function(['inpt', 'target'], d_loss_wrt_pars, mode='FAST_COMPILE') X = np.random.random((10, 30, 2)).astype(theano.config.floatX) Z = np.random.random((30, 1)).astype(theano.config.floatX) f(X, Z) fprime(X, Z)
def test_srnn(): l = SupervisedRecurrentNetwork(2, 3, 1, 'sigmoid', 'identity', 'squared') f = l.function(['inpt', 'target'], 'loss', mode='FAST_COMPILE') d_loss_wrt_pars = T.grad(l.exprs['loss'], l.parameters.flat) fprime = l.function(['inpt', 'target'], d_loss_wrt_pars, mode='FAST_COMPILE') X = np.random.random((10, 3, 2)).astype(theano.config.floatX) Z = np.random.random((10, 3, 1)).astype(theano.config.floatX) f(X, Z) fprime(X, Z)