backbone.py

"""
This file contains our implementation of ReResNet.
@author: Jiaming Han
"""
import e2cnn.nn as enn
import math
import os
import torch
import torch.nn as nn
import torch.utils.checkpoint as cp
from e2cnn import gspaces
from mmcv.cnn import (constant_init, kaiming_init)
from torch.nn.modules.batchnorm import _BatchNorm
import torch.nn.functional as F

# Set default Orientation=8, .i.e, the group C8
# One can change it by passing the env Orientation=xx
Orientation = 8
# keep similar computation or similar params
# One can change it by passing the env fixparams=True
fixparams = False
if 'Orientation' in os.environ:
    Orientation = int(os.environ['Orientation'])
if 'fixparams' in os.environ:
    fixparams = True
print('ReResNet Orientation: {}\tFix Params: {}'.format(Orientation, fixparams))

# define the equivariant group. We use C8 group by default.
gspace = gspaces.Rot2dOnR2(N=Orientation)


def regular_feature_type(gspace: gspaces.GSpace, planes: int):
    """ build a regular feature map with the specified number of channels"""
    assert gspace.fibergroup.order() > 0
    N = gspace.fibergroup.order()
    if fixparams:
        planes *= math.sqrt(N)
    planes = planes / N
    planes = int(planes)
    return enn.FieldType(gspace, [gspace.regular_repr] * planes)


def trivial_feature_type(gspace: gspaces.GSpace, planes: int, fixparams: bool = True):
    """ build a trivial feature map with the specified number of channels"""
    if fixparams:
        planes *= math.sqrt(gspace.fibergroup.order())
    planes = int(planes)
    return enn.FieldType(gspace, [gspace.trivial_repr] * planes)


FIELD_TYPE = {
    "trivial": trivial_feature_type,
    "regular": regular_feature_type,
}


def conv7x7(inplanes, out_planes, stride=2, padding=3, bias=False):
    """7x7 convolution with padding"""
    in_type = enn.FieldType(gspace, inplanes * [gspace.trivial_repr])
    out_type = FIELD_TYPE['regular'](gspace, out_planes)
    return enn.R2Conv(in_type, out_type, 7,
                      stride=stride,
                      padding=padding,
                      bias=bias,
                      sigma=None,
                      frequencies_cutoff=lambda r: 3 * r, )


def conv3x3(inplanes, out_planes, stride=1, padding=1, groups=1, dilation=1):
    """3x3 convolution with padding"""
    in_type = FIELD_TYPE['regular'](gspace, inplanes)
    out_type = FIELD_TYPE['regular'](gspace, out_planes)
    return enn.R2Conv(in_type, out_type, 3,
                      stride=stride,
                      padding=padding,
                      groups=groups,
                      bias=False,
                      dilation=dilation,
                      sigma=None,
                      frequencies_cutoff=lambda r: 3 * r,
                      initialize=False)


def conv1x1(inplanes, out_planes, stride=1):
    """1x1 convolution"""
    in_type = FIELD_TYPE['regular'](gspace, inplanes)
    out_type = FIELD_TYPE['regular'](gspace, out_planes)
    return enn.R2Conv(in_type, out_type, 1,
                      stride=stride,
                      bias=False,
                      sigma=None,
                      frequencies_cutoff=lambda r: 3 * r,
                      initialize=False)


def convnxn(inplanes, outplanes, kernel_size=3, stride=1, padding=0, groups=1, bias=False, dilation=1):
    in_type = FIELD_TYPE['regular'](gspace, inplanes)
    out_type = FIELD_TYPE['regular'](gspace, outplanes)
    return enn.R2Conv(in_type, out_type, kernel_size,
                      stride=stride,
                      padding=padding,
                      groups=groups,
                      bias=bias,
                      dilation=dilation,
                      sigma=None,
                      frequencies_cutoff=lambda r: 3 * r, )


def ennReLU(inplanes):
    in_type = FIELD_TYPE['regular'](gspace, inplanes)
    return enn.ReLU(in_type, inplace=True)


def ennAvgPool(inplanes, kernel_size=1, stride=None, padding=0, ceil_mode=False):
    in_type = FIELD_TYPE['regular'](gspace, inplanes)
    return enn.PointwiseAvgPool(in_type, kernel_size, stride=stride, padding=padding, ceil_mode=ceil_mode)


def ennMaxPool(inplanes, kernel_size, stride=1, padding=0):
    in_type = FIELD_TYPE['regular'](gspace, inplanes)
    return enn.PointwiseMaxPool(in_type, kernel_size=kernel_size, stride=stride, padding=padding)


def build_conv_layer(cfg, *args, **kwargs):
    layer = convnxn(*args, **kwargs)
    return layer


def build_norm_layer(cfg, num_features, postfix=''):
    in_type = FIELD_TYPE['regular'](gspace, num_features)
    return 'bn' + str(postfix), enn.InnerBatchNorm(in_type)


class BasicBlock(enn.EquivariantModule):
    """BasicBlock for ReResNet.
    Args:
        in_channels (int): Input channels of this block.
        out_channels (int): Output channels of this block.
        expansion (int): The ratio of ``out_channels/mid_channels`` where
            ``mid_channels`` is the output channels of conv1. This is a
            reserved argument in BasicBlock and should always be 1. Default: 1.
        stride (int): stride of the block. Default: 1
        dilation (int): dilation of convolution. Default: 1
        downsample (nn.Module): downsample operation on identity branch.
            Default: None.
        style (str): `pytorch` or `caffe`. It is unused and reserved for
            unified API with Bottleneck.
        with_cp (bool): Use checkpoint or not. Using checkpoint will save some
            memory while slowing down the training speed.
        conv_cfg (dict): dictionary to construct and config conv layer.
            Default: None
        norm_cfg (dict): dictionary to construct and config norm layer.
            Default: dict(type='BN')
    """

    def __init__(self,
                 in_channels,
                 out_channels,
                 expansion=1,
                 stride=1,
                 dilation=1,
                 downsample=None,
                 style='pytorch',
                 with_cp=False,
                 conv_cfg=None,
                 norm_cfg=dict(type='BN')):
        super(BasicBlock, self).__init__()
        self.in_type = FIELD_TYPE['regular'](gspace, in_channels)
        self.out_type = FIELD_TYPE['regular'](gspace, out_channels)
        self.in_channels = in_channels
        self.out_channels = out_channels
        self.expansion = expansion
        assert self.expansion == 1
        assert out_channels % expansion == 0
        self.mid_channels = out_channels // expansion
        self.stride = stride
        self.dilation = dilation
        self.style = style
        self.with_cp = with_cp
        self.conv_cfg = conv_cfg
        self.norm_cfg = norm_cfg

        self.norm1_name, norm1 = build_norm_layer(
            norm_cfg, self.mid_channels, postfix=1)
        self.norm2_name, norm2 = build_norm_layer(
            norm_cfg, out_channels, postfix=2)

        self.conv1 = build_conv_layer(
            conv_cfg,
            in_channels,
            self.mid_channels,
            3,
            stride=stride,
            padding=dilation,
            dilation=dilation,
            bias=False)
        self.add_module(self.norm1_name, norm1)
        self.relu1 = ennReLU(self.mid_channels)
        self.conv2 = build_conv_layer(
            conv_cfg,
            self.mid_channels,
            out_channels,
            3,
            padding=1,
            bias=False)
        self.add_module(self.norm2_name, norm2)

        self.relu2 = ennReLU(out_channels)
        self.downsample = downsample

    @property
    def norm1(self):
        return getattr(self, self.norm1_name)

    @property
    def norm2(self):
        return getattr(self, self.norm2_name)

    def forward(self, x):

        def _inner_forward(x):
            identity = x

            out = self.conv1(x)
            out = self.norm1(out)
            out = self.relu1(out)

            out = self.conv2(out)
            out = self.norm2(out)

            if self.downsample is not None:
                identity = self.downsample(x)

            out += identity

            return out

        if self.with_cp and x.requires_grad:
            out = cp.checkpoint(_inner_forward, x)
        else:
            out = _inner_forward(x)

        out = self.relu2(out)

        return out

    def evaluate_output_shape(self, input_shape):
        assert len(input_shape) == 4
        assert input_shape[1] == self.in_type.size
        if self.downsample is not None:
            return self.downsample.evaluate_output_shape(input_shape)
        else:
            return input_shape


class Bottleneck(enn.EquivariantModule):
    """Bottleneck block for ReResNet.
    Args:
        in_channels (int): Input channels of this block.
        out_channels (int): Output channels of this block.
        expansion (int): The ratio of ``out_channels/mid_channels`` where
            ``mid_channels`` is the input/output channels of conv2. Default: 4.
        stride (int): stride of the block. Default: 1
        dilation (int): dilation of convolution. Default: 1
        downsample (nn.Module): downsample operation on identity branch.
            Default: None.
        style (str): ``"pytorch"`` or ``"caffe"``. If set to "pytorch", the
            stride-two layer is the 3x3 conv layer, otherwise the stride-two
            layer is the first 1x1 conv layer. Default: "pytorch".
        with_cp (bool): Use checkpoint or not. Using checkpoint will save some
            memory while slowing down the training speed.
        conv_cfg (dict): dictionary to construct and config conv layer.
            Default: None
        norm_cfg (dict): dictionary to construct and config norm layer.
            Default: dict(type='BN')
    """

    def __init__(self,
                 in_channels,
                 out_channels,
                 expansion=4,
                 stride=1,
                 dilation=1,
                 downsample=None,
                 style='pytorch',
                 with_cp=False,
                 conv_cfg=None,
                 norm_cfg=dict(type='BN')):
        super(Bottleneck, self).__init__()
        assert style in ['pytorch', 'caffe']
        self.in_type = FIELD_TYPE['regular'](gspace, in_channels)
        self.out_type = FIELD_TYPE['regular'](gspace, out_channels)
        self.in_channels = in_channels
        self.out_channels = out_channels
        self.expansion = expansion
        assert out_channels % expansion == 0
        self.mid_channels = out_channels // expansion
        self.stride = stride
        self.dilation = dilation
        self.style = style
        self.with_cp = with_cp
        self.conv_cfg = conv_cfg
        self.norm_cfg = norm_cfg
        if self.style == 'pytorch':
            self.conv1_stride = 1
            self.conv2_stride = stride
        else:
            self.conv1_stride = stride
            self.conv2_stride = 1

        self.norm1_name, norm1 = build_norm_layer(
            norm_cfg, self.mid_channels, postfix=1)
        self.norm2_name, norm2 = build_norm_layer(
            norm_cfg, self.mid_channels, postfix=2)
        self.norm3_name, norm3 = build_norm_layer(
            norm_cfg, out_channels, postfix=3)

        self.conv1 = build_conv_layer(
            conv_cfg,
            in_channels,
            self.mid_channels,
            kernel_size=1,
            stride=self.conv1_stride,
            bias=False)
        self.add_module(self.norm1_name, norm1)
        self.relu1 = ennReLU(self.mid_channels)
        self.conv2 = build_conv_layer(
            conv_cfg,
            self.mid_channels,
            self.mid_channels,
            kernel_size=3,
            stride=self.conv2_stride,
            padding=dilation,
            dilation=dilation,
            bias=False)

        self.add_module(self.norm2_name, norm2)
        self.relu2 = ennReLU(self.mid_channels)
        self.conv3 = build_conv_layer(
            conv_cfg,
            self.mid_channels,
            out_channels,
            kernel_size=1,
            bias=False)
        self.add_module(self.norm3_name, norm3)
        self.relu3 = ennReLU(out_channels)

        self.downsample = downsample

    @property
    def norm1(self):
        return getattr(self, self.norm1_name)

    @property
    def norm2(self):
        return getattr(self, self.norm2_name)

    @property
    def norm3(self):
        return getattr(self, self.norm3_name)

    def forward(self, x):

        def _inner_forward(x):
            identity = x

            out = self.conv1(x)
            out = self.norm1(out)
            out = self.relu1(out)

            out = self.conv2(out)
            out = self.norm2(out)
            out = self.relu2(out)

            out = self.conv3(out)
            out = self.norm3(out)

            if self.downsample is not None:
                identity = self.downsample(x)

            out += identity

            return out

        if self.with_cp and x.requires_grad:
            out = cp.checkpoint(_inner_forward, x)
        else:
            out = _inner_forward(x)

        out = self.relu3(out)

        return out

    def evaluate_output_shape(self, input_shape):
        assert len(input_shape) == 4
        assert input_shape[1] == self.in_type.size
        if self.downsample is not None:
            return self.downsample.evaluate_output_shape(input_shape)
        else:
            return input_shape


def get_expansion(block, expansion=None):
    """Get the expansion of a residual block.
    The block expansion will be obtained by the following order:
    1. If ``expansion`` is given, just return it.
    2. If ``block`` has the attribute ``expansion``, then return
       ``block.expansion``.
    3. Return the default value according the the block type:
       1 for ``BasicBlock`` and 4 for ``Bottleneck``.
    Args:
        block (class): The block class.
        expansion (int | None): The given expansion ratio.
    Returns:
        int: The expansion of the block.
    """
    if isinstance(expansion, int):
        assert expansion > 0
    elif expansion is None:
        if hasattr(block, 'expansion'):
            expansion = block.expansion
        elif issubclass(block, BasicBlock):
            expansion = 1
        elif issubclass(block, Bottleneck):
            expansion = 4
        else:
            raise TypeError(f'expansion is not specified for {block.__name__}')
    else:
        raise TypeError('expansion must be an integer or None')

    return expansion


class ResLayer(nn.Sequential):
    """ResLayer to build ReResNet style backbone.
    Args:
        block (nn.Module): Residual block used to build ResLayer.
        num_blocks (int): Number of blocks.
        in_channels (int): Input channels of this block.
        out_channels (int): Output channels of this block.
        expansion (int, optional): The expansion for BasicBlock/Bottleneck.
            If not specified, it will firstly be obtained via
            ``block.expansion``. If the block has no attribute "expansion",
            the following default values will be used: 1 for BasicBlock and
            4 for Bottleneck. Default: None.
        stride (int): stride of the first block. Default: 1.
        avg_down (bool): Use AvgPool instead of stride conv when
            downsampling in the bottleneck. Default: False
        conv_cfg (dict): dictionary to construct and config conv layer.
            Default: None
        norm_cfg (dict): dictionary to construct and config norm layer.
            Default: dict(type='BN')
    """

    def __init__(self,
                 block,
                 num_blocks,
                 in_channels,
                 out_channels,
                 expansion=None,
                 stride=1,
                 avg_down=False,
                 conv_cfg=None,
                 norm_cfg=dict(type='BN'),
                 **kwargs):
        self.block = block
        self.expansion = get_expansion(block, expansion)

        downsample = None
        if stride != 1 or in_channels != out_channels:
            downsample = []
            conv_stride = stride
            if avg_down and stride != 1:
                conv_stride = 1
                downsample.append(
                    ennAvgPool(
                        in_channels,
                        kernel_size=stride,
                        stride=stride,
                        ceil_mode=True))
            downsample.extend([
                build_conv_layer(
                    conv_cfg,
                    in_channels,
                    out_channels,
                    kernel_size=1,
                    stride=conv_stride,
                    bias=False),
                build_norm_layer(norm_cfg, out_channels)[1]
            ])
            downsample = enn.SequentialModule(*downsample)

        layers = []
        layers.append(
            block(
                in_channels=in_channels,
                out_channels=out_channels,
                expansion=self.expansion,
                stride=stride,
                downsample=downsample,
                conv_cfg=conv_cfg,
                norm_cfg=norm_cfg,
                **kwargs))
        in_channels = out_channels
        for i in range(1, num_blocks):
            layers.append(
                block(
                    in_channels=in_channels,
                    out_channels=out_channels,
                    expansion=self.expansion,
                    stride=1,
                    conv_cfg=conv_cfg,
                    norm_cfg=norm_cfg,
                    **kwargs))
        super(ResLayer, self).__init__(*layers)


class ReResNet(nn.Module):
    """ReResNet backbone.
    Please refer to the `paper <https://arxiv.org/abs/1512.03385>`_ for
    details.
    Args:
        depth (int): Network depth, from {18, 34, 50, 101, 152}.
        in_channels (int): Number of input image channels. Default: 3.
        stem_channels (int): Output channels of the stem layer. Default: 64.
        base_channels (int): Middle channels of the first stage. Default: 64.
        num_stages (int): Stages of the network. Default: 4.
        strides (Sequence[int]): Strides of the first block of each stage.
            Default: ``(1, 2, 2, 2)``.
        dilations (Sequence[int]): Dilation of each stage.
            Default: ``(1, 1, 1, 1)``.
        out_indices (Sequence[int]): Output from which stages. If only one
            stage is specified, a single tensor (feature map) is returned,
            otherwise multiple stages are specified, a tuple of tensors will
            be returned. Default: ``(3, )``.
        style (str): `pytorch` or `caffe`. If set to "pytorch", the stride-two
            layer is the 3x3 conv layer, otherwise the stride-two layer is
            the first 1x1 conv layer.
        deep_stem (bool): Replace 7x7 conv in input stem with 3 3x3 conv.
            Default: False.
        avg_down (bool): Use AvgPool instead of stride conv when
            downsampling in the bottleneck. Default: False.
        frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
            -1 means not freezing any parameters. Default: -1.
        conv_cfg (dict | None): The config dict for conv layers. Default: None.
        norm_cfg (dict): The config dict for norm layers.
        norm_eval (bool): Whether to set norm layers to eval mode, namely,
            freeze running stats (mean and var). Note: Effect on Batch Norm
            and its variants only. Default: False.
        with_cp (bool): Use checkpoint or not. Using checkpoint will save some
            memory while slowing down the training speed. Default: False.
        zero_init_residual (bool): Whether to use zero init for last norm layer
            in resblocks to let them behave as identity. Default: True.
    Example:
        >>> from mmcls.models import ReResNet
        >>> import torch
        >>> self = ReResNet(depth=18)
        >>> self.eval()
        >>> inputs = torch.rand(1, 3, 32, 32)
        >>> level_outputs = self.forward(inputs)
        >>> for level_out in level_outputs:
        ...     print(tuple(level_out.shape))
        (1, 64, 8, 8)
        (1, 128, 4, 4)
        (1, 256, 2, 2)
        (1, 512, 1, 1)
    """

    arch_settings = {
        18: (BasicBlock, (2, 2, 2, 2)),
        34: (BasicBlock, (3, 4, 6, 3)),
        50: (Bottleneck, (3, 4, 6, 3)),
        101: (Bottleneck, (3, 4, 23, 3)),
        152: (Bottleneck, (3, 8, 36, 3))
    }

    def __init__(self,
                 depth,
                 in_channels=3,
                 stem_channels=64,
                 base_channels=64,
                 expansion=None,
                 num_stages=4,
                 strides=(1, 2, 2, 2),
                 dilations=(1, 1, 1, 1),
                 out_indices=(3,),
                 style='pytorch',
                 deep_stem=False,
                 avg_down=False,
                 frozen_stages=-1,
                 conv_cfg=None,
                 norm_cfg=dict(type='BN', requires_grad=True),
                 norm_eval=False,
                 with_cp=False,
                 zero_init_residual=True):
        super(ReResNet, self).__init__()
        self.in_type = enn.FieldType(gspace, 3 * [gspace.trivial_repr])
        if depth not in self.arch_settings:
            raise KeyError(f'invalid depth {depth} for resnet')
        self.depth = depth
        self.stem_channels = stem_channels
        self.base_channels = base_channels
        self.num_stages = num_stages
        assert num_stages >= 1 and num_stages <= 4
        self.strides = strides
        self.dilations = dilations
        assert len(strides) == len(dilations) == num_stages
        self.out_indices = out_indices
        assert max(out_indices) < num_stages
        self.style = style
        self.deep_stem = deep_stem
        self.avg_down = avg_down
        self.frozen_stages = frozen_stages
        self.conv_cfg = conv_cfg
        self.norm_cfg = norm_cfg
        self.with_cp = with_cp
        self.norm_eval = norm_eval
        self.zero_init_residual = zero_init_residual

        self.block, stage_blocks = self.arch_settings[depth]
        self.stage_blocks = stage_blocks[:num_stages]
        self.expansion = get_expansion(self.block, expansion)

        self._make_stem_layer(in_channels, stem_channels)

        self.res_layers = []
        _in_channels = stem_channels
        _out_channels = base_channels * self.expansion
        for i, num_blocks in enumerate(self.stage_blocks):
            stride = strides[i]
            dilation = dilations[i]
            res_layer = self.make_res_layer(
                block=self.block,
                num_blocks=num_blocks,
                in_channels=_in_channels,
                out_channels=_out_channels,
                expansion=self.expansion,
                stride=stride,
                dilation=dilation,
                style=self.style,
                avg_down=self.avg_down,
                with_cp=with_cp,
                conv_cfg=conv_cfg,
                norm_cfg=norm_cfg)
            _in_channels = _out_channels
            _out_channels *= 2
            layer_name = f'layer{i + 1}'
            self.add_module(layer_name, res_layer)
            self.res_layers.append(layer_name)

        #self._freeze_stages()

        self.feat_dim = res_layer[-1].out_channels

        # add global average pooling
        # add global average pooling
        self.mp = enn.GroupPooling(res_layer[-1].out_type)
        #print("res_layer[-1].out_type")
        #self.gap = nn.AdaptiveAvgPool2d(1)
        #self.gap_pointwise = enn.PointwiseAdaptiveAvgPool(res_layer[-1].out_type,(1))

    def make_res_layer(self, **kwargs):
        return ResLayer(**kwargs)

    @property
    def norm1(self):
        return getattr(self, self.norm1_name)

    def _make_stem_layer(self, in_channels, stem_channels):
        if not self.deep_stem:
            self.conv1 = conv7x7(in_channels, stem_channels)
            self.norm1_name, norm1 = build_norm_layer(
                self.norm_cfg, stem_channels, postfix=1)
            self.add_module(self.norm1_name, norm1)
            self.relu = ennReLU(stem_channels)
        self.maxpool = ennMaxPool(stem_channels, kernel_size=3, stride=2, padding=1)

    def _freeze_stages(self):
        if self.frozen_stages >= 0:
            if not self.deep_stem:
                self.norm1.eval()
                for m in [self.conv1, self.norm1]:
                    for param in m.parameters():
                        param.requires_grad = True

        for i in range(1, self.frozen_stages + 1):
            m = getattr(self, f'layer{i}')
            m.eval()
            for param in m.parameters():
                param.requires_grad = True

    def init_weights(self, pretrained=None):
        super(ReResNet, self).init_weights(pretrained)
        if pretrained is None:
            for m in self.modules():
                if isinstance(m, nn.Conv2d):
                    kaiming_init(m)
                elif isinstance(m, (_BatchNorm, nn.GroupNorm)):
                    constant_init(m, 1)

    def forward(self, x):
        if not self.deep_stem:
            x = enn.GeometricTensor(x, self.in_type)
            x = self.conv1(x)
            x = self.norm1(x)
            x = self.relu(x)
        x = self.maxpool(x)

        for i, layer_name in enumerate(self.res_layers):
            res_layer = getattr(self, layer_name)
            x = res_layer(x)
       
        #x = self.mp(x)
        #x = self.gap(x.tensor)
        #x = self.gap_pointwise(x)
        #x = x.tensor
        #x = x.view(x.size(0), -1)
        #x =  F.normalize(x, p=2, dim=1)
        return x.tensor
        
        

    def train(self, mode=True):
        super(ReResNet, self).train(mode)
        #self._freeze_stages()
        if mode and self.norm_eval:
            for m in self.modules():
                # trick: eval have effect on BatchNorm only
                if isinstance(m, _BatchNorm):
                    m.eval()