conv.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.
"""
All layers just related to the neural network.
"""
from __future__ import print_function

import os
import inspect
import warnings
import itertools

import numpy as np
import six

import paddle
from ..mpc_layer_helper import MpcLayerHelper
from ..framework import MpcVariable, check_mpc_variable_and_dtype
from functools import reduce
import paddle

__all__ = [
    'conv2d',
]

def _convert_to_list(value, n, name, dtype):
    """
    Converts a single numerical type or iterable of numerical
    types into an numerical type list.

    Arguments:
      value: The value to validate and convert. Could an int, or any iterable
        of ints.
      n: The size of the list to be returned.
      name: The name of the argument being validated, e.g. "stride" or
        "filter_size". This is only used to format error messages.
      dtype: the numerical type of the element of the list to be returned.

    Returns:
      A list of n dtypes.

    Raises:
      ValueError: If something else than an int/long or iterable thereof was
        passed.
    """
    if isinstance(value, dtype):
        return [value, ] * n
    else:
        try:
            value_list = list(value)
        except TypeError:
            raise ValueError("The " + name +
                             "'s type must be list or tuple. Received: " + str(
                                 value))
        if len(value_list) != n:
            raise ValueError("The " + name + "'s length must be " + str(n) +
                             ". Received: " + str(value))
        for single_value in value_list:
            try:
                dtype(single_value)
            except (ValueError, TypeError):
                raise ValueError(
                    "The " + name + "'s type must be a list or tuple of " + str(
                        n) + " " + str(dtype) + " . Received: " + str(
                            value) + " "
                    "including element " + str(single_value) + " of type" + " "
                    + str(type(single_value)))
        return value_list


def _is_symmetric_padding(padding, data_dim):
    """
    Check whether padding is symmetrical.
    """
    assert len(padding) == data_dim * 2 or len(padding) == data_dim
    is_sys = True
    if len(padding) == data_dim * 2:
        for i in range(data_dim):
            if padding[i * 2] != padding[i * 2 + 1]:
                is_sys = False
    return is_sys


def conv2d(input,
           num_filters,
           filter_size,
           stride=1,
           padding=0,
           dilation=1,
           groups=None,
           param_attr=None,
           bias_attr=None,
           act=None,
           name=None,
           data_format="NCHW"):
    """
    The convolution2D layer calculates the output based on the input, filter
    and strides, paddings, dilations, groups parameters. Input and
    Output are in NCHW or NHWC format, where N is batch size, C is the number of
    channels, H is the height of the feature, and W is the width of the feature.
    Filter is in MCHW format, where M is the number of output image channels,
    C is the number of input image channels, H is the height of the filter,
    and W is the width of the filter. If the groups is greater than 1,
    C will equal the number of input image channels divided by the groups.
    Please refer to UFLDL's `convolution
    <http://ufldl.stanford.edu/tutorial/supervised/FeatureExtractionUsingConvolution/>`_
    for more details.
    If bias attribution and activation type are provided, bias is added to the
    output of the convolution, and the corresponding activation function is
    applied to the final result.

    For each input :math:`X`, the equation is:

    .. math::

        Out = \sigma (W \\ast X + b)

    Where:

    * :math:`X`: Input value, a tensor with NCHW or NHWC format.
    * :math:`W`: Filter value, a tensor with MCHW format.
    * :math:`\\ast`: Convolution operation.
    * :math:`b`: Bias value, a 2-D tensor with shape [M, 1].
    * :math:`\\sigma`: Activation function.
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.

    Example:

        - Input:

          Input shape: :math:`(N, C_{in}, H_{in}, W_{in})`

          Filter shape: :math:`(C_{out}, C_{in}, H_f, W_f)`

        - Output:

          Output shape: :math:`(N, C_{out}, H_{out}, W_{out})`

        Where

        .. math::

            H_{out}&= \\frac{(H_{in} + 2 * paddings[0] - (dilations[0] * (H_f - 1) + 1))}{strides[0]} + 1 \\\\
            W_{out}&= \\frac{(W_{in} + 2 * paddings[1] - (dilations[1] * (W_f - 1) + 1))}{strides[1]} + 1

    Args:
        input (Variable): The input is 4-D Tensor with shape [N, C, H, W], the data type
            of input is float16 or float32 or float64.
        num_filters(int): The number of filter. It is as same as the output
            image channel.
        filter_size (int|tuple): The filter size. If filter_size
            is a tuple, it must contain two integers, (filter_size_height,
            filter_size_width). Otherwise, filter_size_height = filter_size_width =\
            filter_size.
        stride (int|tuple): The stride size. It means the stride in convolution.
            If stride is a tuple, it must contain two integers, (stride_height, stride_width).
            Otherwise, stride_height = stride_width = stride. Default: stride = 1.
        padding (string|int|list|tuple): The padding size. It means the number of zero-paddings
            on both sides for each dimension.If `padding` is a string, either 'VALID' or
            'SAME' which is the padding algorithm. If padding size is a tuple or list,
            it could be in three forms: `[pad_height, pad_width]` or
            `[pad_height_top, pad_height_bottom, pad_width_left, pad_width_right]`, and when
            `data_format` is `"NCHW"`, `padding` can be in the form `[[0,0], [0,0],
            [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right]]`.
            when `data_format` is `"NHWC"`, `pool_padding` can be in the form
            `[[0,0], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right], [0,0]]`.
            Default: padding = 0.
        dilation (int|tuple): The dilation size. It means the spacing between the kernel
            points. If dilation is a tuple, it must contain two integers, (dilation_height,
            dilation_width). Otherwise, dilation_height = dilation_width = dilation.
            Default: dilation = 1.
        groups (int): The groups number of the Conv2d Layer. According to grouped
            convolution in Alex Krizhevsky's Deep CNN paper: when group=2,
            the first half of the filters is only connected to the first half
            of the input channels, while the second half of the filters is only
            connected to the second half of the input channels. Default: groups=1.
        param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights
            of conv2d. If it is set to None or one attribute of ParamAttr, conv2d
            will create ParamAttr as param_attr. If the Initializer of the param_attr
            is not set, the parameter is initialized with :math:`Normal(0.0, std)`,
            and the :math:`std` is :math:`(\\frac{2.0 }{filter\_elem\_num})^{0.5}`. Default: None.
        bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of conv2d.
            If it is set to False, no bias will be added to the output units.
            If it is set to None or one attribute of ParamAttr, conv2d
            will create ParamAttr as bias_attr. If the Initializer of the bias_attr
            is not set, the bias is initialized zero. Default: None.
        act (str): Activation type, if it is set to None, activation is not appended.
            Default: None
        name(str|None): For detailed information, please refer
           to :ref:`api_guide_Name`. Usually name is no need to set and
           None by default.
        data_format (str, optional): Specify the data format of the input, and the data format of the output
            will be consistent with that of the input. An optional string from: `"NCHW"`, `"NHWC"`.
            The default is `"NCHW"`. When it is `"NCHW"`, the data is stored in the order of:
            `[batch_size, input_channels, input_height, input_width]`.

    Returns:
        A Variable holding Tensor representing the conv2d, whose data type is the
        same with input. If act is None, the tensor variable storing the convolution
        result, and if act is not None, the tensor variable storing convolution
        and non-linearity activation result.

    Raises:
        ValueError: If using "depthwise_conv2d" (which is not supported yet).
        ValueError: If `data_format` is not "NCHW" or "NHWC".
        ValueError: If the channel dimmention of the input is less than or equal to zero.
        ValueError: If `padding` is a string, but not "SAME" or "VALID".
        ValueError: If `padding` is a tuple, but the element corresponding to the input's batch size is not 0
            or the element corresponding to the input's channel is not 0.
        ShapeError: If the input is not 4-D Tensor.
        ShapeError: If the input's dimension size and filter's dimension size not equal.
        ShapeError: If the dimension size of input minus the size of `stride` is not 2.
        ShapeError: If the number of input channels is not equal to filter's channels * groups.
        ShapeError: If the number of output channels is not be divided by groups.

    Examples:
        .. code-block:: python

          import paddle.fluid as fluid
          data = fluid.data(name='data', shape=[None, 3, 32, 32], dtype='float32')
          conv2d = fluid.layers.conv2d(input=data, num_filters=2, filter_size=3, act="relu")
    """

    check_mpc_variable_and_dtype(input, 'input', ['int64'],
            'conv2d')
    num_channels = input.shape[1 + 1]
    use_cudnn = False
    if not isinstance(use_cudnn, bool):
        raise ValueError("Attr(use_cudnn) should be True or False. Received "
                         "Attr(use_cudnn): %s. " % str(use_cudnn))

    if data_format not in ["NCHW", "NHWC"]:
        raise ValueError(
            "Attr(data_format) should be 'NCHW' or 'NHWC'. Received "
            "Attr(data_format): %s." % str(data_format))

    channel_last = (data_format == "NHWC")
    num_channels = input.shape[3 + 1] if channel_last else input.shape[1 + 1]
    if num_channels < 0:
        raise ValueError(
            "The channel dimmention of the input(%s) should be defined. "
            "Received: %s." % (str(input.shape), str(num_channels)))
    assert param_attr is not False, "param_attr should not be False here."

    l_type = 'conv2d'
    if (num_channels == groups and num_filters % num_channels == 0 and
            not use_cudnn):
        l_type = 'depthwise_conv2d'
        raise ValueError("l_type"
                "%s is not implemented yet. " % (str(l_type)))

    helper = MpcLayerHelper(l_type, **locals())
    dtype = helper.input_dtype()

    if groups is None:
        num_filter_channels = num_channels
    else:
        if num_channels % groups != 0:
            raise ValueError(
                "the channel of input must be divisible by groups,"
                "received: the channel of input is {}, the shape of input is {}"
                ", the groups is {}".format(num_channels, input.shape, groups))
        num_filter_channels = num_channels // groups

    filter_size = _convert_to_list(filter_size, 2, 'filter_size', np.int)
    stride = _convert_to_list(stride, 2, 'stride', np.int)
    dilation = _convert_to_list(dilation, 2, 'dilation', np.int)

    # padding
    def _update_padding(padding, data_format):
        """ update padding accroding to data_format
        raise ValueError if padding is not supported
        """
        def is_list_or_tuple(ele):
            """ return True if ele is a list or tuple
            """
            if isinstance(ele, list) or isinstance(ele, tuple):
                return True
            return False

        if is_list_or_tuple(padding) and len(padding) == 4:
            if is_list_or_tuple(padding[0]) and (data_format == "NCHW"):
                if not (padding[0] == [0, 0] and padding[1] == [0, 0]):
                    raise ValueError(
                        "Non-zero padding(%s) in the batch or channel dimensions "
                        "is not supported." % str(padding))
                padding = padding[2:4]
                padding = list(itertools.chain(*padding))
            elif is_list_or_tuple(padding[0]) and (data_format == "NHWC"):
                if not (padding[0] == [0, 0] and padding[3] == [0, 0]):
                    raise ValueError(
                        "Non-zero padding(%s) in the batch or channel dimensions "
                        "is not supported." % str(padding))
                padding = padding[1:3]
                padding = list(itertools.chain(*padding))
            padding = _convert_to_list(padding, 4, 'padding', np.int)
            if _is_symmetric_padding(padding, 2):
                padding = [padding[0], padding[2]]

        else:
            padding = _convert_to_list(padding, 2, 'padding', np.int)

        return padding

    padding_algorithm = "EXPLICIT"
    if isinstance(padding, str):
        padding = padding.upper()
        if padding not in ["SAME", "VALID"]:
            raise ValueError(
                "Unknown padding: '%s'. It can only be 'SAME' or 'VALID'." %
                str(padding))
        if padding == "VALID":
            padding_algorithm = "VALID"
            padding = [0, 0]
        elif padding == "SAME":
            padding_algorithm = "SAME"
            padding = [0, 0]

    padding = _update_padding(padding, data_format)

    filter_shape = [num_filters, int(num_filter_channels)] + filter_size

    filter_param = helper.create_mpc_parameter(
        attr=helper.param_attr,
        shape=filter_shape,
        dtype=dtype)

    pre_bias = helper.create_mpc_variable_for_type_inference(dtype)

    helper.append_op(
        type= 'mpc_' + l_type,
        inputs={
            'Input': input,
            'Filter': filter_param,
        },
        outputs={"Output": pre_bias},
        attrs={
            'strides': stride,
            'paddings': padding,
            'dilations': dilation,
            'groups': groups,
            'use_cudnn': use_cudnn,
            'use_mkldnn': False,
            'fuse_relu_before_depthwise_conv': False,
            "padding_algorithm": padding_algorithm,
            "data_format": data_format,
        })

    if data_format == 'NCHW':
        pre_act = helper.append_mpc_bias_op(pre_bias, dim_start=1, dim_end=2 + 1)
    else:
        pre_act = helper.append_mpc_bias_op(pre_bias, dim_start=3, dim_end=4 + 1)

    return helper.append_mpc_activation(pre_act)