resnest.py

Go to the documentation of this file.

# encoding: utf-8
# based on:
# https://github.com/zhanghang1989/ResNeSt/blob/master/resnest/torch/resnest.py
"""ResNeSt models"""
 
import logging
import math
 
import torch
from torch import nn
 
from fastreid.layers import (
    IBN,
    Non_local,
    SplAtConv2d,
    get_norm,
)
from fastreid.utils.checkpoint import get_unexpected_parameters_message, get_missing_parameters_message
from .build import BACKBONE_REGISTRY
 
logger = logging.getLogger(__name__)
_url_format = 'https://s3.us-west-1.wasabisys.com/resnest/torch/{}-{}.pth'
 
_model_sha256 = {name: checksum for checksum, name in [
    ('528c19ca', 'resnest50'),
    ('22405ba7', 'resnest101'),
    ('75117900', 'resnest200'),
    ('0cc87c48', 'resnest269'),
]}
 
 
def short_hash(name):
    if name not in _model_sha256:
        raise ValueError('Pretrained model for {name} is not available.'.format(name=name))
    return _model_sha256[name][:8]
 
 
model_urls = {name: _url_format.format(name, short_hash(name)) for
              name in _model_sha256.keys()
              }
 
 
class Bottleneck(nn.Module):
    """ResNet Bottleneck
    """
    # pylint: disable=unused-argument
    expansion = 4
 
    def __init__(self, inplanes, planes, bn_norm, with_ibn=False, stride=1, downsample=None,
                 radix=1, cardinality=1, bottleneck_width=64,
                 avd=False, avd_first=False, dilation=1, is_first=False,
                 rectified_conv=False, rectify_avg=False,
                 dropblock_prob=0.0, last_gamma=False):
        super(Bottleneck, self).__init__()
        group_width = int(planes * (bottleneck_width / 64.)) * cardinality
        self.conv1 = nn.Conv2d(inplanes, group_width, kernel_size=1, bias=False)
        if with_ibn:
            self.bn1 = IBN(group_width, bn_norm)
        else:
            self.bn1 = get_norm(bn_norm, group_width)
        self.dropblock_prob = dropblock_prob
        self.radix = radix
        self.avd = avd and (stride > 1 or is_first)
        self.avd_first = avd_first
 
        if self.avd:
            self.avd_layer = nn.AvgPool2d(3, stride, padding=1)
            stride = 1
 
        if radix > 1:
            self.conv2 = SplAtConv2d(
                group_width, group_width, kernel_size=3,
                stride=stride, padding=dilation,
                dilation=dilation, groups=cardinality, bias=False,
                radix=radix, rectify=rectified_conv,
                rectify_avg=rectify_avg,
                norm_layer=bn_norm,
                dropblock_prob=dropblock_prob)
        elif rectified_conv:
            from rfconv import RFConv2d
            self.conv2 = RFConv2d(
                group_width, group_width, kernel_size=3, stride=stride,
                padding=dilation, dilation=dilation,
                groups=cardinality, bias=False,
                average_mode=rectify_avg)
            self.bn2 = get_norm(bn_norm, group_width)
        else:
            self.conv2 = nn.Conv2d(
                group_width, group_width, kernel_size=3, stride=stride,
                padding=dilation, dilation=dilation,
                groups=cardinality, bias=False)
            self.bn2 = get_norm(bn_norm, group_width)
 
        self.conv3 = nn.Conv2d(
            group_width, planes * 4, kernel_size=1, bias=False)
        self.bn3 = get_norm(bn_norm, planes * 4)
 
        if last_gamma:
            from torch.nn.init import zeros_
            zeros_(self.bn3.weight)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        self.dilation = dilation
        self.stride = stride
 
    def forward(self, x):
        residual = x
 
        out = self.conv1(x)
        out = self.bn1(out)
        if self.dropblock_prob > 0.0:
            out = self.dropblock1(out)
        out = self.relu(out)
 
        if self.avd and self.avd_first:
            out = self.avd_layer(out)
 
        out = self.conv2(out)
        if self.radix == 1:
            out = self.bn2(out)
            if self.dropblock_prob > 0.0:
                out = self.dropblock2(out)
            out = self.relu(out)
 
        if self.avd and not self.avd_first:
            out = self.avd_layer(out)
 
        out = self.conv3(out)
        out = self.bn3(out)
        if self.dropblock_prob > 0.0:
            out = self.dropblock3(out)
 
        if self.downsample is not None:
            residual = self.downsample(x)
 
        out += residual
        out = self.relu(out)
 
        return out
 
 
class ResNest(nn.Module):
    """ResNet Variants ResNest
    Parameters
    ----------
    block : Block
        Class for the residual block. Options are BasicBlockV1, BottleneckV1.
    layers : list of int
        Numbers of layers in each block
    classes : int, default 1000
        Number of classification classes.
    dilated : bool, default False
        Applying dilation strategy to pretrained ResNet yielding a stride-8 model,
        typically used in Semantic Segmentation.
    norm_layer : object
        Normalization layer used in backbone network (default: :class:`mxnet.gluon.nn.BatchNorm`;
        for Synchronized Cross-GPU BachNormalization).
    Reference:
        - He, Kaiming, et al. "Deep residual learning for image recognition." Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.
        - Yu, Fisher, and Vladlen Koltun. "Multi-scale context aggregation by dilated convolutions."
    """
 
    # pylint: disable=unused-variable
    def __init__(self, last_stride, bn_norm, with_ibn, with_nl, block, layers, non_layers, radix=1,
                 groups=1,
                 bottleneck_width=64,
                 dilated=False, dilation=1,
                 deep_stem=False, stem_width=64, avg_down=False,
                 rectified_conv=False, rectify_avg=False,
                 avd=False, avd_first=False,
                 final_drop=0.0, dropblock_prob=0,
                 last_gamma=False):
        self.cardinality = groups
        self.bottleneck_width = bottleneck_width
        # ResNet-D params
        self.inplanes = stem_width * 2 if deep_stem else 64
        self.avg_down = avg_down
        self.last_gamma = last_gamma
        # ResNeSt params
        self.radix = radix
        self.avd = avd
        self.avd_first = avd_first
 
        super().__init__()
        self.rectified_conv = rectified_conv
        self.rectify_avg = rectify_avg
        if rectified_conv:
            from rfconv import RFConv2d
            conv_layer = RFConv2d
        else:
            conv_layer = nn.Conv2d
        conv_kwargs = {'average_mode': rectify_avg} if rectified_conv else {}
        if deep_stem:
            self.conv1 = nn.Sequential(
                conv_layer(3, stem_width, kernel_size=3, stride=2, padding=1, bias=False, **conv_kwargs),
                get_norm(bn_norm, stem_width),
                nn.ReLU(inplace=True),
                conv_layer(stem_width, stem_width, kernel_size=3, stride=1, padding=1, bias=False, **conv_kwargs),
                get_norm(bn_norm, stem_width),
                nn.ReLU(inplace=True),
                conv_layer(stem_width, stem_width * 2, kernel_size=3, stride=1, padding=1, bias=False, **conv_kwargs),
            )
        else:
            self.conv1 = conv_layer(3, 64, kernel_size=7, stride=2, padding=3,
                                    bias=False, **conv_kwargs)
        self.bn1 = get_norm(bn_norm, self.inplanes)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        self.layer1 = self._make_layer(block, 64, layers[0], 1, bn_norm, with_ibn=with_ibn, is_first=False)
        self.layer2 = self._make_layer(block, 128, layers[1], 2, bn_norm, with_ibn=with_ibn)
        if dilated or dilation == 4:
            self.layer3 = self._make_layer(block, 256, layers[2], 1, bn_norm, with_ibn=with_ibn,
                                           dilation=2, dropblock_prob=dropblock_prob)
            self.layer4 = self._make_layer(block, 512, layers[3], 1, bn_norm, with_ibn=with_ibn,
                                           dilation=4, dropblock_prob=dropblock_prob)
        elif dilation == 2:
            self.layer3 = self._make_layer(block, 256, layers[2], 2, bn_norm, with_ibn=with_ibn,
                                           dilation=1, dropblock_prob=dropblock_prob)
            self.layer4 = self._make_layer(block, 512, layers[3], 1, bn_norm, with_ibn=with_ibn,
                                           dilation=2, dropblock_prob=dropblock_prob)
        else:
            self.layer3 = self._make_layer(block, 256, layers[2], 2, bn_norm, with_ibn=with_ibn,
                                           dropblock_prob=dropblock_prob)
            self.layer4 = self._make_layer(block, 512, layers[3], last_stride, bn_norm, with_ibn=with_ibn,
                                           dropblock_prob=dropblock_prob)
 
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
                m.weight.data.normal_(0, math.sqrt(2. / n))
            elif isinstance(m, nn.BatchNorm2d):
                m.weight.data.fill_(1)
                m.bias.data.zero_()
 
        # fmt: off
        if with_nl: self._build_nonlocal(layers, non_layers, bn_norm)
        else:       self.NL_1_idx = self.NL_2_idx = self.NL_3_idx = self.NL_4_idx = []
        # fmt: on
 
    def _make_layer(self, block, planes, blocks, stride=1, bn_norm="BN", with_ibn=False,
                    dilation=1, dropblock_prob=0.0, is_first=True):
        downsample = None
        if stride != 1 or self.inplanes != planes * block.expansion:
            down_layers = []
            if self.avg_down:
                if dilation == 1:
                    down_layers.append(nn.AvgPool2d(kernel_size=stride, stride=stride,
                                                    ceil_mode=True, count_include_pad=False))
                else:
                    down_layers.append(nn.AvgPool2d(kernel_size=1, stride=1,
                                                    ceil_mode=True, count_include_pad=False))
                down_layers.append(nn.Conv2d(self.inplanes, planes * block.expansion,
                                             kernel_size=1, stride=1, bias=False))
            else:
                down_layers.append(nn.Conv2d(self.inplanes, planes * block.expansion,
                                             kernel_size=1, stride=stride, bias=False))
            down_layers.append(get_norm(bn_norm, planes * block.expansion))
            downsample = nn.Sequential(*down_layers)
 
        layers = []
        if dilation == 1 or dilation == 2:
            layers.append(block(self.inplanes, planes, bn_norm, with_ibn, stride, downsample=downsample,
                                radix=self.radix, cardinality=self.cardinality,
                                bottleneck_width=self.bottleneck_width,
                                avd=self.avd, avd_first=self.avd_first,
                                dilation=1, is_first=is_first, rectified_conv=self.rectified_conv,
                                rectify_avg=self.rectify_avg,
                                dropblock_prob=dropblock_prob,
                                last_gamma=self.last_gamma))
        elif dilation == 4:
            layers.append(block(self.inplanes, planes, bn_norm, with_ibn, stride, downsample=downsample,
                                radix=self.radix, cardinality=self.cardinality,
                                bottleneck_width=self.bottleneck_width,
                                avd=self.avd, avd_first=self.avd_first,
                                dilation=2, is_first=is_first, rectified_conv=self.rectified_conv,
                                rectify_avg=self.rectify_avg,
                                dropblock_prob=dropblock_prob,
                                last_gamma=self.last_gamma))
        else:
            raise RuntimeError("=> unknown dilation size: {}".format(dilation))
 
        self.inplanes = planes * block.expansion
        for i in range(1, blocks):
            layers.append(block(self.inplanes, planes, bn_norm, with_ibn,
                                radix=self.radix, cardinality=self.cardinality,
                                bottleneck_width=self.bottleneck_width,
                                avd=self.avd, avd_first=self.avd_first,
                                dilation=dilation, rectified_conv=self.rectified_conv,
                                rectify_avg=self.rectify_avg,
                                dropblock_prob=dropblock_prob,
                                last_gamma=self.last_gamma))
 
        return nn.Sequential(*layers)
 
    def _build_nonlocal(self, layers, non_layers, bn_norm):
        self.NL_1 = nn.ModuleList(
            [Non_local(256, bn_norm) for _ in range(non_layers[0])])
        self.NL_1_idx = sorted([layers[0] - (i + 1) for i in range(non_layers[0])])
        self.NL_2 = nn.ModuleList(
            [Non_local(512, bn_norm) for _ in range(non_layers[1])])
        self.NL_2_idx = sorted([layers[1] - (i + 1) for i in range(non_layers[1])])
        self.NL_3 = nn.ModuleList(
            [Non_local(1024, bn_norm) for _ in range(non_layers[2])])
        self.NL_3_idx = sorted([layers[2] - (i + 1) for i in range(non_layers[2])])
        self.NL_4 = nn.ModuleList(
            [Non_local(2048, bn_norm) for _ in range(non_layers[3])])
        self.NL_4_idx = sorted([layers[3] - (i + 1) for i in range(non_layers[3])])
 
    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)
 
        NL1_counter = 0
        if len(self.NL_1_idx) == 0:
            self.NL_1_idx = [-1]
        for i in range(len(self.layer1)):
            x = self.layer1[i](x)
            if i == self.NL_1_idx[NL1_counter]:
                _, C, H, W = x.shape
                x = self.NL_1[NL1_counter](x)
                NL1_counter += 1
        # Layer 2
        NL2_counter = 0
        if len(self.NL_2_idx) == 0:
            self.NL_2_idx = [-1]
        for i in range(len(self.layer2)):
            x = self.layer2[i](x)
            if i == self.NL_2_idx[NL2_counter]:
                _, C, H, W = x.shape
                x = self.NL_2[NL2_counter](x)
                NL2_counter += 1
        # Layer 3
        NL3_counter = 0
        if len(self.NL_3_idx) == 0:
            self.NL_3_idx = [-1]
        for i in range(len(self.layer3)):
            x = self.layer3[i](x)
            if i == self.NL_3_idx[NL3_counter]:
                _, C, H, W = x.shape
                x = self.NL_3[NL3_counter](x)
                NL3_counter += 1
        # Layer 4
        NL4_counter = 0
        if len(self.NL_4_idx) == 0:
            self.NL_4_idx = [-1]
        for i in range(len(self.layer4)):
            x = self.layer4[i](x)
            if i == self.NL_4_idx[NL4_counter]:
                _, C, H, W = x.shape
                x = self.NL_4[NL4_counter](x)
                NL4_counter += 1
 
        return x
 
 
@BACKBONE_REGISTRY.register()
def build_resnest_backbone(cfg):
    """
    Create a ResNest instance from config.
    Returns:
        ResNet: a :class:`ResNet` instance.
    """
 
    # fmt: off
    pretrain      = cfg.MODEL.BACKBONE.PRETRAIN
    pretrain_path = cfg.MODEL.BACKBONE.PRETRAIN_PATH
    last_stride   = cfg.MODEL.BACKBONE.LAST_STRIDE
    bn_norm       = cfg.MODEL.BACKBONE.NORM
    with_ibn      = cfg.MODEL.BACKBONE.WITH_IBN
    with_se       = cfg.MODEL.BACKBONE.WITH_SE
    with_nl       = cfg.MODEL.BACKBONE.WITH_NL
    depth         = cfg.MODEL.BACKBONE.DEPTH
    # fmt: on
 
    num_blocks_per_stage = {
        "50x": [3, 4, 6, 3],
        "101x": [3, 4, 23, 3],
        "200x": [3, 24, 36, 3],
        "269x": [3, 30, 48, 8],
    }[depth]
 
    nl_layers_per_stage = {
        "50x": [0, 2, 3, 0],
        "101x": [0, 2, 3, 0],
        "200x": [0, 2, 3, 0],
        "269x": [0, 2, 3, 0],
    }[depth]
 
    stem_width = {
        "50x": 32,
        "101x": 64,
        "200x": 64,
        "269x": 64,
    }[depth]
 
    model = ResNest(last_stride, bn_norm, with_ibn, with_nl, Bottleneck, num_blocks_per_stage,
                    nl_layers_per_stage, radix=2, groups=1, bottleneck_width=64,
                    deep_stem=True, stem_width=stem_width, avg_down=True,
                    avd=True, avd_first=False)
    if pretrain:
        # Load pretrain path if specifically
        if pretrain_path:
            try:
                state_dict = torch.load(pretrain_path, map_location=torch.device('cpu'))
                logger.info(f"Loading pretrained model from {pretrain_path}")
            except FileNotFoundError as e:
                logger.info(f'{pretrain_path} is not found! Please check this path.')
                raise e
            except KeyError as e:
                logger.info("State dict keys error! Please check the state dict.")
                raise e
        else:
            state_dict = torch.hub.load_state_dict_from_url(
                model_urls['resnest' + depth[:-1]], progress=True, check_hash=True, map_location=torch.device('cpu'))
 
        incompatible = model.load_state_dict(state_dict, strict=False)
        if incompatible.missing_keys:
            logger.info(
                get_missing_parameters_message(incompatible.missing_keys)
            )
        if incompatible.unexpected_keys:
            logger.info(
                get_unexpected_parameters_message(incompatible.unexpected_keys)
            )
    return model