EasyCV/easycv/models/segmentation/heads/uper_head.py

# Copyright (c) Alibaba, Inc. and its affiliates.
import warnings

import torch
import torch.nn as nn
import torch.nn.functional as F
from mmcv.cnn import ConvModule

from easycv.models.builder import HEADS
from easycv.models.utils.ops import resize_tensor
from .base import BaseDecodeHead


# Modified from https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/decode_heads/uper_head.py
@HEADS.register_module()
class UPerHead(BaseDecodeHead):
    """Unified Perceptual Parsing for Scene Understanding.

    This head is the implementation of `UPerNet
    <https://arxiv.org/abs/1807.10221>`_.

    Args:
        pool_scales (tuple[int]): Pooling scales used in Pooling Pyramid
            Module applied on the last feature. Default: (1, 2, 3, 6).
    """

    def __init__(self, pool_scales=(1, 2, 3, 6), **kwargs):
        super(UPerHead, self).__init__(
            input_transform='multiple_select', **kwargs)
        # PSP Module
        self.psp_modules = PPM(
            pool_scales,
            self.in_channels[-1],
            self.channels,
            conv_cfg=self.conv_cfg,
            norm_cfg=self.norm_cfg,
            act_cfg=self.act_cfg,
            align_corners=self.align_corners)
        self.bottleneck = ConvModule(
            self.in_channels[-1] + len(pool_scales) * self.channels,
            self.channels,
            3,
            padding=1,
            conv_cfg=self.conv_cfg,
            norm_cfg=self.norm_cfg,
            act_cfg=self.act_cfg)
        # FPN Module
        self.lateral_convs = nn.ModuleList()
        self.fpn_convs = nn.ModuleList()
        for in_channels in self.in_channels[:-1]:  # skip the top layer
            l_conv = ConvModule(
                in_channels,
                self.channels,
                1,
                conv_cfg=self.conv_cfg,
                norm_cfg=self.norm_cfg,
                act_cfg=self.act_cfg,
                inplace=False)
            fpn_conv = ConvModule(
                self.channels,
                self.channels,
                3,
                padding=1,
                conv_cfg=self.conv_cfg,
                norm_cfg=self.norm_cfg,
                act_cfg=self.act_cfg,
                inplace=False)
            self.lateral_convs.append(l_conv)
            self.fpn_convs.append(fpn_conv)

        self.fpn_bottleneck = ConvModule(
            len(self.in_channels) * self.channels,
            self.channels,
            3,
            padding=1,
            conv_cfg=self.conv_cfg,
            norm_cfg=self.norm_cfg,
            act_cfg=self.act_cfg)

    def psp_forward(self, inputs):
        """Forward function of PSP module."""
        x = inputs[-1]
        psp_outs = [x]
        psp_outs.extend(self.psp_modules(x))
        psp_outs = torch.cat(psp_outs, dim=1)
        output = self.bottleneck(psp_outs)

        return output

    def _forward_feature(self, inputs):
        """Forward function for feature maps before classifying each pixel with
        ``self.cls_seg`` fc.

        Args:
            inputs (list[Tensor]): List of multi-level img features.

        Returns:
            feats (Tensor): A tensor of shape (batch_size, self.channels,
                H, W) which is feature map for last layer of decoder head.
        """
        inputs = self._transform_inputs(inputs)

        # build laterals
        laterals = [
            lateral_conv(inputs[i])
            for i, lateral_conv in enumerate(self.lateral_convs)
        ]

        laterals.append(self.psp_forward(inputs))

        # build top-down path
        used_backbone_levels = len(laterals)
        for i in range(used_backbone_levels - 1, 0, -1):
            prev_shape = laterals[i - 1].shape[2:]
            laterals[i - 1] = laterals[i - 1] + resize_tensor(
                laterals[i],
                size=prev_shape,
                mode='bilinear',
                align_corners=self.align_corners)

        # build outputs
        fpn_outs = [
            self.fpn_convs[i](laterals[i])
            for i in range(used_backbone_levels - 1)
        ]
        # append psp feature
        fpn_outs.append(laterals[-1])

        for i in range(used_backbone_levels - 1, 0, -1):
            fpn_outs[i] = resize_tensor(
                fpn_outs[i],
                size=fpn_outs[0].shape[2:],
                mode='bilinear',
                align_corners=self.align_corners)
        fpn_outs = torch.cat(fpn_outs, dim=1)
        feats = self.fpn_bottleneck(fpn_outs)
        return feats

    def forward(self, inputs):
        """Forward function."""
        output = self._forward_feature(inputs)
        output = self.cls_seg(output)
        return output


class PPM(nn.ModuleList):
    """Pooling Pyramid Module used in PSPNet.

    Args:
        pool_scales (tuple[int]): Pooling scales used in Pooling Pyramid
            Module.
        in_channels (int): Input channels.
        channels (int): Channels after modules, before conv_seg.
        conv_cfg (dict|None): Config of conv layers.
        norm_cfg (dict|None): Config of norm layers.
        act_cfg (dict): Config of activation layers.
        align_corners (bool): align_corners argument of F.interpolate.
    """

    def __init__(self, pool_scales, in_channels, channels, conv_cfg, norm_cfg,
                 act_cfg, align_corners, **kwargs):
        super(PPM, self).__init__()
        self.pool_scales = pool_scales
        self.align_corners = align_corners
        self.in_channels = in_channels
        self.channels = channels
        self.conv_cfg = conv_cfg
        self.norm_cfg = norm_cfg
        self.act_cfg = act_cfg
        for pool_scale in pool_scales:
            self.append(
                nn.Sequential(
                    nn.AdaptiveAvgPool2d(pool_scale),
                    ConvModule(
                        self.in_channels,
                        self.channels,
                        1,
                        conv_cfg=self.conv_cfg,
                        norm_cfg=self.norm_cfg,
                        act_cfg=self.act_cfg,
                        **kwargs)))

    def forward(self, x):
        """Forward function."""
        ppm_outs = []
        for ppm in self:
            ppm_out = ppm(x)
            upsampled_ppm_out = resize_tensor(
                ppm_out,
                size=x.size()[2:],
                mode='bilinear',
                align_corners=self.align_corners)
            ppm_outs.append(upsampled_ppm_out)
        return ppm_outs