All ScaledStdConv and StdConv uses default to using F.layernorm so that they work with PyTorch XLA. eps value tweaking is a WIP.

Author: rwightman

timm/models/layers/std_conv.py

@@ -19,17 +19,22 @@ class StdConv2d(nn.Conv2d):
  """
  def __init__(
  self, in_channel, out_channels, kernel_size, stride=1, padding=None, dilation=1,
- groups=1, bias=False, eps=1e-5):
+ groups=1, bias=False, eps=1e-5, use_layernorm=True):
  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)
  self.eps = eps
+ self.use_layernorm = use_layernorm
 
  def get_weight(self):
- std, mean = torch.std_mean(self.weight, dim=[1, 2, 3], keepdim=True, unbiased=False)
- weight = (self.weight - mean) / (std + self.eps)
+ if self.use_layernorm:
+ # NOTE F.layer_norm is being used to compute (self.weight - mean) / (sqrt(var) + self.eps) in one op
+ weight = F.layer_norm(self.weight, self.weight.shape[1:], eps=self.eps)
+ else:
+ std, mean = torch.std_mean(self.weight, dim=[1, 2, 3], keepdim=True, unbiased=False)
+ weight = (self.weight - mean) / (std + self.eps)
  return weight
 
  def forward(self, x):
@@ -45,17 +50,22 @@ class StdConv2dSame(nn.Conv2d):
  """
  def __init__(
  self, in_channel, out_channels, kernel_size, stride=1, padding='SAME', dilation=1,
- groups=1, bias=False, eps=1e-5):
+ groups=1, bias=False, eps=1e-5, use_layernorm=True):
  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)
  self.same_pad = is_dynamic
  self.eps = eps
+ self.use_layernorm = use_layernorm
 
  def get_weight(self):
- std, mean = torch.std_mean(self.weight, dim=[1, 2, 3], keepdim=True, unbiased=False)
- weight = (self.weight - mean) / (std + self.eps)
+ if self.use_layernorm:
+ # NOTE F.layer_norm is being used to compute (self.weight - mean) / (sqrt(var) + self.eps) in one op
+ weight = F.layer_norm(self.weight, self.weight.shape[1:], eps=self.eps)
+ else:
+ std, mean = torch.std_mean(self.weight, dim=[1, 2, 3], keepdim=True, unbiased=False)
+ weight = (self.weight - mean) / (std + self.eps)
  return weight
 
  def forward(self, x):
@@ -76,24 +86,25 @@ class ScaledStdConv2d(nn.Conv2d):
 
  def __init__(
  self, in_channels, out_channels, kernel_size, stride=1, padding=None, dilation=1, groups=1,
- bias=True, gamma=1.0, eps=1e-5, gain_init=1.0, use_layernorm=False):
+ bias=True, gamma=1.0, eps=1e-5, gain_init=1.0, use_layernorm=True):
  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)
  self.gain = nn.Parameter(torch.full((self.out_channels, 1, 1, 1), gain_init))
  self.scale = gamma * self.weight[0].numel() ** -0.5 # gamma * 1 / sqrt(fan-in)
- self.eps = eps ** 2 if use_layernorm else eps
+ self.eps = eps
  self.use_layernorm = use_layernorm # experimental, slightly faster/less GPU memory to hijack LN kernel
 
  def get_weight(self):
  if self.use_layernorm:
- weight = self.scale * F.layer_norm(self.weight, self.weight.shape[1:], eps=self.eps)
+ # NOTE F.layer_norm is being used to compute (self.weight - mean) / (sqrt(var) + self.eps) in one op
+ weight = F.layer_norm(self.weight, self.weight.shape[1:], eps=self.eps)
  else:
  std, mean = torch.std_mean(self.weight, dim=[1, 2, 3], keepdim=True, unbiased=False)
- weight = self.scale * (self.weight - mean) / (std + self.eps)
- return self.gain * weight
+ weight = (self.weight - mean) / (std + self.eps)
+ return weight.mul_(self.gain * self.scale)
 
  def forward(self, x):
  return F.conv2d(x, self.get_weight(), self.bias, self.stride, self.padding, self.dilation, self.groups)
@@ -110,32 +121,25 @@ class ScaledStdConv2dSame(nn.Conv2d):
 
  def __init__(
  self, in_channels, out_channels, kernel_size, stride=1, padding='SAME', dilation=1, groups=1,
- bias=True, gamma=1.0, eps=1e-5, gain_init=1.0, use_layernorm=False):
+ bias=True, gamma=1.0, eps=1e-5, gain_init=1.0, use_layernorm=True):
  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)
  self.gain = nn.Parameter(torch.full((self.out_channels, 1, 1, 1), gain_init))
  self.scale = gamma * self.weight[0].numel() ** -0.5
  self.same_pad = is_dynamic
- self.eps = eps ** 2 if use_layernorm else eps
+ self.eps = eps
  self.use_layernorm = use_layernorm # experimental, slightly faster/less GPU memory to hijack LN kernel
 
- # NOTE an alternate formulation to consider, closer to DeepMind Haiku impl but doesn't seem
- # to make much numerical difference (+/- .002 to .004) in top-1 during eval.
- # def get_weight(self):
- # var, mean = torch.var_mean(self.weight, dim=[1, 2, 3], keepdim=True, unbiased=False)
- # scale = torch.rsqrt((self.weight[0].numel() * var).clamp_(self.eps)) * self.gain
- # weight = (self.weight - mean) * scale
- # return self.gain * weight
-
  def get_weight(self):
  if self.use_layernorm:
- weight = self.scale * F.layer_norm(self.weight, self.weight.shape[1:], eps=self.eps)
+ # NOTE F.layer_norm is being used to compute (self.weight - mean) / (sqrt(var) + self.eps) in one op
+ weight = F.layer_norm(self.weight, self.weight.shape[1:], eps=self.eps)
  else:
  std, mean = torch.std_mean(self.weight, dim=[1, 2, 3], keepdim=True, unbiased=False)
- weight = self.scale * (self.weight - mean) / (std + self.eps)
- return self.gain * weight
+ weight = (self.weight - mean) / (std + self.eps)
+ return weight.mul_(self.gain * self.scale)
 
  def forward(self, x):
  if self.same_pad:

timm/models/nfnet.py

@@ -166,6 +166,8 @@ class NfCfg:
  extra_conv: bool = False # extra 3x3 bottleneck convolution for NFNet models
  gamma_in_act: bool = False
  same_padding: bool = False
+ std_conv_eps: float = 1e-5
+ std_conv_ln: bool = True # use layer-norm impl to normalize in std-conv, works in PyTorch XLA, slightly faster
  skipinit: bool = False # disabled by default, non-trivial performance impact
  zero_init_fc: bool = False
  act_layer: str = 'silu'
@@ -482,10 +484,11 @@ def __init__(self, cfg: NfCfg, num_classes=1000, in_chans=3, global_pool='avg',
  conv_layer = ScaledStdConv2dSame if cfg.same_padding else ScaledStdConv2d
  if cfg.gamma_in_act:
  act_layer = act_with_gamma(cfg.act_layer, gamma=_nonlin_gamma[cfg.act_layer])
- conv_layer = partial(conv_layer, eps=1e-4) # DM weights better with higher eps
+ conv_layer = partial(conv_layer, eps=cfg.std_conv_eps, use_layernorm=cfg.std_conv_ln)
  else:
  act_layer = get_act_layer(cfg.act_layer)
- conv_layer = partial(conv_layer, gamma=_nonlin_gamma[cfg.act_layer])
+ conv_layer = partial(
+ conv_layer, gamma=_nonlin_gamma[cfg.act_layer], eps=cfg.std_conv_eps, use_layernorm=cfg.std_conv_ln)
  attn_layer = partial(get_attn(cfg.attn_layer), **cfg.attn_kwargs) if cfg.attn_layer else None
 
  stem_chs = make_divisible((cfg.stem_chs or cfg.channels[0]) * cfg.width_factor, cfg.ch_div)

timm/models/vision_transformer_hybrid.py

@@ -118,7 +118,7 @@ def _resnetv2(layers=(3, 4, 9), **kwargs):
  padding_same = kwargs.get('padding_same', True)
  if padding_same:
  stem_type = 'same'
- conv_layer = StdConv2dSame
+ conv_layer = partial(StdConv2dSame, eps=1e-5)
  else:
  stem_type = ''
  conv_layer = StdConv2d