nn.py

# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import paddle
from .. import core
from ..layers import utils
from ..layers import nn as F
from .. import dygraph_utils
from . import layers
from ..framework import (
    Variable,
    _non_static_mode,
    OpProtoHolder,
    Parameter,
    _dygraph_tracer,
    _varbase_creator,
    default_main_program,
    _global_flags,
    in_dygraph_mode,
    _in_legacy_dygraph,
)

from ..data_feeder import (
    convert_dtype,
    check_variable_and_dtype,
    check_type,
    check_dtype,
)

from ..param_attr import ParamAttr
from ..initializer import Normal, Constant, NumpyArrayInitializer
from .. import unique_name
from .layer_object_helper import LayerObjectHelper
from ..data_feeder import check_variable_and_dtype, check_type
import numpy as np
import numbers
import logging
import os
import paddle.utils.deprecated as deprecated
from paddle import _C_ops, _legacy_C_ops

__all__ = []


class BatchNorm(layers.Layer):
    r"""

    This interface is used to construct a callable object of the ``BatchNorm`` class.
    For more details, refer to code examples.
    It implements the function of the Batch Normalization Layer and can be used
    as a normalizer function for conv2d and fully connected operations.
    The data is normalized by the mean and variance of the channel based on the current batch data.
    Refer to `Batch Normalization: Accelerating Deep Network Training by Reducing
    Internal Covariate Shift <https://arxiv.org/pdf/1502.03167.pdf>`_
    for more details.

    When use_global_stats = False, the :math:`\mu_{\beta}`
    and :math:`\sigma_{\beta}^{2}` are the statistics of one mini-batch.
    Calculated as follows:

    ..  math::

        \mu_{\beta} &\gets \frac{1}{m} \sum_{i=1}^{m} x_i \qquad &
        //\ mini-batch\ mean \\
        \sigma_{\beta}^{2} &\gets \frac{1}{m} \sum_{i=1}^{m}(x_i - \mu_{\beta})^2 \qquad &
        //\ mini-batch\ variance \\

    - :math:`x` : mini-batch data
    - :math:`m` : the size of the mini-batch data

    When use_global_stats = True, the :math:`\\mu_{\\beta}`
    and :math:`\\sigma_{\\beta}^{2}` are not the statistics of one mini-batch.
    They are global or running statistics (moving_mean and moving_variance). It usually got from the
    pre-trained model. Calculated as follows:

    .. math::
        moving\_mean = moving\_mean * momentum + \mu_{\beta} * (1. - momentum) \quad &// global mean \\
        moving\_variance = moving\_variance * momentum + \sigma_{\beta}^{2} * (1. - momentum) \quad &// global variance \\

    The normalization function formula is as follows:

    ..  math::

        \hat{x_i} &\gets \frac{x_i - \mu_\beta} {\sqrt{\
        \sigma_{\beta}^{2} + \epsilon}} \qquad &//\ normalize \\
        y_i &\gets \gamma \hat{x_i} + \beta \qquad &//\ scale\ and\ shift


    - :math:`\epsilon` : add a smaller value to the variance to prevent division by zero
    - :math:`\gamma` : trainable proportional parameter
    - :math:`\beta` : trainable deviation parameter

    Parameters:
        num_channels(int): Indicate the number of channels of the input ``Tensor``.
        act(str, optional): Activation to be applied to the output of batch normalization. Default: None.
        is_test (bool, optional): A flag indicating whether it is in test phrase or not.
             This flag only has effect on static graph mode. For dygraph mode, please use ``eval()``.
             Default: False.
        momentum(float, optional): The value used for the moving_mean and moving_var computation. Default: 0.9.
        epsilon(float, optional): The small value added to the variance to prevent division by zero. Default: 1e-5.
        param_attr(ParamAttr, optional): The parameter attribute for Parameter `scale`
             of batch_norm. If it is set to None or one attribute of ParamAttr, batch_norm
             will create ParamAttr as param_attr. If the Initializer of the param_attr
             is not set, the parameter is initialized with Xavier. Default: None.
        bias_attr(ParamAttr, optional): The parameter attribute for the bias of batch_norm.
             If it is set to None or one attribute of ParamAttr, batch_norm
             will create ParamAttr as bias_attr. If the Initializer of the bias_attr
             is not set, the bias is initialized zero. Default: None.
        dtype(str, optional): Indicate the data type of the input ``Tensor``,
             which can be float32 or float64. Default: float32.
        data_layout(str, optional): Specify the input data format, the data format can be "NCHW" or "NHWC", where `N` is batch size, `C` is the number of the feature map, `H` is the height of the feature map, `W` is the width of the feature map. Default: NCHW.
        in_place(bool, optional): Make the input and output of batch norm reuse memory. Default: False.
        moving_mean_name(str, optional): The name of moving_mean which store the global Mean. Default: None.
        moving_variance_name(str, optional): The name of the moving_variance which store the global Variance. Default: None.
        do_model_average_for_mean_and_var(bool, optional): Whether parameter mean and variance should do model
            average when model average is enabled. Default: True.
        use_global_stats(bool, optional): Whether to use global mean and
            variance. In inference or test mode, set use_global_stats to true
            or is_test to true, and the behavior is equivalent.
            In train mode, when setting use_global_stats True, the global mean
            and variance are also used during train period. Default: False.
        trainable_statistics(bool, optional): Whether to calculate mean and var in eval mode. In eval mode, when
            setting trainable_statistics True, mean and variance will be calculated by current batch statistics.
            Default: False.

    Returns:
        None

    Examples:
        .. code-block:: python

          import paddle
          import paddle.fluid as fluid
          from paddle.fluid.dygraph.base import to_variable

          x = paddle.rand([3, 10, 3, 7], 'float32')
          with fluid.dygraph.guard():
              x = to_variable(x)
              batch_norm = fluid.BatchNorm(10)
              hidden1 = batch_norm(x)
    """

    def __init__(
        self,
        num_channels,
        act=None,
        is_test=False,
        momentum=0.9,
        epsilon=1e-05,
        param_attr=None,
        bias_attr=None,
        dtype='float32',
        data_layout='NCHW',
        in_place=False,
        moving_mean_name=None,
        moving_variance_name=None,
        do_model_average_for_mean_and_var=True,
        use_global_stats=False,
        trainable_statistics=False,
    ):
        super().__init__()
        self._param_attr = param_attr
        self._bias_attr = bias_attr
        self._act = act
        self._use_mkldnn = _global_flags()["FLAGS_use_mkldnn"]

        assert (
            bias_attr is not False
        ), "bias_attr should not be False in batch_norm."

        if dtype == "float16":
            self._dtype = "float32"
        else:
            self._dtype = dtype

        param_shape = [num_channels]

        # create parameter
        self.weight = self.create_parameter(
            attr=self._param_attr,
            shape=param_shape,
            dtype=self._dtype,
            default_initializer=Constant(1.0),
        )
        self.weight.stop_gradient = (
            use_global_stats and self._param_attr.learning_rate == 0.0
        )

        self.bias = self.create_parameter(
            attr=self._bias_attr,
            shape=param_shape,
            dtype=self._dtype,
            is_bias=True,
        )
        self.bias.stop_gradient = (
            use_global_stats and self._param_attr.learning_rate == 0.0
        )

        self._mean = self.create_parameter(
            attr=ParamAttr(
                name=moving_mean_name,
                initializer=Constant(0.0),
                trainable=False,
                do_model_average=do_model_average_for_mean_and_var,
            ),
            shape=param_shape,
            dtype=self._dtype,
        )
        self._mean.stop_gradient = True

        self._variance = self.create_parameter(
            attr=ParamAttr(
                name=moving_variance_name,
                initializer=Constant(1.0),
                trainable=False,
                do_model_average=do_model_average_for_mean_and_var,
            ),
            shape=param_shape,
            dtype=self._dtype,
        )
        self._variance.stop_gradient = True

        self._in_place = in_place
        self._data_layout = data_layout
        self._momentum = momentum
        self._epsilon = epsilon
        self._is_test = is_test
        self._fuse_with_relu = False
        self._use_global_stats = use_global_stats
        self._trainable_statistics = trainable_statistics

    def forward(self, input):
        # create output
        # mean and mean_out share the same memory
        mean_out = self._mean
        # variance and variance out share the same memory
        variance_out = self._variance

        if _non_static_mode():
            if in_dygraph_mode():
                batch_norm_out, t1, t2, t3, t4, _ = _C_ops.batch_norm(
                    input,
                    self._mean,
                    self._variance,
                    self.weight,
                    self.bias,
                    not self.training,
                    self._momentum,
                    self._epsilon,
                    self._data_layout,
                    self._use_global_stats,
                    self._trainable_statistics,
                )
                return dygraph_utils._append_activation_in_dygraph(
                    batch_norm_out, act=self._act, use_mkldnn=self._use_mkldnn
                )

            elif _in_legacy_dygraph():
                attrs = (
                    "momentum",
                    self._momentum,
                    "epsilon",
                    self._epsilon,
                    "is_test",
                    not self.training,
                    "data_layout",
                    self._data_layout,
                    "use_mkldnn",
                    self._use_mkldnn,
                    "fuse_with_relu",
                    self._fuse_with_relu,
                    "use_global_stats",
                    self._use_global_stats,
                    'trainable_statistics',
                    self._trainable_statistics,
                )
                batch_norm_out, _, _, _, _, _ = _legacy_C_ops.batch_norm(
                    input,
                    self.weight,
                    self.bias,
                    self._mean,
                    self._variance,
                    None,
                    mean_out,
                    variance_out,
                    *attrs
                )

            return dygraph_utils._append_activation_in_dygraph(
                batch_norm_out, act=self._act, use_mkldnn=self._use_mkldnn
            )

        check_variable_and_dtype(
            input, 'input', ['float16', 'float32', 'float64'], 'BatchNorm'
        )

        attrs = {
            "momentum": self._momentum,
            "epsilon": self._epsilon,
            "is_test": self._is_test,
            "data_layout": self._data_layout,
            "use_mkldnn": False,
            "fuse_with_relu": self._fuse_with_relu,
            "use_global_stats": self._use_global_stats,
            "trainable_statistics": self._trainable_statistics,
        }

        inputs = {
            "X": [input],
            "Scale": [self.weight],
            "Bias": [self.bias],
            "Mean": [self._mean],
            "Variance": [self._variance],
        }

        saved_mean = self._helper.create_variable_for_type_inference(
            dtype=self._dtype, stop_gradient=True
        )
        saved_variance = self._helper.create_variable_for_type_inference(
            dtype=self._dtype, stop_gradient=True
        )
        reserve_space = self._helper.create_variable_for_type_inference(
            dtype=self._helper.input_dtype(input), stop_gradient=True
        )

        batch_norm_out = (
            input
            if self._in_place
            else self._helper.create_variable_for_type_inference(self._dtype)
        )

        outputs = {
            "Y": [batch_norm_out],
            "MeanOut": [mean_out],
            "VarianceOut": [variance_out],
            "SavedMean": [saved_mean],
            "SavedVariance": [saved_variance],
        }
        if reserve_space is not None:
            outputs["ReserveSpace"] = [reserve_space]

        self._helper.append_op(
            type="batch_norm", inputs=inputs, outputs=outputs, attrs=attrs
        )

        # Currently, we don't support inplace in dygraph mode
        return self._helper.append_activation(batch_norm_out, self._act)


class RowConv(layers.Layer):
    """
    ***Row-convolution operator***

    The row convolution is called lookahead convolution.  This operator was introduced in the following paper for DeepSpeech2:
    http://www.cs.cmu.edu/~dyogatam/papers/wang+etal.iclrworkshop2016.pdf

    The main motivation is that a bidirectional RNN, useful in DeepSpeech like speech models, learns representation for a sequence by performing a
    forward and a backward pass through the entire sequence. However, unlike
    unidirectional RNNs, bidirectional RNNs are challenging to deploy in an online
    and low-latency setting. The lookahead convolution incorporates information
    from future subsequences in a computationally efficient manner to improve
    unidirectional recurrent neural networks. The row convolution operator is
    different from the 1D sequence convolution, and is computed as follows:

    Given an input sequence X of length t and input dimension D, and a filter (W) of size context * D.

    More details about row_conv please refer to the design document https://github.com/PaddlePaddle/Paddle/issues/2228#issuecomment-303903645 .

    Parameters:
        name_scope(str): The name of this class.
        future_context_size (int): Future context size. Please note, the shape
            of convolution kernel is [future_context_size + 1, D].
        param_attr (ParamAttr): Attributes of parameters, including
            name, initializer etc. Default: None.
        act (str): Non-linear activation to be applied to output variable. Default: None.

    Attributes:
        weight (Parameter): the learnable weights of this layer.

    Returns:
        the output(Out) is a LodTensor, which supports variable time-length input sequences.
        The underlying tensor in this LodTensor is a matrix with shape T x N, i.e., the same shape as X.

    Examples:
        .. code-block:: python

          import paddle.fluid as fluid
          import numpy

          with fluid.dygraph.guard():
              x = numpy.random.random((16)).astype('float32')
              rowConv = fluid.dygraph.nn.RowConv(
                    'RowConv', future_context_size=2)
              ret = rowConv(fluid.dygraph.base.to_variable(x))

    """

    def __init__(
        self, name_scope, future_context_size, param_attr=None, act=None
    ):
        assert (
            not _non_static_mode()
        ), "RowConv is not supported by dynamic graph mode yet!"
        super().__init__(name_scope)
        self._act = act
        self._param_attr = param_attr
        self._future_context_size = future_context_size

    def _build_once(self, input):
        self._dtype = self._helper.input_dtype(input)
        filter_shape = [self._future_context_size + 1, input.shape[1]]
        self.weight = self.create_parameter(
            attr=self._param_attr,
            shape=filter_shape,
            dtype=self._dtype,
            is_bias=False,
        )

    def forward(self, input):
        out = self._helper.create_variable_for_type_inference(self._dtype)
        self._helper.append_op(
            type='row_conv',
            inputs={'X': [input], 'Filter': [self.weight]},
            outputs={'Out': [out]},
        )
        return self._helper.append_activation(out, act=self._act)