""" Normalizer Free RegNet / ResNet (pre-activation) Models Paper: `Characterizing signal propagation to close the performance gap in unnormalized ResNets` - https://arxiv.org/abs/2101.08692 Hacked together by / copyright Ross Wightman, 2021. """ import math from dataclasses import dataclass, field from collections import OrderedDict from typing import Tuple, Optional from functools import partial import torch import torch.nn as nn import torch.nn.functional as F from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD from .helpers import build_model_with_cfg from .registry import register_model from .layers import ClassifierHead, DropPath, AvgPool2dSame, ScaledStdConv2d, get_act_layer, get_attn, make_divisible, get_act_fn def _dcfg(url='', **kwargs): return { 'url': url, 'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': (7, 7), 'crop_pct': 0.875, 'interpolation': 'bicubic', 'mean': IMAGENET_DEFAULT_MEAN, 'std': IMAGENET_DEFAULT_STD, 'first_conv': 'stem.conv', 'classifier': 'head.fc', **kwargs } # FIXME finish default_cfgs = { 'nf_regnet_b0': _dcfg(url=''), 'nf_regnet_b1': _dcfg(url='', input_size=(3, 240, 240)), 'nf_regnet_b2': _dcfg(url='', input_size=(3, 256, 256)), 'nf_regnet_b3': _dcfg(url='', input_size=(3, 272, 272)), 'nf_regnet_b4': _dcfg(url='', input_size=(3, 320, 320)), 'nf_regnet_b5': _dcfg(url='', input_size=(3, 384, 384)), 'nf_resnet26': _dcfg(url='', first_conv='stem.conv'), 'nf_resnet50': _dcfg(url='', first_conv='stem.conv'), 'nf_resnet101': _dcfg(url='', first_conv='stem.conv'), 'nf_seresnet26': _dcfg(url='', first_conv='stem.conv'), 'nf_seresnet50': _dcfg(url='', first_conv='stem.conv'), 'nf_seresnet101': _dcfg(url='', first_conv='stem.conv'), 'nf_ecaresnet26': _dcfg(url='', first_conv='stem.conv'), 'nf_ecaresnet50': _dcfg(url='', first_conv='stem.conv'), 'nf_ecaresnet101': _dcfg(url='', first_conv='stem.conv'), } @dataclass class NfCfg: depths: Tuple[int, int, int, int] channels: Tuple[int, int, int, int] alpha: float = 0.2 gamma_in_act: bool = False stem_type: str = '3x3' stem_chs: Optional[int] = None group_size: Optional[int] = 8 attn_layer: Optional[str] = 'se' attn_kwargs: dict = field(default_factory=lambda: dict(reduction_ratio=0.5, divisor=8)) attn_gain: float = 2.0 # NF correction gain to apply if attn layer is used width_factor: float = 0.75 bottle_ratio: float = 2.25 efficient: bool = True # enables EfficientNet-like options that are used in paper 'nf_regnet_b*' models num_features: int = 1280 # num out_channels for final conv (when enabled in efficient mode) ch_div: int = 8 # round channels % 8 == 0 to keep tensor-core use optimal skipinit: bool = False act_layer: str = 'silu' model_cfgs = dict( # EffNet influenced RegNet defs nf_regnet_b0=NfCfg(depths=(1, 3, 6, 6), channels=(48, 104, 208, 440), num_features=1280), nf_regnet_b1=NfCfg(depths=(2, 4, 7, 7), channels=(48, 104, 208, 440), num_features=1280), nf_regnet_b2=NfCfg(depths=(2, 4, 8, 8), channels=(56, 112, 232, 488), num_features=1416), nf_regnet_b3=NfCfg(depths=(2, 5, 9, 9), channels=(56, 128, 248, 528), num_features=1536), nf_regnet_b4=NfCfg(depths=(2, 6, 11, 11), channels=(64, 144, 288, 616), num_features=1792), nf_regnet_b5=NfCfg(depths=(3, 7, 14, 14), channels=(80, 168, 336, 704), num_features=2048), # ResNet (preact, D style deep stem/avg down) defs nf_resnet26=NfCfg( depths=(2, 2, 2, 2), channels=(256, 512, 1024, 2048), stem_type='7x7_pool', stem_chs=64, width_factor=1.0, bottle_ratio=0.25, efficient=False, group_size=None, act_layer='relu', attn_layer=None,), nf_resnet50=NfCfg( depths=(3, 4, 6, 3), channels=(256, 512, 1024, 2048), stem_type='7x7_pool', stem_chs=64, width_factor=1.0, bottle_ratio=0.25, efficient=False, group_size=None, act_layer='relu', attn_layer=None), nf_resnet101=NfCfg( depths=(3, 4, 6, 3), channels=(256, 512, 1024, 2048), stem_type='7x7_pool', stem_chs=64, width_factor=1.0, bottle_ratio=0.25, efficient=False, group_size=None, act_layer='relu', attn_layer=None), nf_seresnet26=NfCfg( depths=(2, 2, 2, 2), channels=(256, 512, 1024, 2048), stem_type='7x7_pool', stem_chs=64, width_factor=1.0, bottle_ratio=0.25, efficient=False, group_size=None, act_layer='relu', attn_layer='se', attn_kwargs=dict(reduction_ratio=0.25)), nf_seresnet50=NfCfg( depths=(3, 4, 6, 3), channels=(256, 512, 1024, 2048), stem_type='7x7_pool', stem_chs=64, width_factor=1.0, bottle_ratio=0.25, efficient=False, group_size=None, act_layer='relu', attn_layer='se', attn_kwargs=dict(reduction_ratio=0.25)), nf_seresnet101=NfCfg( depths=(3, 4, 6, 3), channels=(256, 512, 1024, 2048), stem_type='7x7_pool', stem_chs=64, width_factor=1.0, bottle_ratio=0.25, efficient=False, group_size=None, act_layer='relu', attn_layer='se', attn_kwargs=dict(reduction_ratio=0.25)), nf_ecaresnet26=NfCfg( depths=(2, 2, 2, 2), channels=(256, 512, 1024, 2048), stem_type='7x7_pool', stem_chs=64, width_factor=1.0, bottle_ratio=0.25, efficient=False, group_size=None, act_layer='relu', attn_layer='eca', attn_kwargs=dict()), nf_ecaresnet50=NfCfg( depths=(3, 4, 6, 3), channels=(256, 512, 1024, 2048), stem_type='7x7_pool', stem_chs=64, width_factor=1.0, bottle_ratio=0.25, efficient=False, group_size=None, act_layer='relu', attn_layer='eca', attn_kwargs=dict()), nf_ecaresnet101=NfCfg( depths=(3, 4, 6, 3), channels=(256, 512, 1024, 2048), stem_type='7x7_pool', stem_chs=64, width_factor=1.0, bottle_ratio=0.25, efficient=False, group_size=None, act_layer='relu', attn_layer='eca', attn_kwargs=dict()), ) class GammaAct(nn.Module): def __init__(self, act_type='relu', gamma: float = 1.0, inplace=False): super().__init__() self.act_fn = get_act_fn(act_type) self.gamma = gamma self.inplace = inplace def forward(self, x): return self.gamma * self.act_fn(x, inplace=self.inplace) def act_with_gamma(act_type, gamma: float = 1.): def _create(inplace=False): return GammaAct(act_type, gamma=gamma, inplace=inplace) return _create class DownsampleAvg(nn.Module): def __init__( self, in_chs, out_chs, stride=1, dilation=1, first_dilation=None, conv_layer=ScaledStdConv2d): """ AvgPool Downsampling as in 'D' ResNet variants. Support for dilation.""" super(DownsampleAvg, self).__init__() avg_stride = stride if dilation == 1 else 1 if stride > 1 or dilation > 1: avg_pool_fn = AvgPool2dSame if avg_stride == 1 and dilation > 1 else nn.AvgPool2d self.pool = avg_pool_fn(2, avg_stride, ceil_mode=True, count_include_pad=False) else: self.pool = nn.Identity() self.conv = conv_layer(in_chs, out_chs, 1, stride=1) def forward(self, x): return self.conv(self.pool(x)) class NormalizationFreeBlock(nn.Module): """Normalization-free pre-activation block. """ def __init__( self, in_chs, out_chs=None, stride=1, dilation=1, first_dilation=None, alpha=1.0, beta=1.0, bottle_ratio=0.25, efficient=True, ch_div=1, group_size=None, attn_layer=None, attn_gain=2.0, act_layer=None, conv_layer=None, drop_path_rate=0., skipinit=False): super().__init__() first_dilation = first_dilation or dilation out_chs = out_chs or in_chs # EfficientNet-like models scale bottleneck from in_chs, otherwise scale from out_chs like ResNet mid_chs = make_divisible(in_chs * bottle_ratio if efficient else out_chs * bottle_ratio, ch_div) groups = 1 if group_size is None else mid_chs // group_size if group_size and group_size % ch_div == 0: mid_chs = group_size * groups # correct mid_chs if group_size divisible by ch_div, otherwise error self.alpha = alpha self.beta = beta self.attn_gain = attn_gain if in_chs != out_chs or stride != 1 or dilation != first_dilation: self.downsample = DownsampleAvg( in_chs, out_chs, stride=stride, dilation=dilation, first_dilation=first_dilation, conv_layer=conv_layer) else: self.downsample = None self.act1 = act_layer() self.conv1 = conv_layer(in_chs, mid_chs, 1) self.act2 = act_layer(inplace=True) self.conv2 = conv_layer(mid_chs, mid_chs, 3, stride=stride, dilation=first_dilation, groups=groups) if attn_layer is not None: self.attn = attn_layer(mid_chs) else: self.attn = None self.act3 = act_layer() self.conv3 = conv_layer(mid_chs, out_chs, 1) self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0 else nn.Identity() self.skipinit_gain = nn.Parameter(torch.tensor(0.)) if skipinit else None def forward(self, x): out = self.act1(x) * self.beta # shortcut branch shortcut = x if self.downsample is not None: shortcut = self.downsample(out) # residual branch out = self.conv1(out) out = self.conv2(self.act2(out)) if self.attn is not None: out = self.attn_gain * self.attn(out) out = self.conv3(self.act3(out)) out = self.drop_path(out) if self.skipinit_gain is None: out = out * self.alpha + shortcut else: # this really slows things down for some reason, TBD out = out * self.alpha * self.skipinit_gain + shortcut return out def create_stem(in_chs, out_chs, stem_type='', conv_layer=None): stem_stride = 2 stem = OrderedDict() assert stem_type in ('', 'deep', '3x3', '7x7', 'deep_pool', '3x3_pool', '7x7_pool') if 'deep' in stem_type: # 3 deep 3x3 conv stack as in ResNet V1D models. NOTE: doesn't work as well here mid_chs = out_chs // 2 stem['conv1'] = conv_layer(in_chs, mid_chs, kernel_size=3, stride=2) stem['conv2'] = conv_layer(mid_chs, mid_chs, kernel_size=3, stride=1) stem['conv3'] = conv_layer(mid_chs, out_chs, kernel_size=3, stride=1) elif '3x3' in stem_type: # 3x3 stem conv as in RegNet stem['conv'] = conv_layer(in_chs, out_chs, kernel_size=3, stride=2) else: # 7x7 stem conv as in ResNet stem['conv'] = conv_layer(in_chs, out_chs, kernel_size=7, stride=2) if 'pool' in stem_type: stem['pool'] = nn.MaxPool2d(3, stride=2, padding=1) stem_stride = 4 return nn.Sequential(stem), stem_stride _nonlin_gamma = dict( silu=1./.5595, relu=(0.5 * (1. - 1. / math.pi)) ** -0.5, identity=1.0 ) class NormalizerFreeNet(nn.Module): """ Normalizer-free ResNets and RegNets As described in `Characterizing signal propagation to close the performance gap in unnormalized ResNets` - https://arxiv.org/abs/2101.08692 This model aims to cover both the NFRegNet-Bx models as detailed in the paper's code snippets and the (preact) ResNet models described earlier in the paper. There are a few differences: * channels are rounded to be divisible by 8 by default (keep tensor core kernels happy), this changes channel dim and param counts slightly from the paper models * activation correcting gamma constants are moved into the ScaledStdConv as it has less performance impact in PyTorch when done with the weight scaling there. This likely wasn't a concern in the JAX impl. * a config option `gamma_in_act` can be enabled to not apply gamma in StdConv as described above, but apply it in each activation. This is slightly slower, and yields slightly different results. * skipinit is disabled by default, it seems to have a rather drastic impact on GPU memory use and throughput for what it is/does. Approx 8-10% throughput loss. """ def __init__(self, cfg: NfCfg, num_classes=1000, in_chans=3, global_pool='avg', output_stride=32, drop_rate=0., drop_path_rate=0.): super().__init__() self.num_classes = num_classes self.drop_rate = drop_rate assert cfg.act_layer in _nonlin_gamma, f"Please add non-linearity constants for activation ({cfg.act_layer})." if cfg.gamma_in_act: act_layer = act_with_gamma(cfg.act_layer, gamma=_nonlin_gamma[cfg.act_layer]) conv_layer = partial(ScaledStdConv2d, bias=True, gain=True) else: act_layer = get_act_layer(cfg.act_layer) conv_layer = partial(ScaledStdConv2d, bias=True, gain=True, gamma=_nonlin_gamma[cfg.act_layer]) attn_layer = partial(get_attn(cfg.attn_layer), **cfg.attn_kwargs) if cfg.attn_layer else None stem_chs = cfg.stem_chs or cfg.channels[0] stem_chs = make_divisible(stem_chs * cfg.width_factor, cfg.ch_div) self.stem, stem_stride = create_stem(in_chans, stem_chs, cfg.stem_type, conv_layer=conv_layer) self.feature_info = [] # NOTE: there will be no stride == 2 feature if stem_stride == 4 dpr = [x.tolist() for x in torch.linspace(0, drop_path_rate, sum(cfg.depths)).split(cfg.depths)] prev_chs = stem_chs net_stride = stem_stride dilation = 1 expected_var = 1.0 stages = [] for stage_idx, stage_depth in enumerate(cfg.depths): stride = 1 if stage_idx == 0 and stem_stride > 2 else 2 self.feature_info += [dict( num_chs=prev_chs, reduction=net_stride, module=f'stages.{stage_idx}.0.act1' if stride == 2 else '')] if net_stride >= output_stride and stride > 1: dilation *= stride stride = 1 net_stride *= stride first_dilation = 1 if dilation in (1, 2) else 2 blocks = [] for block_idx in range(cfg.depths[stage_idx]): first_block = block_idx == 0 and stage_idx == 0 out_chs = make_divisible(cfg.channels[stage_idx] * cfg.width_factor, cfg.ch_div) blocks += [NormalizationFreeBlock( in_chs=prev_chs, out_chs=out_chs, alpha=cfg.alpha, beta=1. / expected_var ** 0.5, # NOTE: beta used as multiplier in block stride=stride if block_idx == 0 else 1, dilation=dilation, first_dilation=first_dilation, group_size=cfg.group_size, bottle_ratio=1. if cfg.efficient and first_block else cfg.bottle_ratio, efficient=cfg.efficient, ch_div=cfg.ch_div, attn_layer=attn_layer, attn_gain=cfg.attn_gain, act_layer=act_layer, conv_layer=conv_layer, drop_path_rate=dpr[stage_idx][block_idx], skipinit=cfg.skipinit, )] if block_idx == 0: expected_var = 1. # expected var is reset after first block of each stage expected_var += cfg.alpha ** 2 # Even if reset occurs, increment expected variance first_dilation = dilation prev_chs = out_chs stages += [nn.Sequential(*blocks)] self.stages = nn.Sequential(*stages) if cfg.efficient and cfg.num_features: # The paper NFRegNet models have an EfficientNet-like final head convolution. self.num_features = make_divisible(cfg.width_factor * cfg.num_features, cfg.ch_div) self.final_conv = conv_layer(prev_chs, self.num_features, 1) else: self.num_features = prev_chs self.final_conv = nn.Identity() # FIXME not 100% clear on gamma subtleties final conv/final act in case where it's in stdconv self.final_act = act_layer() self.feature_info += [dict(num_chs=self.num_features, reduction=net_stride, module='final_act')] self.head = ClassifierHead(self.num_features, num_classes, pool_type=global_pool, drop_rate=self.drop_rate) for n, m in self.named_modules(): if 'fc' in n and isinstance(m, nn.Linear): nn.init.zeros_(m.weight) if m.bias is not None: nn.init.zeros_(m.bias) elif isinstance(m, nn.Conv2d): # as per discussion with paper authors, original in haiku is # hk.initializers.VarianceScaling(1.0, 'fan_in', 'normal')' w/ zero'd bias nn.init.kaiming_normal_(m.weight, mode='fan_in', nonlinearity='linear') if m.bias is not None: nn.init.zeros_(m.bias) def get_classifier(self): return self.head.fc def reset_classifier(self, num_classes, global_pool='avg'): self.head = ClassifierHead(self.num_features, num_classes, pool_type=global_pool, drop_rate=self.drop_rate) def forward_features(self, x): x = self.stem(x) x = self.stages(x) x = self.final_conv(x) x = self.final_act(x) return x def forward(self, x): x = self.forward_features(x) x = self.head(x) return x def _create_normfreenet(variant, pretrained=False, **kwargs): model_cfg = model_cfgs[variant] feature_cfg = dict(flatten_sequential=True) feature_cfg['feature_cls'] = 'hook' # pre-act models need hooks to grab feat from act1 in bottleneck blocks if 'pool' in model_cfg.stem_type: feature_cfg['out_indices'] = (1, 2, 3, 4) # no stride 2, 0 level feat for stride 4 maxpool stems in ResNet return build_model_with_cfg( NormalizerFreeNet, variant, pretrained, model_cfg=model_cfg, default_cfg=default_cfgs[variant], feature_cfg=feature_cfg, **kwargs) @register_model def nf_regnet_b0(pretrained=False, **kwargs): return _create_normfreenet('nf_regnet_b0', pretrained=pretrained, **kwargs) @register_model def nf_regnet_b1(pretrained=False, **kwargs): return _create_normfreenet('nf_regnet_b1', pretrained=pretrained, **kwargs) @register_model def nf_regnet_b2(pretrained=False, **kwargs): return _create_normfreenet('nf_regnet_b2', pretrained=pretrained, **kwargs) @register_model def nf_regnet_b3(pretrained=False, **kwargs): return _create_normfreenet('nf_regnet_b3', pretrained=pretrained, **kwargs) @register_model def nf_regnet_b4(pretrained=False, **kwargs): return _create_normfreenet('nf_regnet_b4', pretrained=pretrained, **kwargs) @register_model def nf_regnet_b5(pretrained=False, **kwargs): return _create_normfreenet('nf_regnet_b5', pretrained=pretrained, **kwargs) @register_model def nf_resnet26(pretrained=False, **kwargs): return _create_normfreenet('nf_resnet26', pretrained=pretrained, **kwargs) @register_model def nf_resnet50(pretrained=False, **kwargs): return _create_normfreenet('nf_resnet50', pretrained=pretrained, **kwargs) @register_model def nf_seresnet26(pretrained=False, **kwargs): return _create_normfreenet('nf_seresnet26', pretrained=pretrained, **kwargs) @register_model def nf_seresnet50(pretrained=False, **kwargs): return _create_normfreenet('nf_seresnet50', pretrained=pretrained, **kwargs) @register_model def nf_ecaresnet26(pretrained=False, **kwargs): return _create_normfreenet('nf_ecaresnet26', pretrained=pretrained, **kwargs) @register_model def nf_ecaresnet50(pretrained=False, **kwargs): return _create_normfreenet('nf_ecaresnet50', pretrained=pretrained, **kwargs)