Python GatedRecurrent.apply方法代码示例

# 需要导入模块: from blocks.bricks.recurrent import GatedRecurrent [as 别名]
# 或者: from blocks.bricks.recurrent.GatedRecurrent import apply [as 别名]
def example2():
    """GRU"""
    x = tensor.tensor3('x')
    dim = 3

    fork = Fork(input_dim=dim, output_dims=[dim, dim*2],name='fork',output_names=["linear","gates"], weights_init=initialization.Identity(),biases_init=Constant(0))
    gru = GatedRecurrent(dim=dim, weights_init=initialization.Identity(),biases_init=Constant(0))

    fork.initialize()
    gru.initialize()

    linear, gate_inputs = fork.apply(x)
    h = gru.apply(linear, gate_inputs)

    f = theano.function([x], h)
    print(f(np.ones((dim, 1, dim), dtype=theano.config.floatX))) 

    doubler = Linear(
                 input_dim=dim, output_dim=dim, weights_init=initialization.Identity(2),
                 biases_init=initialization.Constant(0))
    doubler.initialize()

    lin, gate = fork.apply(doubler.apply(x))
    h_doubler = gru.apply(lin,gate)

    f = theano.function([x], h_doubler)
    print(f(np.ones((dim, 1, dim), dtype=theano.config.floatX))) 

# 需要导入模块: from blocks.bricks.recurrent import GatedRecurrent [as 别名]
# 或者: from blocks.bricks.recurrent.GatedRecurrent import apply [as 别名]
def gru_layer(dim, h, n):
    fork = Fork(output_names=['linear' + str(n), 'gates' + str(n)],
                name='fork' + str(n), input_dim=dim, output_dims=[dim, dim * 2])
    gru = GatedRecurrent(dim=dim, name='gru' + str(n))
    initialize([fork, gru])
    linear, gates = fork.apply(h)
    return gru.apply(linear, gates)

# 需要导入模块: from blocks.bricks.recurrent import GatedRecurrent [as 别名]
# 或者: from blocks.bricks.recurrent.GatedRecurrent import apply [as 别名]
def gru_layer(dim, h, n):
    fork = Fork(
        output_names=["linear" + str(n), "gates" + str(n)],
        name="fork" + str(n),
        input_dim=dim,
        output_dims=[dim, dim * 2],
    )
    gru = GatedRecurrent(dim=dim, name="gru" + str(n))
    initialize([fork, gru])
    linear, gates = fork.apply(h)
    return gru.apply(linear, gates)

# 需要导入模块: from blocks.bricks.recurrent import GatedRecurrent [as 别名]
# 或者: from blocks.bricks.recurrent.GatedRecurrent import apply [as 别名]
class InnerRecurrent(BaseRecurrent, Initializable):
    def __init__(self, inner_input_dim, outer_input_dim, inner_dim, **kwargs):
        self.inner_gru = GatedRecurrent(dim=inner_dim, name='inner_gru')

        self.inner_input_fork = Fork(
            output_names=[name for name in self.inner_gru.apply.sequences
                          if 'mask' not in name],
            input_dim=inner_input_dim, name='inner_input_fork')
        self.outer_input_fork = Fork(
            output_names=[name for name in self.inner_gru.apply.sequences
                          if 'mask' not in name],
            input_dim=outer_input_dim, name='inner_outer_fork')

        super(InnerRecurrent, self).__init__(**kwargs)

        self.children = [
            self.inner_gru, self.inner_input_fork, self.outer_input_fork]

    def _push_allocation_config(self):
        self.inner_input_fork.output_dims = self.inner_gru.get_dims(
            self.inner_input_fork.output_names)
        self.outer_input_fork.output_dims = self.inner_gru.get_dims(
            self.outer_input_fork.output_names)

    @recurrent(sequences=['inner_inputs'], states=['states'],
               contexts=['outer_inputs'], outputs=['states'])
    def apply(self, inner_inputs, states, outer_inputs):
        forked_inputs = self.inner_input_fork.apply(inner_inputs, as_dict=True)
        forked_states = self.outer_input_fork.apply(outer_inputs, as_dict=True)

        gru_inputs = {key: forked_inputs[key] + forked_states[key]
                      for key in forked_inputs.keys()}

        new_states = self.inner_gru.apply(
            iterate=False,
            **dict_union(gru_inputs, {'states': states}))
        return new_states  # mean according to the time axis

    def get_dim(self, name):
        if name == 'states':
            return self.inner_gru.get_dim(name)
        else:
            return AttributeError

# 需要导入模块: from blocks.bricks.recurrent import GatedRecurrent [as 别名]
# 或者: from blocks.bricks.recurrent.GatedRecurrent import apply [as 别名]
class Encoder(Initializable):
    def __init__(self, vocab_size, embedding_dim, state_dim, reverse=True,
                 **kwargs):
        super(Encoder, self).__init__(**kwargs)
        self.vocab_size = vocab_size
        self.embedding_dim = embedding_dim
        self.state_dim = state_dim
        self.reverse = reverse

        self.lookup = LookupTable(name='embeddings')
        self.transition = GatedRecurrent(Tanh(), name='encoder_transition')
        self.fork = Fork([name for name in self.transition.apply.sequences
                          if name != 'mask'], prototype=Linear())

        self.children = [self.lookup, self.transition, self.fork]

    def _push_allocation_config(self):
        self.lookup.length = self.vocab_size
        self.lookup.dim = self.embedding_dim
        self.transition.dim = self.state_dim
        self.fork.input_dim = self.embedding_dim
        self.fork.output_dims = [self.state_dim
                                 for _ in self.fork.output_names]

    @application(inputs=['source_sentence', 'source_sentence_mask'],
                 outputs=['representation'])
    def apply(self, source_sentence, source_sentence_mask):
        # Time as first dimension
        source_sentence = source_sentence.dimshuffle(1, 0)
        source_sentence_mask = source_sentence_mask.T
        if self.reverse:
            source_sentence = source_sentence[::-1]
            source_sentence_mask = source_sentence_mask[::-1]

        embeddings = self.lookup.apply(source_sentence)
        representation = self.transition.apply(**merge(
            self.fork.apply(embeddings, as_dict=True),
            {'mask': source_sentence_mask}
        ))
        return representation[-1]

# 需要导入模块: from blocks.bricks.recurrent import GatedRecurrent [as 别名]
# 或者: from blocks.bricks.recurrent.GatedRecurrent import apply [as 别名]
class Encoder(Initializable):
    """Encoder of RNNsearch model."""

    def __init__(self, blockid, vocab_size, embedding_dim, state_dim, **kwargs):
        super(Encoder, self).__init__(**kwargs)
        self.vocab_size = vocab_size
        self.embedding_dim = embedding_dim
        self.state_dim = state_dim
        self.blockid = blockid

        self.lookup = LookupTable(name='embeddings' + '_' + self.blockid)
        self.gru = GatedRecurrent(activation=Tanh(), dim=state_dim, name = "GatedRNN" + self.blockid)
        self.fwd_fork = Fork(
            [name for name in self.gru.apply.sequences
             if name != 'mask'], prototype=Linear(), name='fwd_fork' + '_' + self.blockid)

        self.children = [self.lookup, self.gru, self.fwd_fork]

    def _push_allocation_config(self):
        self.lookup.length = self.vocab_size
        self.lookup.dim = self.embedding_dim

        self.fwd_fork.input_dim = self.embedding_dim
        self.fwd_fork.output_dims = [self.gru.get_dim(name)
                                     for name in self.fwd_fork.output_names]

    @application(inputs=['source_sentence', 'source_sentence_mask'],
                 outputs=['representation'])
    def apply(self, source_sentence, source_sentence_mask):
        # Time as first dimension
        source_sentence = source_sentence.T
        source_sentence_mask = source_sentence_mask.T

        embeddings = self.lookup.apply(source_sentence)
        grupara =  merge( self.fwd_fork.apply(embeddings, as_dict=True) , {'mask': source_sentence_mask})
        representation = self.gru.apply(**grupara)
        return representation

# 需要导入模块: from blocks.bricks.recurrent import GatedRecurrent [as 别名]
# 或者: from blocks.bricks.recurrent.GatedRecurrent import apply [as 别名]
class Encoder(Initializable):
    def __init__(self, vocab_size, embedding_dim, state_dim, **kwargs):
        super(Encoder, self).__init__(**kwargs)
        self.vocab_size = vocab_size
        self.embedding_dim = embedding_dim
        self.state_dim = state_dim

        self.lookup = LookupTable(name='embeddings')
        self.GRU = GatedRecurrent(activation=Tanh(), dim=state_dim)
        self.children = [self.lookup, self.GRU]

    def _push_allocation_config(self):
        self.lookup.length = self.vocab_size
        self.lookup.dim = self.embedding_dim

    @application(inputs=['source_sentence', 'source_sentence_mask'],
                 outputs=['representation'])
    def apply(self, source_sentence, source_sentence_mask):
        source_sentence = source_sentence.T
        source_sentence_mask = source_sentence_mask.T

        embeddings = self.lookup.apply(source_sentence)
        representation = self.GRU.apply(embeddings, embeddings)
        return representation

# 需要导入模块: from blocks.bricks.recurrent import GatedRecurrent [as 别名]
# 或者: from blocks.bricks.recurrent.GatedRecurrent import apply [as 别名]
class GatedRecurrentWithContext(Initializable):
    def __init__(self, *args, **kwargs):
        self.gated_recurrent = GatedRecurrent(*args, **kwargs)
        self.children = [self.gated_recurrent]

    @application(states=['states'], outputs=['states'],
                 contexts=['readout_context', 'transition_context',
                           'update_context', 'reset_context'])
    def apply(self, transition_context, update_context, reset_context,
              *args, **kwargs):
        kwargs['inputs'] += transition_context
        kwargs['update_inputs'] += update_context
        kwargs['reset_inputs'] += reset_context
        # readout_context was only added for the Readout brick, discard it
        kwargs.pop('readout_context')
        return self.gated_recurrent.apply(*args, **kwargs)

    def get_dim(self, name):
        if name in ['readout_context', 'transition_context',
                    'update_context', 'reset_context']:
            return self.dim
        return self.gated_recurrent.get_dim(name)

    def __getattr__(self, name):
        if name == 'gated_recurrent':
            raise AttributeError
        return getattr(self.gated_recurrent, name)

    @apply.property('sequences')
    def apply_inputs(self):
        sequences = ['mask', 'inputs']
        if self.use_update_gate:
            sequences.append('update_inputs')
        if self.use_reset_gate:
            sequences.append('reset_inputs')
        return sequences

# 需要导入模块: from blocks.bricks.recurrent import GatedRecurrent [as 别名]
# 或者: from blocks.bricks.recurrent.GatedRecurrent import apply [as 别名]
    def __init__(self, config):
        inp = tensor.imatrix('bytes')

        embed = theano.shared(config.embedding_matrix.astype(theano.config.floatX),
                              name='embedding_matrix')
        in_repr = embed[inp.flatten(), :].reshape((inp.shape[0], inp.shape[1], config.repr_dim))
        in_repr.name = 'in_repr'

        bricks = []
        states = []

        # Construct predictive GRU hierarchy
        hidden = []
        costs = []
        next_target = in_repr.dimshuffle(1, 0, 2)
        for i, (hdim, cf, q) in enumerate(zip(config.hidden_dims,
                                                   config.cost_factors,
                                                   config.hidden_q)):
            init_state = theano.shared(numpy.zeros((config.num_seqs, hdim)).astype(theano.config.floatX),
                                       name='st0_%d'%i)

            linear = Linear(input_dim=config.repr_dim, output_dim=3*hdim,
                            name="lstm_in_%d"%i)
            lstm = GatedRecurrent(dim=hdim, activation=config.activation_function,
                        name="lstm_rec_%d"%i)
            linear2 = Linear(input_dim=hdim, output_dim=config.repr_dim, name='lstm_out_%d'%i)
            tanh = Tanh('lstm_out_tanh_%d'%i)
            bricks += [linear, lstm, linear2, tanh]
            if i > 0:
                linear1 = Linear(input_dim=config.hidden_dims[i-1], output_dim=3*hdim,
                                 name='lstm_in2_%d'%i)
                bricks += [linear1]

            next_target = tensor.cast(next_target, dtype=theano.config.floatX)
            inter = linear.apply(theano.gradient.disconnected_grad(next_target))
            if i > 0:
                inter += linear1.apply(theano.gradient.disconnected_grad(hidden[-1][:-1,:,:]))
            new_hidden = lstm.apply(inputs=inter[:,:,:hdim],
                                    gate_inputs=inter[:,:,hdim:],
                                    states=init_state)
            states.append((init_state, new_hidden[-1, :, :]))

            hidden += [tensor.concatenate([init_state[None,:,:], new_hidden],axis=0)]
            pred = tanh.apply(linear2.apply(hidden[-1][:-1,:,:]))
            costs += [numpy.float32(cf) * (-next_target * pred).sum(axis=2).mean()]
            costs += [numpy.float32(cf) * q * abs(pred).sum(axis=2).mean()]
            diff = next_target - pred
            next_target = tensor.ge(diff, 0.5) - tensor.le(diff, -0.5)


        # Construct output from hidden states
        hidden = [s.dimshuffle(1, 0, 2) for s in hidden]

        out_parts = []
        out_dims = config.out_hidden + [config.io_dim]
        for i, (dim, state) in enumerate(zip(config.hidden_dims, hidden)):
            pred_linear = Linear(input_dim=dim, output_dim=out_dims[0],
                                name='pred_linear_%d'%i)
            bricks.append(pred_linear)
            lin = theano.gradient.disconnected_grad(state)
            out_parts.append(pred_linear.apply(lin))

        # Do prediction and calculate cost
        out = sum(out_parts)

        if len(out_dims) > 1:
            out = config.out_hidden_act[0](name='out_act0').apply(out)
            mlp = MLP(dims=out_dims,
                      activations=[x(name='out_act%d'%i) for i, x in enumerate(config.out_hidden_act[1:])]
                                 +[Identity()],
                      name='out_mlp')
            bricks.append(mlp)
            out = mlp.apply(out.reshape((inp.shape[0]*(inp.shape[1]+1),-1))
                           ).reshape((inp.shape[0],inp.shape[1]+1,-1))

        pred = out.argmax(axis=2)

        cost = Softmax().categorical_cross_entropy(inp.flatten(),
                                                   out[:,:-1,:].reshape((inp.shape[0]*inp.shape[1],
                                                                config.io_dim))).mean()
        error_rate = tensor.neq(inp.flatten(), pred[:,:-1].flatten()).mean()

        sgd_cost = cost + sum(costs)
            
        # Initialize all bricks
        for brick in bricks:
            brick.weights_init = config.weights_init
            brick.biases_init = config.biases_init
            brick.initialize()

        # apply noise
        cg = ComputationGraph([sgd_cost, cost, error_rate]+costs)
        if config.weight_noise > 0:
            noise_vars = VariableFilter(roles=[WEIGHT])(cg)
            cg = apply_noise(cg, noise_vars, config.weight_noise)
        sgd_cost = cg.outputs[0]
        cost = cg.outputs[1]
        error_rate = cg.outputs[2]
        costs = cg.outputs[3:]

#.........这里部分代码省略.........

# 需要导入模块: from blocks.bricks.recurrent import GatedRecurrent [as 别名]
# 或者: from blocks.bricks.recurrent.GatedRecurrent import apply [as 别名]
class TestGatedRecurrent(unittest.TestCase):
    def setUp(self):
        self.gated = GatedRecurrent(
            dim=3, activation=Tanh(),
            gate_activation=Tanh(), weights_init=Constant(2))
        self.gated.initialize()
        self.reset_only = GatedRecurrent(
            dim=3, activation=Tanh(),
            gate_activation=Tanh(),
            weights_init=IsotropicGaussian(), seed=1)
        self.reset_only.initialize()

    def test_one_step(self):
        h0 = tensor.matrix('h0')
        x = tensor.matrix('x')
        gi = tensor.matrix('gi')
        h1 = self.gated.apply(x, gi, h0, iterate=False)
        next_h = theano.function(inputs=[h0, x, gi], outputs=[h1])

        h0_val = 0.1 * numpy.array([[1, 1, 0], [0, 1, 1]],
                                   dtype=theano.config.floatX)
        x_val = 0.1 * numpy.array([[1, 2, 3], [4, 5, 6]],
                                  dtype=theano.config.floatX)
        zi_val = (h0_val + x_val) / 2
        ri_val = -x_val
        W_val = 2 * numpy.ones((3, 3), dtype=theano.config.floatX)

        z_val = numpy.tanh(h0_val.dot(W_val) + zi_val)
        r_val = numpy.tanh(h0_val.dot(W_val) + ri_val)
        h1_val = (z_val * numpy.tanh((r_val * h0_val).dot(W_val) + x_val) +
                  (1 - z_val) * h0_val)
        assert_allclose(
            h1_val, next_h(h0_val, x_val, numpy.hstack([zi_val, ri_val]))[0],
            rtol=1e-6)

    def test_many_steps(self):
        x = tensor.tensor3('x')
        gi = tensor.tensor3('gi')
        mask = tensor.matrix('mask')
        h = self.reset_only.apply(x, gi, mask=mask)
        calc_h = theano.function(inputs=[x, gi, mask], outputs=[h])

        x_val = 0.1 * numpy.asarray(list(itertools.permutations(range(4))),
                                    dtype=theano.config.floatX)
        x_val = numpy.ones((24, 4, 3),
                           dtype=theano.config.floatX) * x_val[..., None]
        ri_val = 0.3 - x_val
        zi_val = 2 * ri_val
        mask_val = numpy.ones((24, 4), dtype=theano.config.floatX)
        mask_val[12:24, 3] = 0
        h_val = numpy.zeros((25, 4, 3), dtype=theano.config.floatX)
        W = self.reset_only.state_to_state.get_value()
        Wz = self.reset_only.state_to_gates.get_value()[:, :3]
        Wr = self.reset_only.state_to_gates.get_value()[:, 3:]

        for i in range(1, 25):
            z_val = numpy.tanh(h_val[i - 1].dot(Wz) + zi_val[i - 1])
            r_val = numpy.tanh(h_val[i - 1].dot(Wr) + ri_val[i - 1])
            h_val[i] = numpy.tanh((r_val * h_val[i - 1]).dot(W) +
                                  x_val[i - 1])
            h_val[i] = z_val * h_val[i] + (1 - z_val) * h_val[i - 1]
            h_val[i] = (mask_val[i - 1, :, None] * h_val[i] +
                        (1 - mask_val[i - 1, :, None]) * h_val[i - 1])
        h_val = h_val[1:]
        # TODO Figure out why this tolerance needs to be so big
        assert_allclose(
            h_val,
            calc_h(x_val, numpy.concatenate(
                [zi_val, ri_val], axis=2), mask_val)[0],
            1e-04)

        # Also test that initial state is a parameter
        initial_state, = VariableFilter(roles=[INITIAL_STATE])(
            ComputationGraph(h))
        assert is_shared_variable(initial_state)
        assert initial_state.name == 'initial_state'

# 需要导入模块: from blocks.bricks.recurrent import GatedRecurrent [as 别名]
# 或者: from blocks.bricks.recurrent.GatedRecurrent import apply [as 别名]
class TestGatedRecurrent(unittest.TestCase):
    def setUp(self):
        self.gated = GatedRecurrent(
            dim=3, weights_init=Constant(2),
            activation=Tanh(), gate_activation=Tanh())
        self.gated.initialize()
        self.reset_only = GatedRecurrent(
            dim=3, weights_init=IsotropicGaussian(),
            activation=Tanh(), gate_activation=Tanh(),
            use_update_gate=False, seed=1)
        self.reset_only.initialize()

    def test_one_step(self):
        h0 = tensor.matrix('h0')
        x = tensor.matrix('x')
        z = tensor.matrix('z')
        r = tensor.matrix('r')
        h1 = self.gated.apply(x, z, r, h0, iterate=False)
        next_h = theano.function(inputs=[h0, x, z, r], outputs=[h1])

        h0_val = 0.1 * numpy.array([[1, 1, 0], [0, 1, 1]],
                                   dtype=floatX)
        x_val = 0.1 * numpy.array([[1, 2, 3], [4, 5, 6]],
                                  dtype=floatX)
        zi_val = (h0_val + x_val) / 2
        ri_val = -x_val
        W_val = 2 * numpy.ones((3, 3), dtype=floatX)

        z_val = numpy.tanh(h0_val.dot(W_val) + zi_val)
        r_val = numpy.tanh(h0_val.dot(W_val) + ri_val)
        h1_val = (z_val * numpy.tanh((r_val * h0_val).dot(W_val) + x_val) +
                  (1 - z_val) * h0_val)
        assert_allclose(h1_val, next_h(h0_val, x_val, zi_val, ri_val)[0],
                        rtol=1e-6)

    def test_reset_only_many_steps(self):
        x = tensor.tensor3('x')
        ri = tensor.tensor3('ri')
        mask = tensor.matrix('mask')
        h = self.reset_only.apply(x, reset_inputs=ri, mask=mask)
        calc_h = theano.function(inputs=[x, ri, mask], outputs=[h])

        x_val = 0.1 * numpy.asarray(list(itertools.permutations(range(4))),
                                    dtype=floatX)
        x_val = numpy.ones((24, 4, 3), dtype=floatX) * x_val[..., None]
        ri_val = 0.3 - x_val
        mask_val = numpy.ones((24, 4), dtype=floatX)
        mask_val[12:24, 3] = 0
        h_val = numpy.zeros((25, 4, 3), dtype=floatX)
        W = self.reset_only.state_to_state.get_value()
        U = self.reset_only.state_to_reset.get_value()

        for i in range(1, 25):
            r_val = numpy.tanh(h_val[i - 1].dot(U) + ri_val[i - 1])
            h_val[i] = numpy.tanh((r_val * h_val[i - 1]).dot(W) +
                                  x_val[i - 1])
            h_val[i] = (mask_val[i - 1, :, None] * h_val[i] +
                        (1 - mask_val[i - 1, :, None]) * h_val[i - 1])
        h_val = h_val[1:]
        # TODO Figure out why this tolerance needs to be so big
        assert_allclose(h_val, calc_h(x_val, ri_val,  mask_val)[0], 1e-03)

# 需要导入模块: from blocks.bricks.recurrent import GatedRecurrent [as 别名]
# 或者: from blocks.bricks.recurrent.GatedRecurrent import apply [as 别名]

#.........这里部分代码省略.........
                weights_init=Constant(np.nan),
                dim=self.hidden_dim,
                activation=self.activation,
                gate_activation=self.gate_activation)
    .. [CvMG14] Kyunghyun Cho, Bart van Merriënboer, Çağlar Gülçehre,
        Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua
        Bengio, *Learning Phrase Representations using RNN Encoder-Decoder
        for Statistical Machine Translation*, EMNLP (2014), pp. 1724-1734.

    """
    @lazy(allocation=['hidden_dim', 'state_to_state_init', 'state_to_update_init', 'state_to_reset_init'],
            initialization=['input_to_state_transform', 'input_to_update_transform', 'input_to_reset_transform'])
    def __init__(self, hidden_dim, activation=None, gate_activation=None,
        state_to_state_init=None, state_to_update_init=None, state_to_reset_init=None,
        input_to_state_transform=None, input_to_update_transform=None, input_to_reset_transform=None,
        **kwargs):

        super(GatedRecurrentFull, self).__init__(**kwargs)
        self.hidden_dim = hidden_dim

        self.state_to_state_init = state_to_state_init
        self.state_to_update_init = state_to_update_init
        self.state_to_reset_init = state_to_reset_init

        self.input_to_state_transform = input_to_state_transform
        self.input_to_update_transform = input_to_update_transform
        self.input_to_reset_transform = input_to_reset_transform
        self.input_to_state_transform.name += "_input_to_state_transform"
        self.input_to_update_transform.name += "_input_to_update_transform"
        self.input_to_reset_transform.name += "_input_to_reset_transform"

        self.use_mine = True
        if self.use_mine:
            self.rnn = GatedRecurrentFast(
                    weights_init=Constant(np.nan),
                    dim=self.hidden_dim,
                    activation=activation,
                    gate_activation=gate_activation)
        else:
            self.rnn = GatedRecurrent(
                    weights_init=Constant(np.nan),
                    dim=self.hidden_dim,
                    activation=activation,
                    gate_activation=gate_activation)

        self.children = [self.rnn,
                self.input_to_state_transform, self.input_to_update_transform, self.input_to_reset_transform]
        self.children.extend(self.rnn.children)

    def initialize(self):
        super(GatedRecurrentFull, self).initialize()

        self.input_to_state_transform.initialize()
        self.input_to_update_transform.initialize()
        self.input_to_reset_transform.initialize()

        self.rnn.initialize()

        weight_shape = (self.hidden_dim, self.hidden_dim)
        state_to_state = self.state_to_state_init.generate(rng=self.rng, shape=weight_shape)
        state_to_update= self.state_to_update_init.generate(rng=self.rng, shape=weight_shape)
        state_to_reset = self.state_to_reset_init.generate(rng=self.rng, shape=weight_shape)

        self.rnn.state_to_state.set_value(state_to_state)

        if self.use_mine:
            self.rnn.state_to_update.set_value(state_to_update)
            self.rnn.state_to_reset.set_value(state_to_reset)
        else:
            self.rnn.state_to_gates.set_value(np.hstack((state_to_update, state_to_reset)))

    @application(inputs=['input_'], outputs=['output'])
    def apply(self, input_, mask=None):
        """

        Parameters
        ----------
        inputs_ : :class:`~tensor.TensorVariable`
            sequence to feed into GRU. Axes are mb, sequence, features

        mask : :class:`~tensor.TensorVariable`
            A 1D binary array with 1 or 0 to represent data given available.

        Returns
        -------
        output: :class:`theano.tensor.TensorVariable`
            sequence to feed out. Axes are batch, sequence, features
        """
        states_from_in = self.input_to_state_transform.apply(input_)
        update_from_in = self.input_to_update_transform.apply(input_)
        reset_from_in = self.input_to_reset_transform.apply(input_)

        gate_inputs = tensor.concatenate([update_from_in, reset_from_in], axis=2)

        if self.use_mine:
            output = self.rnn.apply(inputs=states_from_in, update_inputs=update_from_in, reset_inputs=reset_from_in, mask=mask)
        else:
            output = self.rnn.apply(inputs=states_from_in, gate_inputs=gate_inputs)

        return output

# 需要导入模块: from blocks.bricks.recurrent import GatedRecurrent [as 别名]
# 或者: from blocks.bricks.recurrent.GatedRecurrent import apply [as 别名]
iteration = 300 # number of epochs of gradient descent

print "Building Model"
# Symbolic variables
x = tensor.tensor3('x', dtype=floatX)
target = tensor.tensor3('target', dtype=floatX)

# Build the model
linear = Linear(input_dim = n_u, output_dim = n_h, name="first_layer")
rnn = GatedRecurrent(dim=n_h, activation=Tanh())
linear2 = Linear(input_dim = n_h, output_dim = n_y, name="output_layer")
sigm = Sigmoid()

x_transform = linear.apply(x)
h = rnn.apply(x_transform)
predict = sigm.apply(linear2.apply(h))


# only for generation B x h_dim
h_initial = tensor.tensor3('h_initial', dtype=floatX)
h_testing = rnn.apply(x_transform, h_initial, iterate=False)
y_hat_testing = linear2.apply(h_testing)
y_hat_testing = sigm.apply(y_hat_testing)
y_hat_testing.name = 'y_hat_testing'


# Cost function
cost = SquaredError().apply(predict,target)

# Initialization

# 需要导入模块: from blocks.bricks.recurrent import GatedRecurrent [as 别名]
# 或者: from blocks.bricks.recurrent.GatedRecurrent import apply [as 别名]

#.........这里部分代码省略.........

        return initial_h1, last_h1, initial_h2, last_h2, initial_h3, last_h3, \
            initial_w, last_w, initial_k, last_k

    @application
    def compute_cost(
            self, features, features_mask, labels, labels_mask,
            speaker, start_flag, batch_size, raw_audio=None):

        if speaker is None:
            assert not self.use_speaker

        target_features = features[1:]
        mask = features_mask[1:]

        cell_shape = (mask.shape[0], batch_size, self.rnn_h_dim)
        gat_shape = (mask.shape[0], batch_size, 2 * self.rnn_h_dim)
        cell_h1 = tensor.zeros(cell_shape, dtype=floatX)
        cell_h2 = tensor.zeros(cell_shape, dtype=floatX)
        cell_h3 = tensor.zeros(cell_shape, dtype=floatX)
        gat_h1 = tensor.zeros(gat_shape, dtype=floatX)
        gat_h2 = tensor.zeros(gat_shape, dtype=floatX)
        gat_h3 = tensor.zeros(gat_shape, dtype=floatX)

        if self.weak_feedback:
            input_features = features[:-1]

            if self.feedback_noise_level:
                noise = self.theano_rng.normal(
                    size=input_features.shape,
                    avg=0., std=1.)
                input_features += self.noise_level_var * noise

            out_cell_h1, out_gat_h1 = self.out_to_h1.apply(input_features)

            to_normalize = [
                out_cell_h1, out_gat_h1]
            out_cell_h1, out_gat_h1 = \
                [_apply_norm(x, self.layer_norm) for x in to_normalize]

            cell_h1 += out_cell_h1
            gat_h1 += out_gat_h1

        if self.full_feedback:
            assert self.weak_feedback
            out_cell_h2, out_gat_h2 = self.out_to_h2.apply(input_features)
            out_cell_h3, out_gat_h3 = self.out_to_h3.apply(input_features)

            to_normalize = [
                out_cell_h2, out_gat_h2, out_cell_h3, out_gat_h3]
            out_cell_h2, out_gat_h2, out_cell_h3, out_gat_h3 = \
                [_apply_norm(x, self.layer_norm) for x in to_normalize]

            cell_h2 += out_cell_h2
            gat_h2 += out_gat_h2
            cell_h3 += out_cell_h3
            gat_h3 += out_gat_h3

        if self.use_speaker:
            speaker = speaker[:, 0]
            emb_speaker = self.embed_speaker.apply(speaker)
            emb_speaker = tensor.shape_padleft(emb_speaker)

            spk_cell_h1, spk_gat_h1 = self.speaker_to_h1.apply(emb_speaker)
            spk_cell_h2, spk_gat_h2 = self.speaker_to_h2.apply(emb_speaker)
            spk_cell_h3, spk_gat_h3 = self.speaker_to_h3.apply(emb_speaker)