deepy.layers.NeuralLayer

Complexity

Total Complexity

54

Size/Duplication

Total Lines	265
Duplicated Lines	0 %

27 Methods

Rating	Name	Duplication	Size	Complexity
B	__init__()	0	28	1
A	callback_forward_propagation()	0	2	1
A	register_free_parameters()	0	5	1
A	register_monitors()	0	9	3
A	output()	0	5	1
A	create_matrix()	0	6	1
A	create_bias()	0	8	2
A	create_weight()	0	11	3
A	register_updates()	0	8	3
A	setup()	0	6	1
A	create_vector()	0	6	1
A	prepare()	0	5	1
A	register_epoch_callbacks()	0	5	1
D	compute()	0	30	8
A	register_external_inputs()	0	5	1
A	link()	0	13	3
A	call()	0	8	2
A	compute_raw()	0	9	2
A	test_output()	0	5	1
A	register_external_targets()	0	5	1
B	connect()	0	25	6
A	register_parameters()	0	7	2
A	register_training_callbacks()	0	5	1
A	register_training_updates()	0	8	3
A	register()	0	5	1
A	register_inner_layers()	0	3	2
A	register_testing_callbacks()	0	5	1

#!/usr/bin/env python
# -*- coding: utf-8 -*-


import logging as loggers

import numpy as np
import theano

from deepy.utils import FLOATX, global_rand, UniformInitializer

logging = loggers.getLogger(__name__)

class NeuralLayer(object):

    def __init__(self, name="unknown"):
        """
        Create a neural layer.
        """
        self.name = name
        self.input_dim = 0
        self.input_dims = [0]
        self.output_dim = 0

        self._linked_block = None
        self._linked = False

        self.connected = False
        self.updates = []
        self.training_updates = []
        self.free_parameters = []
        self.parameters = []
        self.training_monitors = []
        self.testing_monitors = []
        self._registered_monitors = set()
        self._registered_updates = set()
        self._registered_training_updates = set()
        self.external_inputs = []
        self.external_targets = []
        self.parameter_count = 0
        self.epoch_callbacks = []
        self.training_callbacks = []
        self.testing_callbacks = []

    def connect(self, input_dim=0, input_dims=None, previous_layer=None, network_config=None, no_prepare=False):
        """
        Connect to a previous layer.
        :param no_prepare: if avoid calling setup
        """
        # configure input dimensions
        if input_dims:
            self.input_dims = input_dims
            self.input_dim = input_dims[0]
        else:
            self.input_dim = input_dim
            self.input_dims = [input_dims]
        # set default output dimension
        if self.output_dim == 0:
            self.output_dim = self.input_dim
        self.previous_layer = previous_layer
        self.network_config = network_config
        self.connected = True
        # call prepare
        if not no_prepare:
            self.prepare()
        if self._linked_block and not self._linked:
            self._linked = True
            self._linked_block.register_layer(self)
        return self

    def compute_raw(self, inputs, dims, **kwargs):
        """
        Compute on raw Theano tensors.
        :type inputs: list of TensorLayer
        :type dims: list of int
        """
        from var import NeuralVar
        tensors = [NeuralVar(d, t) for d, t in zip(dims, inputs)]
        return self.compute(*tensors, **kwargs)

    def compute(self, *inputs, **kwargs):
        """
        Take a TensorLayer or tensor as input, compute the result.
        Dimension must be given is the input is a tensor.
        :type inputs:  list of TensorLayer
        :return: TensorLayer
        """
        from var import NeuralVar
        if type(inputs[0]) != NeuralVar:
            raise SystemError("The input of `compute` must be NeuralVar")

        dims = [t.dim() for t in inputs]
        if len(inputs) == 1:
            self.connect(input_dim=dims[0])
        else:
            self.connect(input_dims=dims)
        # convert kwargs
        train_kwargs = {}
        test_kwargs = {}
        for key, val in kwargs.items():
            if type(val) == NeuralVar:
                train_kwargs[key] = val.tensor
                test_kwargs[key] = val.test_tensor
            else:
                train_kwargs[key] = val
                test_kwargs[key] = val

        return NeuralVar(self.output_dim,
                         self.output(*[t.tensor for t in inputs], **train_kwargs),
                         self.test_output(*[t.test_tensor for t in inputs], **test_kwargs))

    def prepare(self):
        """
        Prepare function will be called after connected.
        """
        return self.setup()

    def setup(self):
        """
        !!! DEPRECATED !!!
        Setup function will be called after connected.
        """
        pass

    def output(self, *args, **kwargs):
        """
        Output function.
        """
        raise NotImplementedError("output function of '%s' is not implemented" % self.name)

    def test_output(self, *args, **kwargs):
        """
        Output function in test time.
        """
        return self.output(*args, **kwargs)

    def call(self, x, test=False):
        """
        Call this layer, with a parameter to switch test or not.
        """
        if test:
            return self.test_output(x)
        else:
            return self.output(x)

    def link(self, block):
        """
        Let the given block or network manage the parameters of this layer.
        :param block: Block or NeuralNetwork
        :return: NeuralLayer
        """
        if self._linked_block:
            raise SystemError("One layer can not be linked twice")
        self._linked_block = block
        if self.connected:
            self._linked = True
            block.register_layer(self)
        return self

    def register(self, *layers):
        """
        Register inner layers.
        """
        self.register_inner_layers(*layers)

    def register_inner_layers(self, *layers):
        for layer in layers:
            self.register_parameters(*layer.parameters)

    def register_parameters(self, *parameters):
        """
        Register parameters.
        """
        for param in parameters:
            self.parameter_count += np.prod(param.get_value().shape)
        self.parameters.extend(parameters)

    def register_free_parameters(self, *free_parameters):
        """
        Register free parameters, which means their value will not be learned by trainer.
        """
        return self.free_parameters.extend(free_parameters)

    def register_updates(self, *updates):
        """
        Register updates that will be executed in each iteration.
        """
        for key, node in updates:
            if key not in self._registered_updates:
                self.updates.append((key, node))
                self._registered_updates.add(key)

    def register_training_updates(self, *updates):
        """
        Register updates that will only be executed in training phase.
        """
        for key, node in updates:
            if key not in self._registered_training_updates:
                self.training_updates.append((key, node))
                self._registered_training_updates.add(key)

    def register_monitors(self, *monitors):
        """
        Register monitors they should be tuple of name and Theano variable.
        """
        for key, node in monitors:
            if key not in self._registered_monitors:
                self.training_monitors.append((key, node))
                self.testing_monitors.append((key, node))
                self._registered_monitors.add(key)

    def register_external_inputs(self, *variables):
        """
        Register external input variables.
        """
        self.external_inputs.extend(variables)

    def register_external_targets(self, *variables):
        """
        Register extenal target variables.
        """
        self.external_targets.extend(variables)

    def register_training_callbacks(self, *callbacks):
        """
        Register callback for each iteration in the training.
        """
        self.training_callbacks.extend(callbacks)

    def register_testing_callbacks(self, *callbacks):
        """
        Register callback for each iteration in the testing.
        """
        self.testing_callbacks.extend(callbacks)

    def register_epoch_callbacks(self, *callbacks):
        """
        Register callback which will be called after epoch finished.
        """
        self.epoch_callbacks.extend(callbacks)

    def create_weight(self, input_n=1, output_n=1, suffix="", initializer=None, shape=None):
        if not shape:
            shape = (input_n, output_n)

        if not initializer:
            initializer = UniformInitializer()

        weight = theano.shared(initializer.sample(shape).astype(FLOATX), name='W_{}'.format(suffix))

        logging.info('create weight W_%s: %s', suffix, str(shape))
        return weight

    def create_bias(self, output_n=1, suffix="", value=0., shape=None):
        if not shape:
            shape = (output_n, )
        bs =  np.ones(shape)
        bs *= value
        bias = theano.shared(bs.astype(FLOATX), name='B_{}'.format(suffix))
        logging.info('create bias B_%s: %s', suffix, str(shape))
        return bias

    def create_vector(self, n, name, dtype=FLOATX):
        bs =  np.zeros(n)
        v = theano.shared(bs.astype(dtype), name='{}'.format(name))

        logging.info('create vector %s: %d', name, n)
        return v

    def create_matrix(self, m, n, name):

        matrix = theano.shared(np.zeros((m, n)).astype(FLOATX), name=name)

        logging.info('create matrix %s: %d x %d', name, m, n)
        return matrix

    def callback_forward_propagation(self):
        pass