models/utils/residual.py

# -*- coding: utf-8 -*-
# Residual block as defined in:
# He, Kaiming, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. "Deep residual learning
# for image recognition." In Proceedings of the IEEE conference on computer vision
# and pattern recognition, pp. 770-778. 2016.
#
# Code Snippet adapted from HoverNet implementation (https://github.com/vqdang/hover_net)
#
# @ Fabian Hörst, [email protected]
# Institute for Artifical Intelligence in Medicine,
# University Medicine Essen


import torch
import torch.nn as nn

from collections import OrderedDict

from models.utils.tf_utils import TFSamepaddingLayer


class ResidualBlock(nn.Module):
    """Residual block as defined in:

    He, Kaiming, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. "Deep residual learning
    for image recognition." In Proceedings of the IEEE conference on computer vision
    and pattern recognition, pp. 770-778. 2016.

    """

    def __init__(self, in_ch, unit_ksize, unit_ch, unit_count, stride=1):
        super(ResidualBlock, self).__init__()
        assert len(unit_ksize) == len(unit_ch), "Unbalance Unit Info"

        self.nr_unit = unit_count
        self.in_ch = in_ch
        self.unit_ch = unit_ch

        # ! For inference only so init values for batchnorm may not match tensorflow
        unit_in_ch = in_ch
        self.units = nn.ModuleList()
        for idx in range(unit_count):
            unit_layer = [
                ("preact/bn", nn.BatchNorm2d(unit_in_ch, eps=1e-5)),
                ("preact/relu", nn.ReLU(inplace=True)),
                (
                    "conv1",
                    nn.Conv2d(
                        unit_in_ch,
                        unit_ch[0],
                        unit_ksize[0],
                        stride=1,
                        padding=0,
                        bias=False,
                    ),
                ),
                ("conv1/bn", nn.BatchNorm2d(unit_ch[0], eps=1e-5)),
                ("conv1/relu", nn.ReLU(inplace=True)),
                (
                    "conv2/pad",
                    TFSamepaddingLayer(
                        ksize=unit_ksize[1], stride=stride if idx == 0 else 1
                    ),
                ),
                (
                    "conv2",
                    nn.Conv2d(
                        unit_ch[0],
                        unit_ch[1],
                        unit_ksize[1],
                        stride=stride if idx == 0 else 1,
                        padding=0,
                        bias=False,
                    ),
                ),
                ("conv2/bn", nn.BatchNorm2d(unit_ch[1], eps=1e-5)),
                ("conv2/relu", nn.ReLU(inplace=True)),
                (
                    "conv3",
                    nn.Conv2d(
                        unit_ch[1],
                        unit_ch[2],
                        unit_ksize[2],
                        stride=1,
                        padding=0,
                        bias=False,
                    ),
                ),
            ]
            # * has bna to conclude each previous block so
            # * must not put preact for the first unit of this block
            unit_layer = unit_layer if idx != 0 else unit_layer[2:]
            self.units.append(nn.Sequential(OrderedDict(unit_layer)))
            unit_in_ch = unit_ch[-1]

        if in_ch != unit_ch[-1] or stride != 1:
            self.shortcut = nn.Conv2d(in_ch, unit_ch[-1], 1, stride=stride, bias=False)
        else:
            self.shortcut = None

        self.blk_bna = nn.Sequential(
            OrderedDict(
                [
                    ("bn", nn.BatchNorm2d(unit_in_ch, eps=1e-5)),
                    ("relu", nn.ReLU(inplace=True)),
                ]
            )
        )

    def out_ch(self):
        return self.unit_ch[-1]

    def init_weights(self):
        """Kaiming (HE) initialization for convolutional layers and constant initialization for normalization and linear layers"""
        for m in self.modules():
            classname = m.__class__.__name__

            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")

            if "norm" in classname.lower():
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)

            if "linear" in classname.lower():
                if m.bias is not None:
                    nn.init.constant_(m.bias, 0)

    def forward(self, prev_feat, freeze=False):
        if self.shortcut is None:
            shortcut = prev_feat
        else:
            shortcut = self.shortcut(prev_feat)

        for idx in range(0, len(self.units)):
            new_feat = prev_feat
            if self.training:
                with torch.set_grad_enabled(not freeze):
                    new_feat = self.units[idx](new_feat)
            else:
                new_feat = self.units[idx](new_feat)
            prev_feat = new_feat + shortcut
            shortcut = prev_feat
        feat = self.blk_bna(prev_feat)
        return feat