BatchNormalization.normalization_axes() - Code Metrics - Inspection of "Refactor BatchNormalization to make it more extens..." - mila-udem/blocks - Measure and Improve Code Quality continuously with Scrutinizer

Completed

Pull Request — master (#1120)

by David

created 2016-06-17 00:27 UTC

BatchNormalization.normalization_axes() A

↳ Parent: BatchNormalization

Complexity

Conditions

Size

Total Lines

Duplication

Lines	0
Ratio	0 %

Importance

Changes	1
Bugs	0	Features	0

Metric	Value
cc	3
c	1
b	0
f	0
dl	0
loc	5
rs	9.4285

import collections
from functools import partial

import numpy
from picklable_itertools.extras import equizip
import theano
from theano import tensor
from theano.tensor.nnet import bn

from ..graph import add_annotation
from ..initialization import Constant
from ..roles import (BATCH_NORM_POPULATION_MEAN,
                     BATCH_NORM_POPULATION_STDEV, BATCH_NORM_OFFSET,
                     BATCH_NORM_DIVISOR, BATCH_NORM_MINIBATCH_ESTIMATE,
                     BATCH_NORM_SHIFT_PARAMETER, BATCH_NORM_SCALE_PARAMETER,
                     add_role)
from ..utils import shared_floatx, shared_floatx_nans, is_shared_variable
from .base import lazy, application
from .sequences import Sequence, Feedforward, MLP
from .interfaces import RNGMixin


def _add_batch_axis(var):
    """Prepend a singleton axis to a TensorVariable and name it."""
    new_var = new_var = tensor.shape_padleft(var)
    new_var.name = 'shape_padleft({})'.format(var.name)
    return new_var


def _add_role_and_annotate(var, role, annotations=()):
    """Add a role and zero or more annotations to a variable."""
    add_role(var, role)
    for annotation in annotations:
        add_annotation(var, annotation)


class BatchNormalization(RNGMixin, Feedforward):
    r"""Normalizes activations, parameterizes a scale and shift.

    Parameters
    ----------
    input_dim : int or tuple
        Shape of a single input example. It is assumed that a batch axis
        will be prepended to this.
    broadcastable : tuple, optional
        Tuple of the same length as `input_dim` which specifies which of
        the per-example axes should be averaged over to compute means and
        standard deviations. For example, in order to normalize over all
        spatial locations in a `(batch_index, channels, height, width)`
        image, pass `(False, True, True)`. The batch axis is always
        averaged out.
    conserve_memory : bool, optional
        Use an implementation that stores less intermediate state and
        therefore uses less memory, at the expense of 5-10% speed. Default
        is `True`.
    epsilon : float, optional
       The stabilizing constant for the minibatch standard deviation
       computation (when the brick is run in training mode).
       Added to the variance inside the square root, as in the
       batch normalization paper.
    scale_init : object, optional
        Initialization object to use for the learned scaling parameter
        ($\\gamma$ in [BN]_). By default, uses constant initialization
        of 1.
    shift_init : object, optional
        Initialization object to use for the learned shift parameter
        ($\\beta$ in [BN]_). By default, uses constant initialization of 0.
    mean_only : bool, optional
        Perform "mean-only" batch normalization as described in [SK2016]_.
    learn_scale : bool, optional
        Whether to include a learned scale parameter ($\\gamma$ in [BN]_)
        in this brick. Default is `True`. Has no effect if `mean_only` is
        `True` (i.e. a scale parameter is never learned in mean-only mode).
    learn_shift : bool, optional
        Whether to include a learned shift parameter ($\\beta$ in [BN]_)
        in this brick. Default is `True`.

    Notes
    -----
    In order for trained models to behave sensibly immediately upon
    upon deserialization, by default, this brick runs in *inference* mode,
    using a population mean and population standard deviation (initialized
    to zeros and ones respectively) to normalize activations. It is
    expected that the user will adapt these during training in some
    fashion, independently of the training objective, e.g. by taking a
    moving average of minibatch-wise statistics.

    In order to *train* with batch normalization, one must obtain a
    training graph by transforming the original inference graph. See
    :func:`~blocks.graph.apply_batch_normalization` for a routine to
    transform graphs, and :func:`~blocks.graph.batch_normalization`
    for a context manager that may enable shorter compile times
    (every instance of :class:`BatchNormalization` is itself a context
    manager, entry into which causes applications to be in minibatch
    "training" mode, however it is usually more convenient to use
    :func:`~blocks.graph.batch_normalization` to enable this behaviour
    for all of your graph's :class:`BatchNormalization` bricks at once).

    Note that training in inference mode should be avoided, as this
    brick introduces scales and shift parameters (tagged with the
    `PARAMETER` role) that, in the absence of batch normalization,
    usually makes things unstable. If you must do this, filter for and
    remove `BATCH_NORM_SHIFT_PARAMETER` and `BATCH_NORM_SCALE_PARAMETER`
    from the list of parameters you are training, and this brick should
    behave as a (somewhat expensive) no-op.

    This Brick accepts `scale_init` and `shift_init` arguments but is
    *not* an instance of :class:`~blocks.bricks.Initializable`, and will
    therefore not receive pushed initialization config from any parent
    brick. In almost all cases, you will probably want to stick with the
    defaults (unit scale and zero offset), but you can explicitly pass one
    or both initializers to override this.

    This has the necessary properties to be inserted into a
    :class:`blocks.bricks.conv.ConvolutionalSequence` as-is, in which case
    the `input_dim` should be omitted at construction, to be inferred from
    the layer below.


    .. [BN] Sergey Ioffe and Christian Szegedy. *Batch normalization:
       accelerating deep network training by reducing internal covariate
       shift*. ICML (2015), pp. 448-456.

    .. [SK2016] Tim Salimans and Diederik P. Kingma. *Weight
       normalization: a simple reparameterization to accelerate training
       of deep neural networks*. arXiv 1602.07868.

    """
    @lazy(allocation=['input_dim'])
    def __init__(self, input_dim, broadcastable=None,
                 conserve_memory=True, epsilon=1e-4, scale_init=None,
                 shift_init=None, mean_only=False, learn_scale=True,
                 learn_shift=True, **kwargs):
        self.input_dim = input_dim
        self.broadcastable = broadcastable
        self.conserve_memory = conserve_memory
        self.epsilon = epsilon
        self.learn_scale = learn_scale
        self.learn_shift = learn_shift
        self.scale_init = (Constant(1) if scale_init is None
                           else scale_init)
        self.shift_init = (Constant(0) if shift_init is None
                           else shift_init)
        self.mean_only = mean_only
        self._training_mode = []
        super(BatchNormalization, self).__init__(**kwargs)

    @application(inputs=['input_'], outputs=['output'])
    def apply(self, input_, application_call):
        if self._training_mode:
            mean, stdev = self._compute_training_statistics(input_)
        else:
            mean, stdev = self._prepare_population_statistics()
        # Useful for filtration of calls that were already made in
        # training mode when doing graph transformations.
        # Very important to cast to bool, as self._training_mode is
        # normally a list (to support nested context managers), which would
        # otherwise get passed by reference and be remotely mutated.
        application_call.metadata['training_mode'] = bool(self._training_mode)
        # Useful for retrieving a list of updates for population
        # statistics. Ditch the broadcastable first axis, though, to
        # make it the same dimensions as the population mean and stdev
        # shared variables.
        application_call.metadata['offset'] = mean[0]
        application_call.metadata['divisor'] = stdev[0]
        # Give these quantities roles in the graph.
        _add_role_and_annotate(mean, BATCH_NORM_OFFSET,
                               [self, application_call])
        if self.mean_only:
            stdev = tensor.ones_like(mean)
        else:
            # The annotation/role information is useless if it's a constant.
            _add_role_and_annotate(stdev, BATCH_NORM_DIVISOR,
                                   [self, application_call])
        shift = _add_batch_axis(self.shift)
        scale = _add_batch_axis(self.scale)
        # Heavy lifting is done by the Theano utility function.
        normalized = bn.batch_normalization(input_, scale, shift, mean, stdev,
                                            mode=('low_mem'
                                                  if self.conserve_memory
                                                  else 'high_mem'))
        return normalized

    def __enter__(self):
        self._training_mode.append(True)

    def __exit__(self, *exc_info):
        self._training_mode.pop()

    def _compute_training_statistics(self, input_):
        axes = self.normalization_axes
        mean = input_.mean(axis=axes, keepdims=True)
        add_role(mean, BATCH_NORM_MINIBATCH_ESTIMATE)
        if self.mean_only:
            stdev = tensor.ones_like(mean)
        else:
            var = (tensor.sqr(input_).mean(axis=axes, keepdims=True) -
                   tensor.sqr(mean))
            eps = numpy.cast[theano.config.floatX](self.epsilon)
            stdev = tensor.sqrt(var + eps)
            add_role(stdev, BATCH_NORM_MINIBATCH_ESTIMATE)
        return mean, stdev

    @property
    def normalization_axes(self):
        return (0,) + tuple((i + 1) for i, b in
                            enumerate(self.population_mean.broadcastable)
                            if b)

    def _prepare_population_statistics(self):
        mean = _add_batch_axis(self.population_mean)
        if self.mean_only:
            stdev = tensor.ones_like(self.population_mean)
        else:
            stdev = self.population_stdev
        stdev = _add_batch_axis(stdev)
        return mean, stdev

    def _allocate(self):
        input_dim = ((self.input_dim,)
                     if not isinstance(self.input_dim, collections.Sequence)
                     else self.input_dim)
        broadcastable = (tuple(False for _ in input_dim)
                         if self.broadcastable is None else self.broadcastable)
        if len(input_dim) != len(broadcastable):
            raise ValueError("input_dim and broadcastable must be same length")
        var_dim = tuple(1 if broadcast else dim for dim, broadcast in
                        equizip(input_dim, broadcastable))
        broadcastable = broadcastable

        # "beta", from the Ioffe & Szegedy manuscript.
        if self.learn_shift:
            self.shift = shared_floatx_nans(var_dim, name='batch_norm_shift',
                                            broadcastable=broadcastable)
            add_role(self.shift, BATCH_NORM_SHIFT_PARAMETER)
            self.parameters.append(self.shift)
        else:
            self.shift = tensor.constant(0, dtype=theano.config.floatX)

        if self.learn_scale and not self.mean_only:
            # "gamma", from the Ioffe & Szegedy manuscript.
            self.scale = shared_floatx_nans(var_dim, name='batch_norm_scale',
                                            broadcastable=broadcastable)

            add_role(self.scale, BATCH_NORM_SCALE_PARAMETER)
            self.parameters.append(self.scale)
        else:
            self.scale = tensor.constant(1., dtype=theano.config.floatX)

        self._allocate_population_statistics(var_dim, broadcastable)

    def _allocate_population_statistics(self, var_dim, broadcastable):
        def _allocate_buffer(name, role, value):
            # These aren't technically parameters, in that they should not be
            # learned using the same cost function as other model parameters.
            population_buffer = shared_floatx(value * numpy.ones(var_dim),
                                              broadcastable=broadcastable,
                                              name=name)
            add_role(population_buffer, role)
            # Normally these would get annotated by an AnnotatingList, but they
            # aren't in self.parameters.
            add_annotation(population_buffer, self)
            return population_buffer

        self.population_mean = _allocate_buffer('population_mean',
                                                BATCH_NORM_POPULATION_MEAN,
                                                numpy.zeros(var_dim))

        self.population_stdev = _allocate_buffer('population_stdev',
                                                 BATCH_NORM_POPULATION_STDEV,
                                                 numpy.ones(var_dim))

    def _initialize(self):
        # We gate with is_shared_variable rather than relying on
        # learn_scale and learn_shift so as to avoid the unlikely but nasty
        # scenario where those flags are changed post-allocation but
        # pre-initialization. This ensures that such a change simply has no
        # effect rather than doing an inconsistent combination of things.
        if is_shared_variable(self.shift):
            self.shift_init.initialize(self.shift, self.rng)
        if is_shared_variable(self.scale):
            self.scale_init.initialize(self.scale, self.rng)

    # Needed for the Feedforward interface.
    @property
    def output_dim(self):
        return self.input_dim

    # The following properties allow for BatchNormalization bricks
    # to be used directly inside of a ConvolutionalSequence.
    @property
    def image_size(self):
        return self.input_dim[-2:]

    @image_size.setter
    def image_size(self, value):
        if not isinstance(self.input_dim, collections.Sequence):
            self.input_dim = (None,) + tuple(value)
        else:
            self.input_dim = (self.input_dim[0],) + tuple(value)

    @property
    def num_channels(self):
        return self.input_dim[0]

    @num_channels.setter
    def num_channels(self, value):
        if not isinstance(self.input_dim, collections.Sequence):
            self.input_dim = (value,) + (None, None)
        else:
            self.input_dim = (value,) + self.input_dim[-2:]

    def get_dim(self, name):
        if name in ('input', 'output'):
            return self.input_dim
        else:
            raise KeyError

    @property
    def num_output_channels(self):
        return self.num_channels


class SpatialBatchNormalization(BatchNormalization):
    """Convenient subclass for batch normalization across spatial inputs.

    Parameters
    ----------
    input_dim : int or tuple
        The input size of a single example. Must be length at least 2.
        It's assumed that the first axis of this tuple is a "channels"
        axis, which should not be summed over, and all remaining
        dimensions are spatial dimensions.

    Notes
    -----
    See :class:`BatchNormalization` for more details (and additional
    keyword arguments).

    """
    def _allocate(self):
        if not isinstance(self.input_dim,
                          collections.Sequence) or len(self.input_dim) < 2:
            raise ValueError('expected input_dim to be length >= 2 '
                             'e.g. (channels, height, width)')
        self.broadcastable = (False,) + ((True,) * (len(self.input_dim) - 1))
        super(SpatialBatchNormalization, self)._allocate()


class BatchNormalizedMLP(MLP):
    """Convenient subclass for building an MLP with batch normalization.

    Parameters
    ----------
    conserve_memory : bool, optional, by keyword only
        See :class:`BatchNormalization`.
    mean_only : bool, optional, by keyword only
        See :class:`BatchNormalization`.
    learn_scale : bool, optional, by keyword only
        See :class:`BatchNormalization`.
    learn_shift : bool, optional, by keyword only
        See :class:`BatchNormalization`.

    Notes
    -----
    All other parameters are the same as :class:`~blocks.bricks.MLP`. Each
    activation brick is wrapped in a :class:`~blocks.bricks.Sequence`
    containing an appropriate :class:`BatchNormalization` brick and
    the activation that follows it.

    By default, the contained :class:`~blocks.bricks.Linear` bricks will
    not contain any biases, as they could be canceled out by the biases
    in the :class:`BatchNormalization` bricks being added. Pass
    `use_bias` with a value of `True` if you really want this for some
    reason.

    `mean_only`, `learn_scale` and `learn_shift` are pushed down to
    all created :class:`BatchNormalization` bricks as allocation
    config.

    """
    @lazy(allocation=['dims'])
    def __init__(self, activations, dims, *args, **kwargs):
        self._conserve_memory = kwargs.pop('conserve_memory', True)
        self.mean_only = kwargs.pop('mean_only', False)
        self.learn_scale = kwargs.pop('learn_scale', True)
        self.learn_shift = kwargs.pop('learn_shift', True)

        activations = [
            Sequence([
                (BatchNormalization(conserve_memory=self._conserve_memory)
                 .apply),
                act.apply
            ], name='batch_norm_activation_{}'.format(i))
            for i, act in enumerate(activations)
        ]
        # Batch normalization bricks incorporate a bias, so there's no
        # need for our Linear bricks to have them.
        kwargs.setdefault('use_bias', False)
        super(BatchNormalizedMLP, self).__init__(activations, dims, *args,
                                                 **kwargs)

    def _nested_brick_property_getter(self, property_name):
        return getattr(self, '_' + property_name)

    def _nested_brick_property_setter(self, value, property_name):
        setattr(self, '_' + property_name, value)
        for act in self.activations:
            assert isinstance(act.children[0], BatchNormalization)
            setattr(act.children[0], property_name, value)

    # conserve_memory is a bit special in that it can be modified
    # after construction/allocation and still have a valid effect on
    # apply(). Thus we propagate down all property sets.
    conserve_memory = property(partial(_nested_brick_property_getter,
                                       property_name='conserve_memory'),
                               partial(_nested_brick_property_setter,
                                       property_name='conserve_memory'),
                               doc="Conserve memory.")

    def _push_allocation_config(self):
        super(BatchNormalizedMLP, self)._push_allocation_config()
        # Do the extra allocation pushing for the BatchNormalization
        # bricks. They need as their input dimension the output dimension
        # of each linear transformation.  Exclude the first dimension,
        # which is the input dimension.
        for act, dim in equizip(self.activations, self.dims[1:]):
            assert isinstance(act.children[0], BatchNormalization)
            act.children[0].input_dim = dim
            for attr in ['mean_only', 'learn_scale', 'learn_shift']:
                setattr(act.children[0], attr, getattr(self, attr))


1			import collections
2			from functools import partial
3
4			import numpy
5			from picklable_itertools.extras import equizip
6			import theano
7			from theano import tensor
8			from theano.tensor.nnet import bn
9
10			from ..graph import add_annotation
11			from ..initialization import Constant
12			from ..roles import (BATCH_NORM_POPULATION_MEAN,
13			BATCH_NORM_POPULATION_STDEV, BATCH_NORM_OFFSET,
14			BATCH_NORM_DIVISOR, BATCH_NORM_MINIBATCH_ESTIMATE,
15			BATCH_NORM_SHIFT_PARAMETER, BATCH_NORM_SCALE_PARAMETER,
16			add_role)
17			from ..utils import shared_floatx, shared_floatx_nans, is_shared_variable
18			from .base import lazy, application
19			from .sequences import Sequence, Feedforward, MLP
20			from .interfaces import RNGMixin
21
22
23			def _add_batch_axis(var):
24			"""Prepend a singleton axis to a TensorVariable and name it."""
25			new_var = new_var = tensor.shape_padleft(var)
26			new_var.name = 'shape_padleft({})'.format(var.name)
27			return new_var
28
29
30			def _add_role_and_annotate(var, role, annotations=()):
31			"""Add a role and zero or more annotations to a variable."""
32			add_role(var, role)
33			for annotation in annotations:
34			add_annotation(var, annotation)
35
36
37			class BatchNormalization(RNGMixin, Feedforward):
38			r"""Normalizes activations, parameterizes a scale and shift.
39
40			Parameters
41			----------
42			input_dim : int or tuple
43			Shape of a single input example. It is assumed that a batch axis
44			will be prepended to this.
45			broadcastable : tuple, optional
46			Tuple of the same length as `input_dim` which specifies which of
47			the per-example axes should be averaged over to compute means and
48			standard deviations. For example, in order to normalize over all
49			spatial locations in a `(batch_index, channels, height, width)`
50			image, pass `(False, True, True)`. The batch axis is always
51			averaged out.
52			conserve_memory : bool, optional
53			Use an implementation that stores less intermediate state and
54			therefore uses less memory, at the expense of 5-10% speed. Default
55			is `True`.
56			epsilon : float, optional
57			The stabilizing constant for the minibatch standard deviation
58			computation (when the brick is run in training mode).
59			Added to the variance inside the square root, as in the
60			batch normalization paper.
61			scale_init : object, optional
62			Initialization object to use for the learned scaling parameter
63			($\\gamma$ in [BN]_). By default, uses constant initialization
64			of 1.
65			shift_init : object, optional
66			Initialization object to use for the learned shift parameter
67			($\\beta$ in [BN]_). By default, uses constant initialization of 0.
68			mean_only : bool, optional
69			Perform "mean-only" batch normalization as described in [SK2016]_.
70			learn_scale : bool, optional
71			Whether to include a learned scale parameter ($\\gamma$ in [BN]_)
72			in this brick. Default is `True`. Has no effect if `mean_only` is
73			`True` (i.e. a scale parameter is never learned in mean-only mode).
74			learn_shift : bool, optional
75			Whether to include a learned shift parameter ($\\beta$ in [BN]_)
76			in this brick. Default is `True`.
77
78			Notes
79			-----
80			In order for trained models to behave sensibly immediately upon
81			upon deserialization, by default, this brick runs in inference mode,
82			using a population mean and population standard deviation (initialized
83			to zeros and ones respectively) to normalize activations. It is
84			expected that the user will adapt these during training in some
85			fashion, independently of the training objective, e.g. by taking a
86			moving average of minibatch-wise statistics.
87
88			In order to train with batch normalization, one must obtain a
89			training graph by transforming the original inference graph. See
90			:func:`~blocks.graph.apply_batch_normalization` for a routine to
91			transform graphs, and :func:`~blocks.graph.batch_normalization`
92			for a context manager that may enable shorter compile times
93			(every instance of :class:`BatchNormalization` is itself a context
94			manager, entry into which causes applications to be in minibatch
95			"training" mode, however it is usually more convenient to use
96			:func:`~blocks.graph.batch_normalization` to enable this behaviour
97			for all of your graph's :class:`BatchNormalization` bricks at once).
98
99			Note that training in inference mode should be avoided, as this
100			brick introduces scales and shift parameters (tagged with the
101			`PARAMETER` role) that, in the absence of batch normalization,
102			usually makes things unstable. If you must do this, filter for and
103			remove `BATCH_NORM_SHIFT_PARAMETER` and `BATCH_NORM_SCALE_PARAMETER`
104			from the list of parameters you are training, and this brick should
105			behave as a (somewhat expensive) no-op.
106
107			This Brick accepts `scale_init` and `shift_init` arguments but is
108			not an instance of :class:`~blocks.bricks.Initializable`, and will
109			therefore not receive pushed initialization config from any parent
110			brick. In almost all cases, you will probably want to stick with the
111			defaults (unit scale and zero offset), but you can explicitly pass one
112			or both initializers to override this.
113
114			This has the necessary properties to be inserted into a
115			:class:`blocks.bricks.conv.ConvolutionalSequence` as-is, in which case
116			the `input_dim` should be omitted at construction, to be inferred from
117			the layer below.
118
119
120			.. [BN] Sergey Ioffe and Christian Szegedy. *Batch normalization:
121			accelerating deep network training by reducing internal covariate
122			shift*. ICML (2015), pp. 448-456.
123
124			.. [SK2016] Tim Salimans and Diederik P. Kingma. *Weight
125			normalization: a simple reparameterization to accelerate training
126			of deep neural networks*. arXiv 1602.07868.
127
128			"""
129			@lazy(allocation=['input_dim'])
130			def __init__(self, input_dim, broadcastable=None,
131			conserve_memory=True, epsilon=1e-4, scale_init=None,
132			shift_init=None, mean_only=False, learn_scale=True,
133			learn_shift=True, **kwargs):
134			self.input_dim = input_dim
135			self.broadcastable = broadcastable
136			self.conserve_memory = conserve_memory
137			self.epsilon = epsilon
138			self.learn_scale = learn_scale
139			self.learn_shift = learn_shift
140			self.scale_init = (Constant(1) if scale_init is None
141			else scale_init)
142			self.shift_init = (Constant(0) if shift_init is None
143			else shift_init)
144			self.mean_only = mean_only
145			self._training_mode = []
146			super(BatchNormalization, self).__init__(**kwargs)
147
148			@application(inputs=['input_'], outputs=['output'])
149			def apply(self, input_, application_call):
150			if self._training_mode:
151			mean, stdev = self._compute_training_statistics(input_)
152			else:
153			mean, stdev = self._prepare_population_statistics()
154			# Useful for filtration of calls that were already made in
155			# training mode when doing graph transformations.
156			# Very important to cast to bool, as self._training_mode is
157			# normally a list (to support nested context managers), which would
158			# otherwise get passed by reference and be remotely mutated.
159			application_call.metadata['training_mode'] = bool(self._training_mode)
160			# Useful for retrieving a list of updates for population
161			# statistics. Ditch the broadcastable first axis, though, to
162			# make it the same dimensions as the population mean and stdev
163			# shared variables.
164			application_call.metadata['offset'] = mean[0]
165			application_call.metadata['divisor'] = stdev[0]
166			# Give these quantities roles in the graph.
167			_add_role_and_annotate(mean, BATCH_NORM_OFFSET,
168			[self, application_call])
169			if self.mean_only:
170			stdev = tensor.ones_like(mean)
171			else:
172			# The annotation/role information is useless if it's a constant.
173			_add_role_and_annotate(stdev, BATCH_NORM_DIVISOR,
174			[self, application_call])
175			shift = _add_batch_axis(self.shift)
176			scale = _add_batch_axis(self.scale)
177			# Heavy lifting is done by the Theano utility function.
178			normalized = bn.batch_normalization(input_, scale, shift, mean, stdev,
179			mode=('low_mem'
180			if self.conserve_memory
181			else 'high_mem'))
182			return normalized
183
184			def __enter__(self):
185			self._training_mode.append(True)
186
187			def __exit__(self, *exc_info):
188			self._training_mode.pop()
189
190			def _compute_training_statistics(self, input_):
191			axes = self.normalization_axes
192			mean = input_.mean(axis=axes, keepdims=True)
193			add_role(mean, BATCH_NORM_MINIBATCH_ESTIMATE)
194			if self.mean_only:
195			stdev = tensor.ones_like(mean)
196			else:
197			var = (tensor.sqr(input_).mean(axis=axes, keepdims=True) -
198			tensor.sqr(mean))
199			eps = numpy.cast[theano.config.floatX](self.epsilon)
200			stdev = tensor.sqrt(var + eps)
201			add_role(stdev, BATCH_NORM_MINIBATCH_ESTIMATE)
202			return mean, stdev
203
204			@property
205			def normalization_axes(self):
206			return (0,) + tuple((i + 1) for i, b in
207			enumerate(self.population_mean.broadcastable)
208			if b)
209
210			def _prepare_population_statistics(self):
211			mean = _add_batch_axis(self.population_mean)
212			if self.mean_only:
213			stdev = tensor.ones_like(self.population_mean)
214			else:
215			stdev = self.population_stdev
216			stdev = _add_batch_axis(stdev)
217			return mean, stdev
218
219			def _allocate(self):
220			input_dim = ((self.input_dim,)
221			if not isinstance(self.input_dim, collections.Sequence)
222			else self.input_dim)
223			broadcastable = (tuple(False for _ in input_dim)
224			if self.broadcastable is None else self.broadcastable)
225			if len(input_dim) != len(broadcastable):
226			raise ValueError("input_dim and broadcastable must be same length")
227			var_dim = tuple(1 if broadcast else dim for dim, broadcast in
228			equizip(input_dim, broadcastable))
229			broadcastable = broadcastable
230
231			# "beta", from the Ioffe & Szegedy manuscript.
232			if self.learn_shift:
233			self.shift = shared_floatx_nans(var_dim, name='batch_norm_shift',
234			broadcastable=broadcastable)
235			add_role(self.shift, BATCH_NORM_SHIFT_PARAMETER)
236			self.parameters.append(self.shift)
237			else:
238			self.shift = tensor.constant(0, dtype=theano.config.floatX)
239
240			if self.learn_scale and not self.mean_only:
241			# "gamma", from the Ioffe & Szegedy manuscript.
242			self.scale = shared_floatx_nans(var_dim, name='batch_norm_scale',
243			broadcastable=broadcastable)
244
245			add_role(self.scale, BATCH_NORM_SCALE_PARAMETER)
246			self.parameters.append(self.scale)
247			else:
248			self.scale = tensor.constant(1., dtype=theano.config.floatX)
249
250			self._allocate_population_statistics(var_dim, broadcastable)
251
252			def _allocate_population_statistics(self, var_dim, broadcastable):
253			def _allocate_buffer(name, role, value):
254			# These aren't technically parameters, in that they should not be
255			# learned using the same cost function as other model parameters.
256			population_buffer = shared_floatx(value * numpy.ones(var_dim),
257			broadcastable=broadcastable,
258			name=name)
259			add_role(population_buffer, role)
260			# Normally these would get annotated by an AnnotatingList, but they
261			# aren't in self.parameters.
262			add_annotation(population_buffer, self)
263			return population_buffer
264
265			self.population_mean = _allocate_buffer('population_mean',
266			BATCH_NORM_POPULATION_MEAN,
267			numpy.zeros(var_dim))
268
269			self.population_stdev = _allocate_buffer('population_stdev',
270			BATCH_NORM_POPULATION_STDEV,
271			numpy.ones(var_dim))
272
273			def _initialize(self):
274			# We gate with is_shared_variable rather than relying on
275			# learn_scale and learn_shift so as to avoid the unlikely but nasty
276			# scenario where those flags are changed post-allocation but
277			# pre-initialization. This ensures that such a change simply has no
278			# effect rather than doing an inconsistent combination of things.
279			if is_shared_variable(self.shift):
280			self.shift_init.initialize(self.shift, self.rng)
281			if is_shared_variable(self.scale):
282			self.scale_init.initialize(self.scale, self.rng)
283
284			# Needed for the Feedforward interface.
285			@property
286			def output_dim(self):
287			return self.input_dim
288
289			# The following properties allow for BatchNormalization bricks
290			# to be used directly inside of a ConvolutionalSequence.
291			@property
292			def image_size(self):
293			return self.input_dim[-2:]
294
295			@image_size.setter
296			def image_size(self, value):
297			if not isinstance(self.input_dim, collections.Sequence):
298			self.input_dim = (None,) + tuple(value)
299			else:
300			self.input_dim = (self.input_dim[0],) + tuple(value)
301
302			@property
303			def num_channels(self):
304			return self.input_dim[0]
305
306			@num_channels.setter
307			def num_channels(self, value):
308			if not isinstance(self.input_dim, collections.Sequence):
309			self.input_dim = (value,) + (None, None)
310			else:
311			self.input_dim = (value,) + self.input_dim[-2:]
312
313			def get_dim(self, name):
314			if name in ('input', 'output'):
315			return self.input_dim
316			else:
317			raise KeyError
318
319			@property
320			def num_output_channels(self):
321			return self.num_channels
322
323
324			class SpatialBatchNormalization(BatchNormalization):
325			"""Convenient subclass for batch normalization across spatial inputs.
326
327			Parameters
328			----------
329			input_dim : int or tuple
330			The input size of a single example. Must be length at least 2.
331			It's assumed that the first axis of this tuple is a "channels"
332			axis, which should not be summed over, and all remaining
333			dimensions are spatial dimensions.
334
335			Notes
336			-----
337			See :class:`BatchNormalization` for more details (and additional
338			keyword arguments).
339
340			"""
341			def _allocate(self):
342			if not isinstance(self.input_dim,
343			collections.Sequence) or len(self.input_dim) < 2:
344			raise ValueError('expected input_dim to be length >= 2 '
345			'e.g. (channels, height, width)')
346			self.broadcastable = (False,) + ((True,) * (len(self.input_dim) - 1))
347			super(SpatialBatchNormalization, self)._allocate()
348
349
350			class BatchNormalizedMLP(MLP):
351			"""Convenient subclass for building an MLP with batch normalization.
352
353			Parameters
354			----------
355			conserve_memory : bool, optional, by keyword only
356			See :class:`BatchNormalization`.
357			mean_only : bool, optional, by keyword only
358			See :class:`BatchNormalization`.
359			learn_scale : bool, optional, by keyword only
360			See :class:`BatchNormalization`.
361			learn_shift : bool, optional, by keyword only
362			See :class:`BatchNormalization`.
363
364			Notes
365			-----
366			All other parameters are the same as :class:`~blocks.bricks.MLP`. Each
367			activation brick is wrapped in a :class:`~blocks.bricks.Sequence`
368			containing an appropriate :class:`BatchNormalization` brick and
369			the activation that follows it.
370
371			By default, the contained :class:`~blocks.bricks.Linear` bricks will
372			not contain any biases, as they could be canceled out by the biases
373			in the :class:`BatchNormalization` bricks being added. Pass
374			`use_bias` with a value of `True` if you really want this for some
375			reason.
376
377			`mean_only`, `learn_scale` and `learn_shift` are pushed down to
378			all created :class:`BatchNormalization` bricks as allocation
379			config.
380
381			"""
382			@lazy(allocation=['dims'])
383			def __init__(self, activations, dims, args, *kwargs):
384			self._conserve_memory = kwargs.pop('conserve_memory', True)
385			self.mean_only = kwargs.pop('mean_only', False)
386			self.learn_scale = kwargs.pop('learn_scale', True)
387			self.learn_shift = kwargs.pop('learn_shift', True)
388
389			activations = [
390			Sequence([
391			(BatchNormalization(conserve_memory=self._conserve_memory)
392			.apply),
393			act.apply
394			], name='batch_norm_activation_{}'.format(i))
395			for i, act in enumerate(activations)
396			]
397			# Batch normalization bricks incorporate a bias, so there's no
398			# need for our Linear bricks to have them.
399			kwargs.setdefault('use_bias', False)
400			super(BatchNormalizedMLP, self).__init__(activations, dims, *args,
401			**kwargs)
402
403			def _nested_brick_property_getter(self, property_name):
404			return getattr(self, '_' + property_name)
405
406			def _nested_brick_property_setter(self, value, property_name):
407			setattr(self, '_' + property_name, value)
408			for act in self.activations:
409			assert isinstance(act.children[0], BatchNormalization)
410			setattr(act.children[0], property_name, value)
411
412			# conserve_memory is a bit special in that it can be modified
413			# after construction/allocation and still have a valid effect on
414			# apply(). Thus we propagate down all property sets.
415			conserve_memory = property(partial(_nested_brick_property_getter,
416			property_name='conserve_memory'),
417			partial(_nested_brick_property_setter,
418			property_name='conserve_memory'),
419			doc="Conserve memory.")
420
421			def _push_allocation_config(self):
422			super(BatchNormalizedMLP, self)._push_allocation_config()
423			# Do the extra allocation pushing for the BatchNormalization
424			# bricks. They need as their input dimension the output dimension
425			# of each linear transformation. Exclude the first dimension,
426			# which is the input dimension.
427			for act, dim in equizip(self.activations, self.dims[1:]):
428			assert isinstance(act.children[0], BatchNormalization)
429			act.children[0].input_dim = dim
430			for attr in ['mean_only', 'learn_scale', 'learn_shift']:
431			setattr(act.children[0], attr, getattr(self, attr))
432

mila-udem / blocks

Pull Request — master (#1120)

BatchNormalization.normalization_axes() A

Complexity

Size

Duplication

Importance

Duplication Side-by-Side

Filter issues like