MegEngine/imperative/python/megengine/module/conv.py

# MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
#
# Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
#
# Unless required by applicable law or agreed to in writing,
# software distributed under the License is distributed on an
# "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
from abc import abstractmethod
from typing import Tuple, Union

import numpy as np

from ..functional import (
    conv1d,
    conv2d,
    conv3d,
    conv_transpose2d,
    conv_transpose3d,
    deformable_conv2d,
    local_conv2d,
    relu,
)
from ..tensor import Parameter
from ..utils.tuple_function import _pair, _pair_nonzero, _triple, _triple_nonzero
from . import init
from .module import Module


class _ConvNd(Module):
    """base class for convolution modules, including transposed conv"""

    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        kernel_size: Union[int, Tuple[int, int]],
        stride: Union[int, Tuple[int, int]],
        padding: Union[int, Tuple[int, int]],
        dilation: Union[int, Tuple[int, int]],
        groups: int,
        bias: bool = True,
        **kwargs
    ):
        super().__init__(**kwargs)
        if in_channels % groups != 0:
            raise ValueError("in_channels must be divisible by groups")
        if out_channels % groups != 0:
            raise ValueError("out_channels must be divisible by groups")
        self.in_channels = in_channels
        self.out_channels = out_channels
        self.kernel_size = kernel_size
        self.stride = stride
        self.padding = padding
        self.dilation = dilation
        self.groups = groups

        self.weight = Parameter(np.zeros(self._infer_weight_shape(), dtype=np.float32))
        self.bias = None
        if bias:
            self.bias = Parameter(np.zeros(self._infer_bias_shape(), dtype=np.float32))
        self.reset_parameters()

    @abstractmethod
    def _get_fanin(self):
        pass

    def reset_parameters(self) -> None:
        fanin = self._get_fanin()
        std = np.sqrt(1 / fanin)
        init.normal_(self.weight, 0.0, std)
        if self.bias is not None:
            init.zeros_(self.bias)

    @abstractmethod
    def _infer_weight_shape(self):
        pass

    @abstractmethod
    def _infer_bias_shape(self):
        pass

    def _module_info_string(self):
        s = "{in_channels}, {out_channels}, kernel_size={kernel_size}"

        if self.stride != (1,) * len(self.stride):
            s += ", stride={stride}"
        if self.padding != (0,) * len(self.padding):
            s += ", padding={padding}"
        if self.dilation != (1,) * len(self.dilation):
            s += ", dilation={dilation}"
        if self.groups != 1:
            s += ", groups={groups}"
        if self.bias is None:
            s += ", bias=False"
        return s.format(**self.__dict__)


class Conv1d(_ConvNd):

    r"""
    Applies a 1D convolution over an input tensor.

    For instance, given an input of the size :math:`(N, C_{\text{in}}, H)`,
    this layer generates an output of the size
    :math:`(N, C_{\text{out}}, H_{\text{out}})` through the
    process described as below:

    .. math::
        \text{out}(N_i, C_{\text{out}_j}) = \text{bias}(C_{\text{out}_j}) +
        \sum_{k = 0}^{C_{\text{in}} - 1} \text{weight}(C_{\text{out}_j}, k) \star \text{input}(N_i, k)

    where :math:`\star` is the valid 1D cross-correlation operator,
    :math:`N` is batch size, :math:`C` denotes number of channels, and
    :math:`H` is length of 1D data element.


    When `groups == in_channels` and `out_channels == K * in_channels`,
    where K is a positive integer, this operation is also known as depthwise
    convolution.

    In other words, for an input of size :math:`(N, C_{in}, H_{in})`,
    a depthwise convolution with a depthwise multiplier `K`, can be constructed
    by arguments :math:`(in\_channels=C_{in}, out\_channels=C_{in} \times K, ..., groups=C_{in})`.

    :param in_channels: number of input channels.
    :param out_channels: number of output channels.
    :param kernel_size: size of weight on spatial dimensions. If kernel_size is
        an :class:`int`, the actual kernel size would be
        `(kernel_size, kernel_size)`. Default: 1
    :param stride: stride of the 1D convolution operation. Default: 1
    :param padding: size of the paddings added to the input on both sides of its
        spatial dimensions. Only zero-padding is supported. Default: 0
    :param dilation: dilation of the 1D convolution operation. Default: 1
    :param groups: number of groups into which the input and output channels are divided,
        so as to perform a "grouped convolution". When ``groups`` is not 1,
        ``in_channels`` and ``out_channels`` must be divisible by ``groups``,
        and there would be an extra dimension at the beginning of the weight's
        shape. Specifically, the shape of weight would be `(groups,
        out_channel // groups, in_channels // groups, *kernel_size)`.
    :param bias: whether to add a bias onto the result of convolution. Default:
        True
    :param conv_mode: Supports `cross_correlation`. Default:
        `cross_correlation`
    :param compute_mode: When set to "default", no special requirements will be
        placed on the precision of intermediate results. When set to "float32",
        "float32" would be used for accumulator and intermediate result, but only
        effective when input and output are of float16 dtype.

    Examples:

    .. testcode::

        import numpy as np
        import megengine as mge
        import megengine.module as M

        m = M.Conv1d(in_channels=3, out_channels=1, kernel_size=3)
        inp = mge.tensor(np.arange(0, 24).astype("float32").reshape(2, 3, 4))
        oup = m(inp)
        print(oup.numpy().shape)

    Outputs:

    .. testoutput::

        (2, 1, 2)

    """

    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        kernel_size: int,
        stride: int = 1,
        padding: int = 0,
        dilation: int = 1,
        groups: int = 1,
        bias: bool = True,
        conv_mode: str = "cross_correlation",
        compute_mode: str = "default",
        **kwargs
    ):
        kernel_size = kernel_size
        stride = stride
        padding = padding
        dilation = dilation
        self.conv_mode = conv_mode
        self.compute_mode = compute_mode
        super().__init__(
            in_channels,
            out_channels,
            kernel_size,
            stride,
            padding,
            dilation,
            groups,
            bias,
            **kwargs,
        )

    def _get_fanin(self):
        kh = self.kernel_size
        ic = self.in_channels
        return kh * ic

    def _infer_weight_shape(self):
        group = self.groups
        ichl = self.in_channels
        ochl = self.out_channels
        kh = self.kernel_size
        if group == 1:
            # Assume format is NCH(W=1)
            return (ochl, ichl, kh)

        assert (
            ichl % group == 0 and ochl % group == 0
        ), "invalid config: input_channels={} output_channels={} group={}".format(
            ichl, ochl, group
        )
        # Assume format is NCH(W=1)
        return (group, ochl // group, ichl // group, kh)

    def _infer_bias_shape(self):
        # Assume format is NCH(W=1)
        return (1, self.out_channels, 1)

    def calc_conv(self, inp, weight, bias):
        return conv1d(
            inp,
            weight,
            bias,
            self.stride,
            self.padding,
            self.dilation,
            self.groups,
            self.conv_mode,
            self.compute_mode,
        )

    def forward(self, inp):
        return self.calc_conv(inp, self.weight, self.bias)


class Conv2d(_ConvNd):
    r"""
    Applies a 2D convolution over an input tensor.

    For instance, given an input of the size :math:`(N, C_{\text{in}}, H, W)`,
    this layer generates an output of the size
    :math:`(N, C_{\text{out}}, H_{\text{out}}, W_{\text{out}})` through the
    process described as below:

    .. math::
        \text{out}(N_i, C_{\text{out}_j}) = \text{bias}(C_{\text{out}_j}) +
        \sum_{k = 0}^{C_{\text{in}} - 1} \text{weight}(C_{\text{out}_j}, k) \star \text{input}(N_i, k)

    where :math:`\star` is the valid 2D cross-correlation operator,
    :math:`N` is batch size, :math:`C` denotes number of channels,
    :math:`H` is height of input planes in pixels, and :math:`W` is
    width in pixels.

    In general, output feature maps' shapes can be inferred as follows:

    input: :math:`(N, C_{\text{in}}, H_{\text{in}}, W_{\text{in}})`

    output: :math:`(N, C_{\text{out}}, H_{\text{out}}, W_{\text{out}})` where

    .. math::
        \text{H}_{out} = \lfloor \frac{\text{H}_{in} + 2 * \text{padding[0]} - 
        \text{dilation[0]} * (\text{kernel_size[0]} - 1) - 1}{\text{stride[0]}} + 1 \rfloor

    .. math::
        \text{W}_{out} = \lfloor \frac{\text{W}_{in} + 2 * \text{padding[1]} - 
        \text{dilation[1]} * (\text{kernel_size[1]} - 1) - 1}{\text{stride[1]}} + 1 \rfloor

    When `groups == in_channels` and `out_channels == K * in_channels`,
    where K is a positive integer, this operation is also known as depthwise
    convolution.

    In other words, for an input of size :math:`(N, C_{in}, H_{in}, W_{in})`,
    a depthwise convolution with a depthwise multiplier `K`, can be constructed
    by arguments :math:`(in\_channels=C_{in}, out\_channels=C_{in} \times K, ..., groups=C_{in})`.

    :param in_channels: number of input channels.
    :param out_channels: number of output channels.
    :param kernel_size: size of weight on spatial dimensions. If kernel_size is
        an :class:`int`, the actual kernel size would be
        `(kernel_size, kernel_size)`. Default: 1
    :param stride: stride of the 2D convolution operation. Default: 1
    :param padding: size of the paddings added to the input on both sides of its
        spatial dimensions. Only zero-padding is supported. Default: 0
    :param dilation: dilation of the 2D convolution operation. Default: 1
    :param groups: number of groups into which the input and output channels are divided,
        so as to perform a "grouped convolution". When ``groups`` is not 1,
        ``in_channels`` and ``out_channels`` must be divisible by ``groups``,
        and there would be an extra dimension at the beginning of the weight's
        shape. Specifically, the shape of weight would be `(groups,
        out_channel // groups, in_channels // groups, *kernel_size)`.
    :param bias: whether to add a bias onto the result of convolution. Default:
        True
    :param conv_mode: Supports `cross_correlation`. Default:
        `cross_correlation`
    :param compute_mode: When set to "default", no special requirements will be
        placed on the precision of intermediate results. When set to "float32",
        "float32" would be used for accumulator and intermediate result, but only
        effective when input and output are of float16 dtype.

    Examples:

    .. testcode::

        import numpy as np
        import megengine as mge
        import megengine.module as M

        m = M.Conv2d(in_channels=3, out_channels=1, kernel_size=3)
        inp = mge.tensor(np.arange(0, 96).astype("float32").reshape(2, 3, 4, 4))
        oup = m(inp)
        print(oup.numpy().shape)

    Outputs:

    .. testoutput::

        (2, 1, 2, 2)

    """

    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        kernel_size: Union[int, Tuple[int, int]],
        stride: Union[int, Tuple[int, int]] = 1,
        padding: Union[int, Tuple[int, int]] = 0,
        dilation: Union[int, Tuple[int, int]] = 1,
        groups: int = 1,
        bias: bool = True,
        conv_mode: str = "cross_correlation",
        compute_mode: str = "default",
        **kwargs
    ):
        kernel_size = _pair_nonzero(kernel_size)
        stride = _pair_nonzero(stride)
        padding = _pair(padding)
        dilation = _pair_nonzero(dilation)
        self.conv_mode = conv_mode
        self.compute_mode = compute_mode
        super().__init__(
            in_channels,
            out_channels,
            kernel_size,
            stride,
            padding,
            dilation,
            groups,
            bias,
            **kwargs,
        )

    def _get_fanin(self):
        kh, kw = self.kernel_size
        ic = self.in_channels
        return kh * kw * ic

    def _infer_weight_shape(self):
        group = self.groups
        ichl = self.in_channels
        ochl = self.out_channels
        kh, kw = self.kernel_size
        if group == 1:
            # Assume format is NCHW
            return (ochl, ichl, kh, kw)

        assert (
            ichl % group == 0 and ochl % group == 0
        ), "invalid config: input_channels={} output_channels={} group={}".format(
            ichl, ochl, group
        )
        # Assume format is NCHW
        return (group, ochl // group, ichl // group, kh, kw)

    def _infer_bias_shape(self):
        # Assume format is NCHW
        return (1, self.out_channels, 1, 1)

    def calc_conv(self, inp, weight, bias):
        return conv2d(
            inp,
            weight,
            bias,
            self.stride,
            self.padding,
            self.dilation,
            self.groups,
            self.conv_mode,
            self.compute_mode,
        )

    def forward(self, inp):
        return self.calc_conv(inp, self.weight, self.bias)


class Conv3d(_ConvNd):

    r"""
    Applies a 3D convolution over an input tensor.

    For instance, given an input of the size :math:`(N, C_{\text{in}}, T, H, W)`,
    this layer generates an output of the size
    :math:`(N, C_{\text{out}}, T_{\text{out}}}, H_{\text{out}}}, W_{\text{out}}})` through the
    process described as below:

    .. math::
        \text{out}(N_i, C_{\text{out}_j}) = \text{bias}(C_{\text{out}_j}) +
        \sum_{k = 0}^{C_{\text{in}} - 1} \text{weight}(C_{\text{out}_j}, k) \star \text{input}(N_i, k)

    where :math:`\star` is the valid 3D cross-correlation operator,
    :math:`N` is batch size, :math:`C` denotes number of channels


    When `groups == in_channels` and `out_channels == K * in_channels`,
    where K is a positive integer, this operation is also known as depthwise
    convolution.

    In other words, for an input of size :math:`(N, C_{in}, T_{int}, H_{in}, W_{in})`,
    a depthwise convolution with a depthwise multiplier `K`, can be constructed
    by arguments :math:`(in\_channels=C_{in}, out\_channels=C_{in} \times K, ..., groups=C_{in})`.

    :param in_channels: number of input channels.
    :param out_channels: number of output channels.
    :param kernel_size: size of weight on spatial dimensions. If kernel_size is
        an :class:`int`, the actual kernel size would be
        `(kernel_size, kernel_size, kernel_size)`. Default: 1
    :param stride: stride of the 3D convolution operation. Default: 1
    :param padding: size of the paddings added to the input on both sides of its
        spatial dimensions. Only zero-padding is supported. Default: 0
    :param dilation: dilation of the 3D convolution operation. Default: 1
    :param groups: number of groups into which the input and output channels are divided,
        so as to perform a "grouped convolution". When ``groups`` is not 1,
        ``in_channels`` and ``out_channels`` must be divisible by ``groups``,
        and there would be an extra dimension at the beginning of the weight's
        shape. Specifically, the shape of weight would be `(groups,
        out_channel // groups, in_channels // groups, *kernel_size)`.
    :param bias: whether to add a bias onto the result of convolution. Default:
        True
    :param conv_mode: Supports `cross_correlation`. Default:
        `cross_correlation`

    Examples:

    .. testcode::

        import numpy as np
        import megengine as mge
        import megengine.module as M

        m = M.Conv3d(in_channels=3, out_channels=1, kernel_size=3)
        inp = mge.tensor(np.arange(0, 384).astype("float32").reshape(2, 3, 4, 4, 4))
        oup = m(inp)
        print(oup.numpy().shape)

    Outputs:

    .. testoutput::

        (2, 1, 2, 2, 2)

    """

    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        kernel_size: Union[int, Tuple[int, int, int]],
        stride: Union[int, Tuple[int, int, int]] = 1,
        padding: Union[int, Tuple[int, int, int]] = 0,
        dilation: Union[int, Tuple[int, int, int]] = 1,
        groups: int = 1,
        bias: bool = True,
        conv_mode: str = "cross_correlation",
    ):
        kernel_size = _triple_nonzero(kernel_size)
        stride = _triple_nonzero(stride)
        padding = _triple(padding)
        dilation = _triple_nonzero(dilation)
        self.conv_mode = conv_mode
        super().__init__(
            in_channels,
            out_channels,
            kernel_size,
            stride,
            padding,
            dilation,
            groups,
            bias,
        )

    def _get_fanin(self):
        kt, kh, kw = self.kernel_size
        ic = self.in_channels
        return kt * kh * kw * ic

    def _infer_weight_shape(self):
        group = self.groups
        ichl = self.in_channels
        ochl = self.out_channels
        kt, kh, kw = self.kernel_size
        if group == 1:
            # Assume format is NCTHW
            return (ochl, ichl, kt, kh, kw)

        assert (
            ichl % group == 0 and ochl % group == 0
        ), "invalid config: input_channels={} output_channels={} group={}".format(
            ichl, ochl, group
        )
        # Assume format is NCTHW
        return (group, ochl // group, ichl // group, kt, kh, kw)

    def _infer_bias_shape(self):
        # Assume format is NCTHW
        return (1, self.out_channels, 1, 1, 1)

    def calc_conv(self, inp, weight, bias):
        return conv3d(
            inp,
            weight,
            bias,
            self.stride,
            self.padding,
            self.dilation,
            self.groups,
            self.conv_mode,
        )

    def forward(self, inp):
        return self.calc_conv(inp, self.weight, self.bias)


class ConvTranspose2d(_ConvNd):
    r"""
    Applies a 2D transposed convolution over an input tensor.

    This module is also known as a deconvolution or a fractionally-strided convolution.
    :class:`ConvTranspose2d` can be seen as the gradient of :class:`Conv2d` operation
    with respect to its input.

    Convolution usually reduces the size of input, while transposed convolution works
    the opposite way, transforming a smaller input to a larger output while preserving the
    connectivity pattern.

    :param in_channels: number of input channels.
    :param out_channels: number of output channels.
    :param kernel_size: size of weight on spatial dimensions. If ``kernel_size`` is
        an :class:`int`, the actual kernel size would be
        ``(kernel_size, kernel_size)``. Default: 1
    :param stride: stride of the 2D convolution operation. Default: 1
    :param padding: size of the paddings added to the input on both sides of its
        spatial dimensions. Only zero-padding is supported. Default: 0
    :param dilation: dilation of the 2D convolution operation. Default: 1
    :param groups: number of groups into which the input and output channels are divided,
        so as to perform a "grouped convolution". When ``groups`` is not 1,
        ``in_channels`` and ``out_channels`` must be divisible by ``groups``,
        and there would be an extra dimension at the beginning of the weight's
        shape. Specifically, the shape of weight would be ``(groups,
        out_channels // groups, in_channels // groups, *kernel_size)``. Default: 1
    :param bias: wether to add a bias onto the result of convolution. Default:
        True
    :param conv_mode: Supports `cross_correlation`. Default:
        `cross_correlation`
    :param compute_mode: When set to "default", no special requirements will be
        placed on the precision of intermediate results. When set to "float32",
        "float32" would be used for accumulator and intermediate result, but only
        effective when input and output are of float16 dtype.
    """

    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        kernel_size: Union[int, Tuple[int, int]],
        stride: Union[int, Tuple[int, int]] = 1,
        padding: Union[int, Tuple[int, int]] = 0,
        dilation: Union[int, Tuple[int, int]] = 1,
        groups: int = 1,
        bias: bool = True,
        conv_mode: str = "cross_correlation",
        compute_mode: str = "default",
        **kwargs
    ):
        kernel_size = _pair_nonzero(kernel_size)
        stride = _pair_nonzero(stride)
        padding = _pair(padding)
        dilation = _pair_nonzero(dilation)
        self.conv_mode = conv_mode
        self.compute_mode = compute_mode
        super().__init__(
            in_channels,
            out_channels,
            kernel_size,
            stride,
            padding,
            dilation,
            groups,
            bias,
            **kwargs,
        )

    def _get_fanin(self):
        kh, kw = self.kernel_size
        oc = self.out_channels
        return kh * kw * oc

    def _infer_weight_shape(self):
        group = self.groups
        ichl = self.in_channels
        ochl = self.out_channels
        kh, kw = self.kernel_size
        if group == 1:
            # Assume format is NCHW
            return (ichl, ochl, kh, kw)

        assert (
            ichl % group == 0 and ochl % group == 0
        ), "invalid config: input_channels={} output_channels={} group={}".format(
            ichl, ochl, group
        )
        # Assume format is NCHW
        return (group, ichl // group, ochl // group, kh, kw)

    def _infer_bias_shape(self):
        # Assume format is NCHW
        return (1, self.out_channels, 1, 1)

    def forward(self, inp):
        return conv_transpose2d(
            inp,
            self.weight,
            self.bias,
            self.stride,
            self.padding,
            self.dilation,
            self.groups,
            self.conv_mode,
            self.compute_mode,
        )


class LocalConv2d(Conv2d):
    r"""
    Applies a spatial convolution with untied kernels over an groupped channeled input 4D tensor.
    It is also known as the locally connected layer.

    :param in_channels: number of input channels.
    :param out_channels: number of output channels.
    :param input_height: the height of the input images.
    :param input_width: the width of the input images.
    :param kernel_size: size of weight on spatial dimensions. If kernel_size is
        an :class:`int`, the actual kernel size would be
        `(kernel_size, kernel_size)`. Default: 1
    :param stride: stride of the 2D convolution operation. Default: 1
    :param padding: size of the paddings added to the input on both sides of its
        spatial dimensions. Only zero-padding is supported. Default: 0
    :param groups: number of groups into which the input and output channels are divided,
        so as to perform a "grouped convolution". When ``groups`` is not 1,
        ``in_channels`` and ``out_channels`` must be divisible by ``groups``.
        The shape of weight is `(groups, output_height, output_width,
        in_channels // groups, *kernel_size, out_channels // groups)`.
    """

    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        input_height: int,
        input_width: int,
        kernel_size: Union[int, Tuple[int, int]],
        stride: Union[int, Tuple[int, int]] = 1,
        padding: Union[int, Tuple[int, int]] = 0,
        dilation: Union[int, Tuple[int, int]] = 1,
        groups: int = 1,
        conv_mode: str = "cross_correlation",
        **kwargs
    ):
        self.input_height = input_height
        self.input_width = input_width
        super().__init__(
            in_channels,
            out_channels,
            kernel_size,
            stride,
            padding,
            dilation,
            groups,
            bias=False,
            **kwargs,
        )

    def _infer_weight_shape(self):
        group = self.groups
        output_height = (
            self.input_height + self.padding[0] * 2 - self.kernel_size[0]
        ) // self.stride[0] + 1
        output_width = (
            self.input_width + self.padding[1] * 2 - self.kernel_size[1]
        ) // self.stride[1] + 1
        # Assume format is NCHW
        return (
            group,
            output_height,
            output_width,
            self.in_channels // group,
            self.kernel_size[0],
            self.kernel_size[1],
            self.out_channels // group,
        )

    def forward(self, inp):
        return local_conv2d(
            inp,
            self.weight,
            None,
            self.stride,
            self.padding,
            self.dilation,
            self.conv_mode,
        )


class ConvRelu2d(Conv2d):
    r"""
    A fused :class:`~.Module` including :class:`~.module.Conv2d` and :func:`~.relu`.
    Could be replaced with :class:`~.QATModule` version :class:`~.qat.ConvRelu2d` using :func:`~.quantize.quantize_qat`.
    """

    def forward(self, inp):
        return relu(self.calc_conv(inp, self.weight, self.bias))


class DeformableConv2d(_ConvNd):
    """
    Deformable Convolution.

    :param in_channels: number of input channels.
    :param out_channels: number of output channels.
    :param kernel_size: size of weight on spatial dimensions. If kernel_size is
        an :class:`int`, the actual kernel size would be
        `(kernel_size, kernel_size)`. Default: 1
    :param stride: stride of the 2D convolution operation. Default: 1
    :param padding: size of the paddings added to the input on both sides of its
        spatial dimensions. Only zero-padding is supported. Default: 0
    :param dilation: dilation of the 2D convolution operation. Default: 1
    :param groups: number of groups into which the input and output channels are divided,
        so as to perform a "grouped convolution". When ``groups`` is not 1,
        ``in_channels`` and ``out_channels`` must be divisible by ``groups``,
        and there would be an extra dimension at the beginning of the weight's
        shape. Specifically, the shape of weight would be `(groups,
        out_channel // groups, in_channels // groups, *kernel_size)`.
    :param bias: whether to add a bias onto the result of convolution. Default:
        True
    :param conv_mode: Supports `cross_correlation`. Default:
        `cross_correlation`
    :param compute_mode: When set to "default", no special requirements will be
        placed on the precision of intermediate results. When set to "float32",
        "float32" would be used for accumulator and intermediate result, but only
        effective when input and output are of float16 dtype.
    """

    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        kernel_size: Union[int, Tuple[int, int]],
        stride: Union[int, Tuple[int, int]] = 1,
        padding: Union[int, Tuple[int, int]] = 0,
        dilation: Union[int, Tuple[int, int]] = 1,
        groups: int = 1,
        bias: bool = True,
        conv_mode: str = "cross_correlation",
        compute_mode: str = "default",
        **kwargs
    ):
        kernel_size = _pair_nonzero(kernel_size)
        stride = _pair_nonzero(stride)
        padding = _pair(padding)
        dilation = _pair_nonzero(dilation)
        self.conv_mode = conv_mode
        self.compute_mode = compute_mode
        super().__init__(
            in_channels,
            out_channels,
            kernel_size,
            stride,
            padding,
            dilation,
            groups,
            bias,
            **kwargs,
        )

    def _get_fanin(self):
        kh, kw = self.kernel_size
        ic = self.in_channels
        return kh * kw * ic

    def _infer_weight_shape(self):
        group = self.groups
        ichl = self.in_channels
        ochl = self.out_channels
        kh, kw = self.kernel_size
        if group == 1:
            # Assume format is NCHW
            return (ochl, ichl, kh, kw)

        assert (
            ichl % group == 0 and ochl % group == 0
        ), "invalid config: input_channels={} output_channels={} group={}".format(
            ichl, ochl, group
        )
        # Assume format is NCHW
        return (group, ochl // group, ichl // group, kh, kw)

    def _infer_bias_shape(self):
        # Assume format is NCHW
        return (1, self.out_channels, 1, 1)

    def calc_conv(self, inp, weight, offset, mask, bias):
        return deformable_conv2d(
            inp,
            weight,
            offset,
            mask,
            bias,
            self.stride,
            self.padding,
            self.dilation,
            self.groups,
            self.conv_mode,
            self.compute_mode,
        )

    def forward(self, inp, offset, mask):
        return self.calc_conv(inp, self.weight, offset, mask, self.bias)


class ConvTranspose3d(_ConvNd):
    r"""
    Applies a 3D transposed convolution over an input tensor.

    Only support the case that group = 1 and conv_mode = "cross_correlation".

    :class:`ConvTranspose3d` can be seen as the gradient of :class:`Conv3d` operation
    with respect to its input.

    Convolution3D usually reduces the size of input, while transposed convolution3d
    works the opposite way, transforming a smaller input to a larger output while
    preserving the connectivity pattern.

    :param in_channels: number of input channels.
    :param out_channels: number of output channels.
    :param kernel_size: size of weight on spatial dimensions. If ``kernel_size`` is
        an :class:`int`, the actual kernel size would be
        ``(kernel_size, kernel_size, kernel_size)``. Default: 1
    :param stride: stride of the 3D convolution operation. Default: 1
    :param padding: size of the paddings added to the input on all sides of its
        spatial dimensions. Only zero-padding is supported. Default: 0
    :param dilation: dilation of the 3D convolution operation. Default: 1
    :param bias: wether to add a bias onto the result of convolution. Default:
        True
    """

    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        kernel_size: Union[int, Tuple[int, int, int]],
        stride: Union[int, Tuple[int, int, int]] = 1,
        padding: Union[int, Tuple[int, int, int]] = 0,
        dilation: Union[int, Tuple[int, int, int]] = 1,
        bias: bool = True,
    ):
        kernel_size = _triple_nonzero(kernel_size)
        stride = _triple_nonzero(stride)
        padding = _triple(padding)
        dilation = _triple_nonzero(dilation)
        super().__init__(
            in_channels=in_channels,
            out_channels=out_channels,
            kernel_size=kernel_size,
            stride=stride,
            padding=padding,
            dilation=dilation,
            groups=1,
            bias=bias,
        )

    def _get_fanin(self):
        kt, kh, kw = self.kernel_size
        ic = self.in_channels
        return kt * kh * kw * ic

    def _infer_weight_shape(self):
        ichl = self.in_channels
        ochl = self.out_channels
        kt, kh, kw = self.kernel_size
        return (ochl, ichl, kt, kh, kw)

    def _infer_bias_shape(self):
        # Assume format is NCTHW
        return (1, self.out_channels, 1, 1, 1)

    def forward(self, inp):
        return conv_transpose3d(
            inp, self.weight, self.bias, self.stride, self.padding, self.dilation,
        )