Python GatedRecurrent.weights_init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 weights_init [as 别名]
def main(mode, save_path, steps, num_batches):
    num_states = MarkovChainDataset.num_states

    if mode == "train":
        # Experiment configuration
        rng = numpy.random.RandomState(1)
        batch_size = 50
        seq_len = 100
        dim = 10
        feedback_dim = 8

        # Build the bricks and initialize them
        transition = GatedRecurrent(name="transition", dim=dim,
                                    activation=Tanh())
        generator = SequenceGenerator(
            Readout(readout_dim=num_states, source_names=["states"],
                    emitter=SoftmaxEmitter(name="emitter"),
                    feedback_brick=LookupFeedback(
                        num_states, feedback_dim, name='feedback'),
                    name="readout"),
            transition,
            weights_init=IsotropicGaussian(0.01), biases_init=Constant(0),
            name="generator")
        generator.push_initialization_config()
        transition.weights_init = Orthogonal()
        generator.initialize()

        # Give an idea of what's going on.
        logger.info("Parameters:\n" +
                    pprint.pformat(
                        [(key, value.get_value().shape) for key, value
                         in Selector(generator).get_params().items()],
                        width=120))
        logger.info("Markov chain entropy: {}".format(
            MarkovChainDataset.entropy))
        logger.info("Expected min error: {}".format(
            -MarkovChainDataset.entropy * seq_len))

        # Build the cost computation graph.
        x = tensor.lmatrix('data')
        cost = aggregation.mean(generator.cost_matrix(x[:, :]).sum(),
                                x.shape[1])
        cost.name = "sequence_log_likelihood"

        algorithm = GradientDescent(
            cost=cost, params=list(Selector(generator).get_params().values()),
            step_rule=Scale(0.001))
        main_loop = MainLoop(
            algorithm=algorithm,
            data_stream=DataStream(
                MarkovChainDataset(rng, seq_len),
                iteration_scheme=ConstantScheme(batch_size)),
            model=Model(cost),
            extensions=[FinishAfter(after_n_batches=num_batches),
                        TrainingDataMonitoring([cost], prefix="this_step",
                                               after_batch=True),
                        TrainingDataMonitoring([cost], prefix="average",
                                               every_n_batches=100),
                        Checkpoint(save_path, every_n_batches=500),
                        Printing(every_n_batches=100)])
        main_loop.run()
    elif mode == "sample":
        main_loop = cPickle.load(open(save_path, "rb"))
        generator = main_loop.model

        sample = ComputationGraph(generator.generate(
            n_steps=steps, batch_size=1, iterate=True)).get_theano_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(theano.config.floatX)
        freqs /= freqs.sum()
        print("Frequencies:\n {} vs {}".format(freqs,
                                               MarkovChainDataset.equilibrium))

        trans_freqs = numpy.zeros((num_states, num_states),
                                  dtype=theano.config.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, MarkovChainDataset.trans_prob))
    else:
        assert False

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

# 需要导入模块: from blocks.bricks.recurrent import GatedRecurrent [as 别名]
# 或者: from blocks.bricks.recurrent.GatedRecurrent import weights_init [as 别名]
def train():

    if os.path.isfile('trainingdata.tar'):
        with open('trainingdata.tar', 'rb') as f:
            main = load(f)
    else:
        hidden_size = 512
        filename = 'warpeace.hdf5'

        encoder = HDF5CharEncoder('warpeace_input.txt', 1000)
        encoder.write(filename)
        alphabet_len = encoder.length

        x = theano.tensor.lmatrix('x')

        readout = Readout(
            readout_dim=alphabet_len,
            feedback_brick=LookupFeedback(alphabet_len, hidden_size, name='feedback'),
            source_names=['states'],
            emitter=RandomSoftmaxEmitter(),
            name='readout'
        )

        transition = GatedRecurrent(
            activation=Tanh(),
            dim=hidden_size)
        transition.weights_init = IsotropicGaussian(0.01)

        gen = SequenceGenerator(readout=readout,
                                transition=transition,
                                weights_init=IsotropicGaussian(0.01),
                                biases_init=Constant(0),
                                name='sequencegenerator')

        gen.push_initialization_config()
        gen.initialize()

        cost = gen.cost(outputs=x)
        cost.name = 'cost'

        cg = ComputationGraph(cost)

        algorithm = GradientDescent(cost=cost,
                                    parameters=cg.parameters,
                                    step_rule=Scale(0.5))

        train_set = encoder.get_dataset()
        train_stream = DataStream.default_stream(
            train_set, iteration_scheme=SequentialScheme(
                train_set.num_examples, batch_size=128))

        main = MainLoop(
            model=Model(cost),
            data_stream=train_stream,
            algorithm=algorithm,
            extensions=[
                FinishAfter(),
                Printing(),
                Checkpoint('trainingdata.tar', every_n_epochs=10),
                ShowOutput(every_n_epochs=10)
            ])

    main.run()

# 需要导入模块: from blocks.bricks.recurrent import GatedRecurrent [as 别名]
# 或者: from blocks.bricks.recurrent.GatedRecurrent import weights_init [as 别名]
def main(mode, save_path, steps, num_batches, load_params):
    chars = (list(string.ascii_uppercase) + list(range(10)) +
             [' ', '.', ',', '\'', '"', '!', '?', '<UNK>'])
    char_to_ind = {char: i for i, char in enumerate(chars)}
    ind_to_char = {v: k for k, v in char_to_ind.iteritems()}

    train_dataset = TextFile(['/Tmp/serdyuk/data/wsj_text_train'],
                             char_to_ind, bos_token=None, eos_token=None,
                             level='character')
    valid_dataset = TextFile(['/Tmp/serdyuk/data/wsj_text_valid'],
                             char_to_ind, bos_token=None, eos_token=None,
                             level='character')

    vocab_size = len(char_to_ind)
    logger.info('Dictionary size: {}'.format(vocab_size))
    if mode == 'continue':
        continue_training(save_path)
        return
    elif mode == "sample":
        main_loop = load(open(save_path, "rb"))
        generator = main_loop.model.get_top_bricks()[-1]

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

        states, outputs, costs = [data[:, 0] for data in sample()]
        print("".join([ind_to_char[s] for s in outputs]))

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

        freqs = numpy.bincount(outputs).astype(floatX)
        freqs /= freqs.sum()

        trans_freqs = numpy.zeros((vocab_size, vocab_size), dtype=floatX)
        for a, b in zip(outputs, outputs[1:]):
            trans_freqs[a, b] += 1
        trans_freqs /= trans_freqs.sum(axis=1)[:, None]
        return

    # Experiment configuration
    batch_size = 20
    dim = 650
    feedback_dim = 650

    valid_stream = valid_dataset.get_example_stream()
    valid_stream = Batch(valid_stream,
                         iteration_scheme=ConstantScheme(batch_size))
    valid_stream = Padding(valid_stream)
    valid_stream = Mapping(valid_stream, _transpose)

    # Build the bricks and initialize them

    transition = GatedRecurrent(name="transition", dim=dim,
                                activation=Tanh())
    generator = SequenceGenerator(
        Readout(readout_dim=vocab_size, source_names=transition.apply.states,
                emitter=SoftmaxEmitter(name="emitter"),
                feedback_brick=LookupFeedback(
                    vocab_size, feedback_dim, name='feedback'),
                name="readout"),
        transition,
        weights_init=Uniform(std=0.04), biases_init=Constant(0),
        name="generator")
    generator.push_initialization_config()
    transition.weights_init = Orthogonal()
    transition.push_initialization_config()
    generator.initialize()

    # Build the cost computation graph.
    features = tensor.lmatrix('features')
    features_mask = tensor.matrix('features_mask')
    cost_matrix = generator.cost_matrix(
        features, mask=features_mask)
    batch_cost = cost_matrix.sum()
    cost = aggregation.mean(
        batch_cost,
        features.shape[1])
    cost.name = "sequence_log_likelihood"
    char_cost = aggregation.mean(
        batch_cost, features_mask.sum())
    char_cost.name = 'character_log_likelihood'
    ppl = 2 ** (cost / numpy.log(2))
    ppl.name = 'ppl'
    bits_per_char = char_cost / tensor.log(2)
    bits_per_char.name = 'bits_per_char'
    length = features.shape[0]
    length.name = 'length'

    model = Model(batch_cost)
    if load_params:
        params = load_parameter_values(save_path)
        model.set_parameter_values(params)

    if mode == "train":
        # Give an idea of what's going on.
        logger.info("Parameters:\n" +
                    pprint.pformat(
                        [(key, value.get_value().shape) for key, value
                         in Selector(generator).get_parameters().items()],
#.........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 weights_init [as 别名]
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(
        "save_path", default="sine",
        help="The part to save PyLearn2 model")
    parser.add_argument(
        "--steps", type=int, default=100,
        help="Number of steps to plot")
    parser.add_argument(
        "--reset", action="store_true", default=False,
        help="Start training from scratch")
    args = parser.parse_args()

    num_states = ChainDataset.num_states

    if args.mode == "train":
        # Experiment configuration
        rng = numpy.random.RandomState(1)
        batch_size = 50
        seq_len = 100
        dim = 10
        feedback_dim = 8

        # Build the bricks and initialize them
        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.push_initialization_config()
        transition.weights_init = Orthogonal()
        generator.initialize()

        logger.debug("Parameters:\n" +
                     pprint.pformat(
                         [(key, value.get_value().shape) for key, value
                          in Selector(generator).get_params().items()],
                         width=120))
        logger.debug("Markov chain entropy: {}".format(
            ChainDataset.entropy))
        logger.debug("Expected min error: {}".format(
            -ChainDataset.entropy * seq_len * batch_size))

        if os.path.isfile(args.save_path) and not args.reset:
            model = Pylearn2Model.load(args.save_path)
        else:
            model = Pylearn2Model(generator)

        # Build the cost computation graph.
        # Note: would be probably nicer to make cost part of the model.
        x = tensor.ltensor3('x')
        cost = Pylearn2Cost(model.brick.cost(x[:, :, 0]).sum())

        dataset = ChainDataset(rng, seq_len)
        sgd = SGD(learning_rate=0.0001, cost=cost,
                  batch_size=batch_size, batches_per_iter=10,
                  monitoring_dataset=dataset,
                  monitoring_batch_size=batch_size,
                  monitoring_batches=1,
                  learning_rule=Pylearn2LearningRule(
                      SGDLearningRule(),
                      dict(training_objective=cost.cost)))
        train = Pylearn2Train(dataset, model, algorithm=sgd,
                              save_path=args.save_path, save_freq=10)
        train.main_loop()
    elif args.mode == "sample":
        model = Pylearn2Model.load(args.save_path)
        generator = model.brick

        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,
                                               ChainDataset.equilibrium))

        trans_freqs = numpy.zeros((num_states, num_states), dtype=floatX)
#.........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 weights_init [as 别名]
def main(mode, save_path, steps, time_budget, reset):

    num_states = ChainDataset.num_states

    if mode == "train":
        # Experiment configuration
        rng = numpy.random.RandomState(1)
        batch_size = 50
        seq_len = 100
        dim = 10
        feedback_dim = 8

        # Build the bricks and initialize them
        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.push_initialization_config()
        transition.weights_init = Orthogonal()
        generator.initialize()

        logger.info("Parameters:\n" +
                    pprint.pformat(
                        [(key, value.get_value().shape) for key, value
                         in Selector(generator).get_params().items()],
                        width=120))
        logger.info("Markov chain entropy: {}".format(
            ChainDataset.entropy))
        logger.info("Expected min error: {}".format(
            -ChainDataset.entropy * seq_len * batch_size))

        if os.path.isfile(save_path) and not reset:
            model = Pylearn2Model.load(save_path)
        else:
            model = Pylearn2Model(generator)

        # Build the cost computation graph.
        # Note: would be probably nicer to make cost part of the model.
        x = tensor.ltensor3('x')
        cost = Pylearn2Cost(model.brick.cost(x[:, :, 0]).sum())

        dataset = ChainDataset(rng, seq_len)
        sgd = SGD(learning_rate=0.0001, cost=cost,
                  batch_size=batch_size, batches_per_iter=10,
                  monitoring_dataset=dataset,
                  monitoring_batch_size=batch_size,
                  monitoring_batches=1,
                  learning_rule=Pylearn2LearningRule(
                      SGDLearningRule(),
                      dict(training_objective=cost.cost)))
        train = Pylearn2Train(dataset, model, algorithm=sgd,
                              save_path=save_path, save_freq=10)
        train.main_loop(time_budget=time_budget)
    elif mode == "sample":
        model = Pylearn2Model.load(save_path)
        generator = model.brick

        sample = ComputationGraph(generator.generate(
            n_steps=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,
                                               ChainDataset.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, ChainDataset.trans_prob))
    else:
        assert False

# 需要导入模块: from blocks.bricks.recurrent import GatedRecurrent [as 别名]
# 或者: from blocks.bricks.recurrent.GatedRecurrent import weights_init [as 别名]
        Readout(readout_dim = vocab_size,
                source_names = ["states"], # transition.apply.states ???
                emitter = SoftmaxEmitter(name = "emitter"),
                feedback_brick = LookupFeedback(
                    vocab_size,
                    feedback_dim,
                    name = 'feedback'
                ),
                name = "readout"),
        transition,
        weights_init = IsotropicGaussian(0.01),
        biases_init  = Constant(0),
        name = "generator"
    )
    generator.push_initialization_config()
    transition.weights_init = Orthogonal()
    generator.initialize()

    # Build the cost computation graph.
    x = tensor.lmatrix('inchar')

    cost = generator.cost(outputs=x)
    cost.name = "sequence_cost"

    algorithm = GradientDescent(
        cost = cost,
        parameters = list(Selector(generator).get_parameters().values()),
        step_rule = Adam(),
        # because we want use all the stuff in the training data
        on_unused_sources = 'ignore'
    )