""" Convolution with Weight Standardization (StdConv and ScaledStdConv) StdConv: @article{weightstandardization, author = {Siyuan Qiao and Huiyu Wang and Chenxi Liu and Wei Shen and Alan Yuille}, title = {Weight Standardization}, journal = {arXiv preprint arXiv:1903.10520}, year = {2019}, } Code: https://github.com/joe-siyuan-qiao/WeightStandardization ScaledStdConv: Paper: `Characterizing signal propagation to close the performance gap in unnormalized ResNets` - https://arxiv.org/abs/2101.08692 Official Deepmind JAX code: https://github.com/deepmind/deepmind-research/tree/master/nfnets Hacked together by / copyright Ross Wightman, 2021. """ from typing import Optional, Tuple, Union import torch import torch.nn as nn import torch.nn.functional as F from ._fx import register_notrace_module from .padding import get_padding, get_padding_value, pad_same class StdConv2d(nn.Conv2d): """Conv2d with Weight Standardization. Used for BiT ResNet-V2 models. Paper: `Micro-Batch Training with Batch-Channel Normalization and Weight Standardization` - https://arxiv.org/abs/1903.10520v2 """ def __init__( self, in_channel: int, out_channels: int, kernel_size: Union[int, Tuple[int, int]], stride: Union[int, Tuple[int, int]] = 1, padding: Optional[Union[int, Tuple[int, int]]] = None, dilation: Union[int, Tuple[int, int]] = 1, groups: int = 1, bias: bool = False, eps: float = 1e-6, device=None, dtype=None, ): if padding is None: padding = get_padding(kernel_size, stride, dilation) super().__init__( in_channel, out_channels, kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias, device=device, dtype=dtype) self.eps = eps def forward(self, x): weight = F.batch_norm( self.weight.reshape(1, self.out_channels, -1), None, # running_mean None, # running_var training=True, momentum=0., eps=self.eps, ).reshape_as(self.weight) x = F.conv2d(x, weight, self.bias, self.stride, self.padding, self.dilation, self.groups) return x @register_notrace_module class StdConv2dSame(nn.Conv2d): """Conv2d with Weight Standardization. TF compatible SAME padding. Used for ViT Hybrid model. Paper: `Micro-Batch Training with Batch-Channel Normalization and Weight Standardization` - https://arxiv.org/abs/1903.10520v2 """ def __init__( self, in_channel: int, out_channels: int, kernel_size: Union[int, Tuple[int, int]], stride: Union[int, Tuple[int, int]] = 1, padding: str = 'SAME', dilation: Union[int, Tuple[int, int]] = 1, groups: int = 1, bias: bool = False, eps: float = 1e-6, device=None, dtype=None, ): padding, is_dynamic = get_padding_value(padding, kernel_size, stride=stride, dilation=dilation) super().__init__( in_channel, out_channels, kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias, device=device, dtype=dtype) self.same_pad = is_dynamic self.eps = eps def forward(self, x): if self.same_pad: x = pad_same(x, self.kernel_size, self.stride, self.dilation) weight = F.batch_norm( self.weight.reshape(1, self.out_channels, -1), None, # running_mean None, # running_var training=True, momentum=0., eps=self.eps, ).reshape_as(self.weight) x = F.conv2d(x, weight, self.bias, self.stride, self.padding, self.dilation, self.groups) return x class ScaledStdConv2d(nn.Conv2d): """Conv2d layer with Scaled Weight Standardization. Paper: `Characterizing signal propagation to close the performance gap in unnormalized ResNets` - https://arxiv.org/abs/2101.08692 NOTE: the operations used in this impl differ slightly from the DeepMind Haiku impl. The impact is minor. """ def __init__( self, in_channels: int, out_channels: int, kernel_size: Union[int, Tuple[int, int]], stride: Union[int, Tuple[int, int]] = 1, padding: Optional[Union[int, Tuple[int, int], str]] = None, dilation: Union[int, Tuple[int, int]] = 1, groups: int = 1, bias: bool = True, gamma: float = 1.0, eps: float = 1e-6, gain_init: float = 1.0, device=None, dtype=None, ): dd = {'device': device, 'dtype': dtype} if padding is None: padding = get_padding(kernel_size, stride, dilation) super().__init__( in_channels, out_channels, kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias, **dd) self.scale = gamma * self.weight[0].numel() ** -0.5 # gamma * 1 / sqrt(fan-in) self.eps = eps self.gain_init = gain_init self.gain = nn.Parameter(torch.empty((self.out_channels, 1, 1, 1), **dd)) self.reset_parameters() def reset_parameters(self) -> None: # Only initialize gain if it exists (for the second call) if hasattr(self, 'gain'): torch.nn.init.constant_(self.gain, self.gain_init) # Also reset parent parameters if needed super().reset_parameters() # On first call (from super().__init__), do nothing def forward(self, x): weight = F.batch_norm( self.weight.reshape(1, self.out_channels, -1), None, # running_mean None, # running_var weight=(self.gain * self.scale).view(-1), training=True, momentum=0., eps=self.eps, ).reshape_as(self.weight) return F.conv2d(x, weight, self.bias, self.stride, self.padding, self.dilation, self.groups) @register_notrace_module class ScaledStdConv2dSame(nn.Conv2d): """Conv2d layer with Scaled Weight Standardization and Tensorflow-like SAME padding support Paper: `Characterizing signal propagation to close the performance gap in unnormalized ResNets` - https://arxiv.org/abs/2101.08692 NOTE: the operations used in this impl differ slightly from the DeepMind Haiku impl. The impact is minor. """ def __init__( self, in_channels: int, out_channels: int, kernel_size: Union[int, Tuple[int, int]], stride: Union[int, Tuple[int, int]] = 1, padding: str = 'SAME', dilation: Union[int, Tuple[int, int]] = 1, groups: int = 1, bias: bool = True, gamma: float = 1.0, eps: float = 1e-6, gain_init: float = 1.0, device=None, dtype=None, ): dd = {'device': device, 'dtype': dtype} padding, is_dynamic = get_padding_value(padding, kernel_size, stride=stride, dilation=dilation) super().__init__( in_channels, out_channels, kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias, **dd) self.scale = gamma * self.weight[0].numel() ** -0.5 self.same_pad = is_dynamic self.eps = eps self.gain_init = gain_init self.gain = nn.Parameter(torch.empty((self.out_channels, 1, 1, 1), **dd)) self.reset_parameters() def reset_parameters(self) -> None: # Only initialize gain if it exists (for the second call) if hasattr(self, 'gain'): torch.nn.init.constant_(self.gain, self.gain_init) # Also reset parent parameters if needed super().reset_parameters() # On first call (from super().__init__), do nothing def forward(self, x): if self.same_pad: x = pad_same(x, self.kernel_size, self.stride, self.dilation) weight = F.batch_norm( self.weight.reshape(1, self.out_channels, -1), None, # running_mean None, # running_var weight=(self.gain * self.scale).view(-1), training=True, momentum=0., eps=self.eps, ).reshape_as(self.weight) return F.conv2d(x, weight, self.bias, self.stride, self.padding, self.dilation, self.groups)