conv_module.py

import warnings

import torch.nn as nn
from mmcv.cnn import kaiming_init, constant_init

from .conv_ws import ConvWS2d
from .norm import build_norm_layer

conv_cfg = {
    'Conv': nn.Conv2d,
    'ConvWS': ConvWS2d,
    # TODO: octave conv
}


def build_conv_layer(cfg, *args, **kwargs):
    """ Build convolution layer

    Args:
        cfg (None or dict): cfg should contain:
            type (str): identify conv layer type.
            layer args: args needed to instantiate a conv layer.

    Returns:
        layer (nn.Module): created conv layer
    """
    if cfg is None:
        cfg_ = dict(type='Conv')
    else:
        assert isinstance(cfg, dict) and 'type' in cfg
        cfg_ = cfg.copy()

    layer_type = cfg_.pop('type')
    if layer_type not in conv_cfg:
        raise KeyError('Unrecognized norm type {}'.format(layer_type))
    else:
        conv_layer = conv_cfg[layer_type]

    layer = conv_layer(*args, **kwargs, **cfg_)

    return layer


class ConvModule(nn.Module):
    """Conv-Norm-Activation block.

    Args:
        in_channels (int): Same as nn.Conv2d.
        out_channels (int): Same as nn.Conv2d.
        kernel_size (int or tuple[int]): Same as nn.Conv2d.
        stride (int or tuple[int]): Same as nn.Conv2d.
        padding (int or tuple[int]): Same as nn.Conv2d.
        dilation (int or tuple[int]): Same as nn.Conv2d.
        groups (int): Same as nn.Conv2d.
        bias (bool or str): If specified as `auto`, it will be decided by the
            norm_cfg. Bias will be set as True if norm_cfg is None, otherwise
            False.
        conv_cfg (dict): Config dict for convolution layer.
        norm_cfg (dict): Config dict for normalization layer.
        activation (str or None): Activation type, "ReLU" by default.
        inplace (bool): Whether to use inplace mode for activation.
        activate_last (bool): Whether to apply the activation layer in the
            last. (Do not use this flag since the behavior and api may be
            changed in the future.)
    """

    def __init__(self,
                 in_channels,
                 out_channels,
                 kernel_size,
                 stride=1,
                 padding=0,
                 dilation=1,
                 groups=1,
                 bias='auto',
                 conv_cfg=None,
                 norm_cfg=None,
                 activation='relu',
                 inplace=True,
                 activate_last=True):
        super(ConvModule, self).__init__()
        assert conv_cfg is None or isinstance(conv_cfg, dict)
        assert norm_cfg is None or isinstance(norm_cfg, dict)
        self.conv_cfg = conv_cfg
        self.norm_cfg = norm_cfg
        self.activation = activation
        self.inplace = inplace
        self.activate_last = activate_last

        self.with_norm = norm_cfg is not None
        self.with_activatation = activation is not None
        # if the conv layer is before a norm layer, bias is unnecessary.
        if bias == 'auto':
            bias = False if self.with_norm else True
        self.with_bias = bias

        if self.with_norm and self.with_bias:
            warnings.warn('ConvModule has norm and bias at the same time')

        # build convolution layer
        self.conv = build_conv_layer(
            conv_cfg,
            in_channels,
            out_channels,
            kernel_size,
            stride=stride,
            padding=padding,
            dilation=dilation,
            groups=groups,
            bias=bias)
        # export the attributes of self.conv to a higher level for convenience
        self.in_channels = self.conv.in_channels
        self.out_channels = self.conv.out_channels
        self.kernel_size = self.conv.kernel_size
        self.stride = self.conv.stride
        self.padding = self.conv.padding
        self.dilation = self.conv.dilation
        self.transposed = self.conv.transposed
        self.output_padding = self.conv.output_padding
        self.groups = self.conv.groups

        # build normalization layers
        if self.with_norm:
            norm_channels = out_channels if self.activate_last else in_channels
            self.norm_name, norm = build_norm_layer(norm_cfg, norm_channels)
            self.add_module(self.norm_name, norm)

        # build activation layer
        if self.with_activatation:
            if self.activation not in ['relu']:
                raise ValueError('{} is currently not supported.'.format(
                    self.activation))
            if self.activation == 'relu':
                self.activate = nn.ReLU(inplace=inplace)

        # Use msra init by default
        self.init_weights()

    @property
    def norm(self):
        return getattr(self, self.norm_name)

    def init_weights(self):
        nonlinearity = 'relu' if self.activation is None else self.activation
        kaiming_init(self.conv, nonlinearity=nonlinearity)
        if self.with_norm:
            constant_init(self.norm, 1, bias=0)

    def forward(self, x, activate=True, norm=True):
        if self.activate_last:
            x = self.conv(x)
            if norm and self.with_norm:
                x = self.norm(x)
            if activate and self.with_activatation:
                x = self.activate(x)
        else:
            # WARN: this may be removed or modified
            if norm and self.with_norm:
                x = self.norm(x)
            if activate and self.with_activatation:
                x = self.activate(x)
            x = self.conv(x)
        return x