Python sequence_generators.SequenceGenerator类代码示例

def test_sequence_generator():
    # Disclaimer: here we only check shapes, not values.

    output_dim = 1
    dim = 20
    batch_size = 30
    n_steps = 10

    transition = GatedRecurrent(
        name="transition", activation=Tanh(), dim=dim,
        weights_init=Orthogonal())
    generator = SequenceGenerator(
        LinearReadout(readout_dim=output_dim, source_names=["states"],
                      emitter=TestEmitter(name="emitter"), name="readout"),
        transition,
        weights_init=IsotropicGaussian(0.01), biases_init=Constant(0),
        name="generator")
    generator.initialize()

    y = tensor.tensor3('y')
    mask = tensor.matrix('mask')
    costs = generator.cost(y, mask)
    assert costs.ndim == 2
    costs_val = theano.function([y, mask], [costs])(
        numpy.zeros((n_steps, batch_size, output_dim), dtype=floatX),
        numpy.ones((n_steps, batch_size), dtype=floatX))[0]
    assert costs_val.shape == (n_steps, batch_size)

    states, outputs, costs = [variable.eval() for variable in
                              generator.generate(
                                  iterate=True, batch_size=batch_size,
                                  n_steps=n_steps)]
    assert states.shape == (n_steps, batch_size, dim)
    assert outputs.shape == (n_steps, batch_size, output_dim)
    assert costs.shape == (n_steps, batch_size)

def getRnnGenerator(vocab_size,hidden_dim,input_dim=512):
    """
    "Apply" the RNN to the input x
    For initializing the network, the vocab size needs to be known
    Default of the hidden layer is set tot 512 like Karpathy
    """
    generator = SequenceGenerator(
        Readout(readout_dim = vocab_size,
                source_names = ["states"], # transition.apply.states ???
                emitter = SoftmaxEmitter(name="emitter"),
                feedback_brick = LookupFeedback(
                    vocab_size,
                    input_dim,
                    name = 'feedback'
                ),
                name = "readout"
        ),
        MySimpleRecurrent(
            name = "transition",
            activation = Tanh(),
            dim = hidden_dim
        ),
        weights_init = IsotropicGaussian(0.01),
        biases_init  = Constant(0),
        name = "generator"
    )
    generator.push_initialization_config()
    generator.transition.weights_init = IsotropicGaussian(0.01)
    generator.initialize()
    
    return generator

def test_integer_sequence_generator():
    # Disclaimer: here we only check shapes, not values.

    readout_dim = 5
    feedback_dim = 3
    dim = 20
    batch_size = 30
    n_steps = 10

    transition = GatedRecurrent(
        name="transition", activation=Tanh(), dim=dim,
        weights_init=Orthogonal())
    generator = SequenceGenerator(
        LinearReadout(readout_dim=readout_dim, source_names=["states"],
                      emitter=SoftmaxEmitter(name="emitter"),
                      feedbacker=LookupFeedback(readout_dim, feedback_dim),
                      name="readout"),
        transition,
        weights_init=IsotropicGaussian(0.01), biases_init=Constant(0),
        name="generator")
    generator.initialize()

    y = tensor.lmatrix('y')
    mask = tensor.matrix('mask')
    costs = generator.cost(y, mask)
    assert costs.ndim == 2
    costs_val = theano.function([y, mask], [costs])(
        numpy.zeros((n_steps, batch_size), dtype='int64'),
        numpy.ones((n_steps, batch_size), dtype=floatX))[0]
    assert costs_val.shape == (n_steps, batch_size)

    states, outputs, costs = generator.generate(
        iterate=True, batch_size=batch_size, n_steps=n_steps)
    states_val, outputs_val, costs_val = theano.function(
        [], [states, outputs, costs],
        updates=costs.owner.inputs[0].owner.tag.updates)()
    assert states_val.shape == (n_steps, batch_size, dim)
    assert outputs_val.shape == (n_steps, batch_size)
    assert outputs_val.dtype == 'int64'
    assert costs_val.shape == (n_steps, batch_size)

def test_recurrentstack_sequence_generator():
    """Test RecurrentStack behaviour inside a SequenceGenerator.

    """
    floatX = theano.config.floatX
    rng = numpy.random.RandomState(1234)

    output_dim = 1
    dim = 20
    batch_size = 30
    n_steps = 10

    depth=2
    transitions = [LSTM(dim=dim) for _ in range(depth)]
    transition = RecurrentStack(transitions,fast=True,
                                weights_init=Constant(2),
                                biases_init=Constant(0))
    generator = SequenceGenerator(
        Readout(readout_dim=output_dim, source_names=["states_%d"%(depth-1)],
                emitter=TestEmitter()),
        transition,
        weights_init=IsotropicGaussian(0.1), biases_init=Constant(0.0),
        seed=1234)
    generator.initialize()

    y = tensor.tensor3('y')

    cost = generator.cost(y)

    # Check that all states can be accessed and not just the state connected
    # to readout.
    cg = ComputationGraph(cost)
    from blocks.roles import INPUT, OUTPUT
    dropout_target = VariableFilter(roles=[INNER_OUTPUT],
                                    # bricks=transitions,
                                    # name_regex='*'
                                    )(cg.variables)
    assert_equal(len(dropout_target), depth)

    def __init__(self, vocab_size, embedding_dim, state_dim,
                 representation_dim, theano_seed=None, **kwargs):
        super(Decoder, self).__init__(**kwargs)
        self.vocab_size = vocab_size
        self.embedding_dim = embedding_dim
        self.state_dim = state_dim
        self.representation_dim = representation_dim
        self.theano_seed = theano_seed

        # Initialize gru with special initial state
        self.transition = GRUInitialState(
            attended_dim=state_dim, dim=state_dim,
            activation=Tanh(), name='decoder')

        # Initialize the attention mechanism
        self.attention = SequenceContentAttention(
            state_names=self.transition.apply.states,
            attended_dim=representation_dim,
            match_dim=state_dim, name="attention")

        # Initialize the readout, note that SoftmaxEmitter emits -1 for
        # initial outputs which is used by LookupFeedBackWMT15
        readout = Readout(
            source_names=['states', 'feedback',
                          self.attention.take_glimpses.outputs[0]],
            readout_dim=self.vocab_size,
            emitter=SoftmaxEmitter(initial_output=-1, theano_seed=theano_seed),
            feedback_brick=LookupFeedbackWMT15(vocab_size, embedding_dim),
            post_merge=InitializableFeedforwardSequence(
                [Bias(dim=state_dim, name='maxout_bias').apply,
                 Maxout(num_pieces=2, name='maxout').apply,
                 Linear(input_dim=state_dim / 2, output_dim=embedding_dim,
                        use_bias=False, name='softmax0').apply,
                 Linear(input_dim=embedding_dim, name='softmax1').apply]),
            merged_dim=state_dim)

        # Build sequence generator accordingly
        self.sequence_generator = SequenceGenerator(
            readout=readout,
            transition=self.transition,
            attention=self.attention,
            fork=Fork([name for name in self.transition.apply.sequences
                       if name != 'mask'], prototype=Linear())
        )

        self.children = [self.sequence_generator]

    def __init__(self, vocab_size, embedding_dim, 
            state_dim, theano_seed=None, **kwargs):
        super(Decoder, self).__init__(**kwargs)
        self.vocab_size = vocab_size
        self.embedding_dim = embedding_dim
        self.theano_seed = theano_seed

        self.transition = GatedRecurrent(dim=state_dim, 
                activation=Tanh(), name='decoder')

        readout = Readout(
                source_names=['states'],
                readout_dim=self.vocab_size,
                merged_dim=state_dim)

        self.sequence_generator = SequenceGenerator(
                readout=readout,
                transition=self.transition,
                fork=Fork([name for name in self.transition.apply.sequences
                    if name != 'mask'], prototype=Linear()))

        self.children = [self.sequence_generator]

    def __init__(self, vocab_size, embedding_dim, state_dim,
                 representation_dim, **kwargs):
        super(Decoder, self).__init__(**kwargs)
        self.vocab_size = vocab_size
        self.embedding_dim = embedding_dim
        self.state_dim = state_dim
        self.representation_dim = representation_dim

        self.transition = GRUInitialState(
            attended_dim=state_dim, dim=state_dim,
            activation=Tanh(), name='decoder')
        self.attention = SequenceContentAttention(
            state_names=self.transition.apply.states,
            attended_dim=representation_dim,
            match_dim=state_dim, name="attention")

        readout = Readout(
            source_names=['states', 'feedback', self.attention.take_glimpses.outputs[0]],
            readout_dim=self.vocab_size,
            emitter=SoftmaxEmitter(initial_output=-1),
            feedback_brick=LookupFeedbackWMT15(vocab_size, embedding_dim),
            post_merge=InitializableFeedforwardSequence(
                [Bias(dim=state_dim, name='maxout_bias').apply,
                 Maxout(num_pieces=2, name='maxout').apply,
                 Linear(input_dim=state_dim / 2, output_dim=embedding_dim,
                        use_bias=False, name='softmax0').apply,
                 Linear(input_dim=embedding_dim, name='softmax1').apply]),
            merged_dim=state_dim,
            merge_prototype=Linear(use_bias=True))

        self.sequence_generator = SequenceGenerator(
            readout=readout,
            transition=self.transition,
            attention=self.attention,
            fork=Fork([name for name in self.transition.apply.sequences
                       if name != 'mask'], prototype=Linear())
        )

        self.children = [self.sequence_generator]

    def __init__(self, vocab_size, embedding_dim, state_dim,
                 representation_dim, **kwargs):
        super(Decoder, self).__init__(**kwargs)
        self.vocab_size = vocab_size
        self.embedding_dim = embedding_dim
        self.state_dim = state_dim
        self.representation_dim = representation_dim

        readout = Readout(
            source_names=['states', 'feedback', 'readout_context'],
            readout_dim=self.vocab_size,
            emitter=SoftmaxEmitter(),
            feedback_brick=LookupFeedback(vocab_size, embedding_dim),
            post_merge=InitializableFeedforwardSequence(
                [Bias(dim=1000).apply,
                 Maxout(num_pieces=2).apply,
                 Linear(input_dim=state_dim / 2, output_dim=100,
                        use_bias=False).apply,
                 Linear(input_dim=100).apply]),
            merged_dim=1000)

        self.transition = GatedRecurrentWithContext(Tanh(), dim=state_dim,
                                                    name='decoder')
        # Readout will apply the linear transformation to 'readout_context'
        # with a Merge brick, so no need to fork it here
        self.fork = Fork([name for name in
                          self.transition.apply.contexts +
                          self.transition.apply.states
                          if name != 'readout_context'], prototype=Linear())
        self.tanh = Tanh()

        self.sequence_generator = SequenceGenerator(
            readout=readout, transition=self.transition,
            fork_inputs=[name for name in self.transition.apply.sequences
                         if name != 'mask'],
        )

        self.children = [self.fork, self.sequence_generator, self.tanh]

def build_model(alphabet_size, config):
    layers = config['lstm_layers']
    dimensions = [config['lstm_dim_' + str(i)] for i in range(layers)]
    uniform_width = config['lstm_init_width']
    stack = []
    for dim in dimensions:
        stack.append(LSTM(dim=dim, use_bias=True, 
                          weights_init = Uniform(width=uniform_width),
                          forget_init=Constant(1.)))
    recurrent_stack = RecurrentStack(stack, name='transition')

    readout = Readout(readout_dim=alphabet_size,
                      source_names=['states#' + str(layers - 1)],
                      emitter=SoftmaxEmitter(name='emitter'),
                      feedback_brick=LookupFeedback(alphabet_size,
                                                    feedback_dim=alphabet_size,
                                                    name='feedback'),
                      name='readout')

    generator = SequenceGenerator(readout=readout,
                                  transition=recurrent_stack,
                                  weights_init=Uniform(width=uniform_width),
                                  biases_init=Constant(0),
                                  name='generator')
    generator.push_initialization_config()
    generator.initialize()

    x = tensor.lmatrix('features')
    mask = tensor.fmatrix('features_mask')
    cost_matrix = generator.cost_matrix(x, mask=mask)

    log2e = math.log(math.e, 2)
    if 'batch_length' in config:
        length = config['batch_length'] - config['batch_overlap']

        cost = log2e * aggregation.mean(cost_matrix[:,-length:].sum(), 
                                    mask[:,-length:].sum())
    else:
        cost = log2e * aggregation.mean(cost_matrix[:,:].sum(), 
                                    mask[:,:].sum())
        
    cost.name = 'bits_per_character'

    return generator, cost

class Decoder(Initializable):
    def __init__(self, vocab_size, embedding_dim, 
            state_dim, theano_seed=None, **kwargs):
        super(Decoder, self).__init__(**kwargs)
        self.vocab_size = vocab_size
        self.embedding_dim = embedding_dim
        self.theano_seed = theano_seed

        self.transition = GatedRecurrent(dim=state_dim, 
                activation=Tanh(), name='decoder')

        readout = Readout(
                source_names=['states'],
                readout_dim=self.vocab_size,
                merged_dim=state_dim)

        self.sequence_generator = SequenceGenerator(
                readout=readout,
                transition=self.transition,
                fork=Fork([name for name in self.transition.apply.sequences
                    if name != 'mask'], prototype=Linear()))

        self.children = [self.sequence_generator]

    @application(inputs=['representation', 'source_sentence_mask',
                         'target_sentence_mask', 'target_sentence'],
                 outputs=['cost'])
    def cost(self, representation, source_sentence_mask,
            target_sentence, target_sentence_mask):
        source_sentence_mask = source_sentence_mask.T
        target_sentence = target_sentence.T
        target_sentence_mask = target_sentence_mask.T

        cost = self.sequence_generator.cost_matrix(**{
            'mask': target_sentence_mask,
            'outputs': target_sentence})

def test_sequence_generator():
    """Test a sequence generator with no contexts and continuous outputs.

    Such sequence generators can be used to model e.g. dynamical systems.

    """
    rng = numpy.random.RandomState(1234)

    output_dim = 1
    dim = 20
    batch_size = 30
    n_steps = 10

    transition = SimpleRecurrent(activation=Tanh(), dim=dim,
                                 weights_init=Orthogonal())
    generator = SequenceGenerator(
        Readout(readout_dim=output_dim, source_names=["states"],
                emitter=TestEmitter()),
        transition,
        weights_init=IsotropicGaussian(0.1), biases_init=Constant(0.0),
        seed=1234)
    generator.initialize()

    # Test 'cost_matrix' method
    y = tensor.tensor3('y')
    mask = tensor.matrix('mask')
    costs = generator.cost_matrix(y, mask)
    assert costs.ndim == 2
    y_test = rng.uniform(size=(n_steps, batch_size, output_dim)).astype(floatX)
    m_test = numpy.ones((n_steps, batch_size), dtype=floatX)
    costs_val = theano.function([y, mask], [costs])(y_test, m_test)[0]
    assert costs_val.shape == (n_steps, batch_size)
    assert_allclose(costs_val.sum(), 115.593, rtol=1e-5)

    # Test 'cost' method
    cost = generator.cost(y, mask)
    assert cost.ndim == 0
    cost_val = theano.function([y, mask], [cost])(y_test, m_test)
    assert_allclose(cost_val, 3.8531, rtol=1e-5)

    # Test 'AUXILIARY' variable 'per_sequence_element' in 'cost' method
    cg = ComputationGraph([cost])
    var_filter = VariableFilter(roles=[AUXILIARY])
    aux_var_name = '_'.join([generator.name, generator.cost.name,
                             'per_sequence_element'])
    cost_per_el = [el for el in var_filter(cg.variables)
                   if el.name == aux_var_name][0]
    assert cost_per_el.ndim == 0
    cost_per_el_val = theano.function([y, mask], [cost_per_el])(y_test, m_test)
    assert_allclose(cost_per_el_val, 0.38531, rtol=1e-5)

    # Test 'generate' method
    states, outputs, costs = [variable.eval() for variable in
                              generator.generate(
                                  states=rng.uniform(
                                      size=(batch_size, dim)).astype(floatX),
                                  iterate=True, batch_size=batch_size,
                                  n_steps=n_steps)]
    assert states.shape == (n_steps, batch_size, dim)
    assert outputs.shape == (n_steps, batch_size, output_dim)
    assert costs.shape == (n_steps, batch_size)
    assert_allclose(outputs.sum(), -0.33683, rtol=1e-5)
    assert_allclose(states.sum(), 15.7909, rtol=1e-5)
    # There is no generation cost in this case, since generation is
    # deterministic
    assert_allclose(costs.sum(), 0.0)

def main():
    logging.basicConfig(
        level=logging.DEBUG,
        format="%(asctime)s: %(name)s: %(levelname)s: %(message)s")

    parser = argparse.ArgumentParser(
        "Case study of language modeling with RNN",
        formatter_class=argparse.ArgumentDefaultsHelpFormatter)
    parser.add_argument(
        "mode", choices=["train", "sample"],
        help="The mode to run. Use `train` to train a new model"
             " and `sample` to sample a sequence generated by an"
             " existing one.")
    parser.add_argument(
        "prefix", default="sine",
        help="The prefix for model, timing and state files")
    parser.add_argument(
        "state", nargs="?", default="",
        help="Changes to Groundhog state")
    parser.add_argument("--path", help="Path to a language dataset")
    parser.add_argument("--dict", help="Path to the dataset dictionary")
    parser.add_argument("--restart", help="Start anew")
    parser.add_argument(
        "--reset", action="store_true", default=False,
        help="Reset the hidden state between batches")
    parser.add_argument(
        "--steps", type=int, default=100,
        help="Number of steps to plot for the 'sample' mode"
             " OR training sequence length for the 'train' mode.")
    args = parser.parse_args()
    logger.debug("Args:\n" + str(args))

    dim = 200
    num_chars = 50

    transition = GatedRecurrent(
        name="transition", activation=Tanh(), dim=dim,
        weights_init=Orthogonal())
    generator = SequenceGenerator(
        LinearReadout(readout_dim=num_chars, source_names=["states"],
                      emitter=SoftmaxEmitter(name="emitter"),
                      feedbacker=LookupFeedback(
                          num_chars, dim, name='feedback'),
                      name="readout"),
        transition,
        weights_init=IsotropicGaussian(0.01), biases_init=Constant(0),
        name="generator")
    generator.allocate()
    logger.debug("Parameters:\n" +
                 pprint.pformat(
                     [(key, value.get_value().shape) for key, value
                      in Selector(generator).get_params().items()],
                     width=120))

    if args.mode == "train":
        batch_size = 1
        seq_len = args.steps

        generator.initialize()

        # Build cost computation graph that uses the saved hidden states.
        # An issue: for Groundhog this is completely transparent, that's
        # why it does not carry the hidden state over the period when
        # validation in done. We should find a way to fix in the future.
        x = tensor.lmatrix('x')
        init_states = shared_floatx_zeros((batch_size, dim),
                                          name='init_states')
        reset = tensor.scalar('reset')
        cost = ComputationGraph(
            generator.cost(x, states=init_states * reset).sum())
        # TODO: better search routine
        states = [v for v in cost.variables
                  if hasattr(v.tag, 'application_call')
                  and v.tag.application_call.brick == generator.transition
                  and (v.tag.application_call.application ==
                       generator.transition.apply)
                  and v.tag.role == VariableRole.OUTPUT
                  and v.tag.name == 'states']
        assert len(states) == 1
        states = states[0]

        gh_model = GroundhogModel(generator, cost)
        gh_model.properties.append(
            ('bpc', cost.outputs[0] * numpy.log(2) / seq_len))
        gh_model.properties.append(('mean_init_state', init_states.mean()))
        gh_model.properties.append(('reset', reset))
        if not args.reset:
            gh_model.updates.append((init_states, states[-1]))

        state = GroundhogState(args.prefix, batch_size,
                               learning_rate=0.0001).as_dict()
        changes = eval("dict({})".format(args.state))
        state.update(changes)

        def output_format(x, y, reset):
            return dict(x=x[:, None], reset=reset)
        train, valid, test = [
            LMIterator(batch_size=batch_size,
                       use_infinite_loop=mode == 'train',
                       path=args.path,
#.........这里部分代码省略.........

def test_with_attention():
    """Test a sequence generator with continuous outputs and attention."""
    rng = numpy.random.RandomState(1234)

    inp_dim = 2
    inp_len = 10
    attended_dim = 3
    attended_len = 11
    batch_size = 4
    n_steps = 30

    # For values
    def rand(size):
        return rng.uniform(size=size).astype(floatX)

    # For masks
    def generate_mask(length, batch_size):
        mask = numpy.ones((length, batch_size), dtype=floatX)
        # To make it look like read data
        for i in range(batch_size):
            mask[1 + rng.randint(0, length - 1):, i] = 0.0
        return mask

    output_vals = rand((inp_len, batch_size, inp_dim))
    output_mask_vals = generate_mask(inp_len, batch_size)
    attended_vals = rand((attended_len, batch_size, attended_dim))
    attended_mask_vals = generate_mask(attended_len, batch_size)

    transition = TestTransition(
        dim=inp_dim, attended_dim=attended_dim, activation=Identity())
    attention = SequenceContentAttention(
        state_names=transition.apply.states, match_dim=inp_dim)
    generator = SequenceGenerator(
        Readout(
            readout_dim=inp_dim,
            source_names=[transition.apply.states[0],
                          attention.take_glimpses.outputs[0]],
            emitter=TestEmitter()),
        transition=transition,
        attention=attention,
        weights_init=IsotropicGaussian(0.1), biases_init=Constant(0),
        add_contexts=False, seed=1234)
    generator.initialize()

    # Test 'cost_matrix' method
    attended = tensor.tensor3("attended")
    attended_mask = tensor.matrix("attended_mask")
    outputs = tensor.tensor3('outputs')
    mask = tensor.matrix('mask')
    costs = generator.cost_matrix(outputs, mask,
                                  attended=attended,
                                  attended_mask=attended_mask)
    costs_vals = costs.eval({outputs: output_vals,
                             mask: output_mask_vals,
                             attended: attended_vals,
                             attended_mask: attended_mask_vals})
    assert costs_vals.shape == (inp_len, batch_size)
    assert_allclose(costs_vals.sum(), 13.5042, rtol=1e-5)

    # Test `generate` method
    results = (
        generator.generate(n_steps=n_steps, batch_size=attended.shape[1],
                           attended=attended, attended_mask=attended_mask))
    assert len(results) == 5
    states_vals, outputs_vals, glimpses_vals, weights_vals, costs_vals = (
        theano.function([attended, attended_mask], results)
        (attended_vals, attended_mask_vals))
    assert states_vals.shape == (n_steps, batch_size, inp_dim)
    assert states_vals.shape == outputs_vals.shape
    assert glimpses_vals.shape == (n_steps, batch_size, attended_dim)
    assert weights_vals.shape == (n_steps, batch_size, attended_len)
    assert costs_vals.shape == (n_steps, batch_size)
    assert_allclose(states_vals.sum(), 23.4172, rtol=1e-5)
    # There is no generation cost in this case, since generation is
    # deterministic
    assert_allclose(costs_vals.sum(), 0.0, rtol=1e-5)
    assert_allclose(weights_vals.sum(), 120.0, rtol=1e-5)
    assert_allclose(glimpses_vals.sum(), 199.2402, rtol=1e-5)
    assert_allclose(outputs_vals.sum(), -11.6008, rtol=1e-5)

def main():
    logging.basicConfig(
        level=logging.DEBUG,
        format="%(asctime)s: %(name)s: %(levelname)s: %(message)s")

    parser = argparse.ArgumentParser(
        "Case study of generating a Markov chain with RNN.",
        formatter_class=argparse.ArgumentDefaultsHelpFormatter)
    parser.add_argument(
        "mode", choices=["train", "sample"],
        help="The mode to run. Use `train` to train a new model"
             " and `sample` to sample a sequence generated by an"
             " existing one.")
    parser.add_argument(
        "prefix", default="sine",
        help="The prefix for model, timing and state files")
    parser.add_argument(
        "--steps", type=int, default=100,
        help="Number of steps to plot")
    args = parser.parse_args()

    dim = 10
    num_states = ChainIterator.num_states
    feedback_dim = 8

    transition = GatedRecurrent(name="transition", activation=Tanh(), dim=dim)
    generator = SequenceGenerator(
        LinearReadout(readout_dim=num_states, source_names=["states"],
                      emitter=SoftmaxEmitter(name="emitter"),
                      feedbacker=LookupFeedback(
                          num_states, feedback_dim, name='feedback'),
                      name="readout"),
        transition,
        weights_init=IsotropicGaussian(0.01), biases_init=Constant(0),
        name="generator")
    generator.allocate()
    logger.debug("Parameters:\n" +
                 pprint.pformat(
                     [(key, value.get_value().shape) for key, value
                      in Selector(generator).get_params().items()],
                     width=120))

    if args.mode == "train":
        rng = numpy.random.RandomState(1)
        batch_size = 50

        generator.push_initialization_config()
        transition.weights_init = Orthogonal()
        generator.initialize()
        logger.debug("transition.weights_init={}".format(
            transition.weights_init))

        cost = generator.cost(tensor.lmatrix('x')).sum()
        gh_model = GroundhogModel(generator, cost)
        state = GroundhogState(args.prefix, batch_size,
                               learning_rate=0.0001).as_dict()
        data = ChainIterator(rng, 100, batch_size)
        trainer = SGD(gh_model, state, data)
        main_loop = MainLoop(data, None, None, gh_model, trainer, state, None)
        main_loop.main()
    elif args.mode == "sample":
        load_params(generator,  args.prefix + "model.npz")

        sample = ComputationGraph(generator.generate(
            n_steps=args.steps, batch_size=1, iterate=True)).function()

        states, outputs, costs = [data[:, 0] for data in sample()]

        numpy.set_printoptions(precision=3, suppress=True)
        print("Generation cost:\n{}".format(costs.sum()))

        freqs = numpy.bincount(outputs).astype(floatX)
        freqs /= freqs.sum()
        print("Frequencies:\n {} vs {}".format(freqs,
                                               ChainIterator.equilibrium))

        trans_freqs = numpy.zeros((num_states, num_states), dtype=floatX)
        for a, b in zip(outputs, outputs[1:]):
            trans_freqs[a, b] += 1
        trans_freqs /= trans_freqs.sum(axis=1)[:, None]
        print("Transition frequencies:\n{}\nvs\n{}".format(
            trans_freqs, ChainIterator.trans_prob))
    else:
        assert False

def main(name, epochs, batch_size, learning_rate,
         dim, mix_dim, old_model_name, max_length, bokeh, GRU, dropout,
         depth, max_grad, step_method, epsilon, sample, skip, uniform, top):

    #----------------------------------------------------------------------
    datasource = name

    def shnum(x):
        """ Convert a positive float into a short tag-usable string
             E.g.: 0 -> 0, 0.005 -> 53, 100 -> 1-2
        """
        return '0' if x <= 0 else '%s%d' % (("%e"%x)[0], -np.floor(np.log10(x)))

    jobname = "%s-%dX%dm%dd%dr%sb%de%s" % (datasource, depth, dim, mix_dim,
                                           int(dropout*10),
                                           shnum(learning_rate), batch_size,
                                           shnum(epsilon))
    if max_length != 600:
        jobname += '-L%d'%max_length

    if GRU:
        jobname += 'g'
    if max_grad != 5.:
        jobname += 'G%g'%max_grad
    if step_method != 'adam':
        jobname += step_method
    if skip:
        jobname += 'D'
        assert depth > 1
    if top:
        jobname += 'T'
        assert depth > 1
    if uniform > 0.:
        jobname += 'u%d'%int(uniform*100)

    if debug:
        jobname += ".debug"

    if sample:
        print("Sampling")
    else:
        print("\nRunning experiment %s" % jobname)
    if old_model_name:
        print("starting from model %s"%old_model_name)

    #----------------------------------------------------------------------
    transitions = [GatedRecurrent(dim=dim) if GRU else LSTM(dim=dim)
                   for _ in range(depth)]
    if depth > 1:
        transition = RecurrentStack(transitions, name="transition",
                                    fast=True, skip_connections=skip or top)
        if skip:
            source_names=['states'] + ['states_%d'%d for d in range(1,depth)]
        else:
            source_names=['states_%d'%(depth-1)]
    else:
        transition = transitions[0]
        transition.name = "transition"
        source_names=['states']

    emitter = SketchEmitter(mix_dim=mix_dim,
                            epsilon=epsilon,
                            name="emitter")
    readout = Readout(
        readout_dim=emitter.get_dim('inputs'),
        source_names=source_names,
        emitter=emitter,
        name="readout")
    normal_inputs = [name for name in transition.apply.sequences
                     if 'mask' not in name]
    fork = Fork(normal_inputs, prototype=Linear(use_bias=True))
    generator = SequenceGenerator(readout=readout, transition=transition,
                                  fork=fork)

    # Initialization settings
    if uniform > 0.:
        generator.weights_init = Uniform(width=uniform*2.)
    else:
        generator.weights_init = OrthogonalGlorot()
    generator.biases_init = Constant(0)

    # Build the cost computation graph [steps, batch_size, 3]
    x = T.tensor3('features', dtype=floatX)
    if debug:
        x.tag.test_value = np.ones((max_length,batch_size,3)).astype(floatX)
    x = x[:max_length,:,:]  # has to be after setting test_value
    cost = generator.cost(x)
    cost.name = "sequence_log_likelihood"

    # Give an idea of what's going on
    model = Model(cost)
    params = model.get_params()
    logger.info("Parameters:\n" +
                pprint.pformat(
                    [(key, value.get_value().shape) for key, value
                     in params.items()],
                    width=120))
    model_size = 0
    for v in params.itervalues():
        s = v.get_value().shape
#.........这里部分代码省略.........