Cleanup experimental vit weight init a bit

timm/models/layers/__init__.py

@@ -31,4 +31,4 @@ from .split_attn import SplitAttnConv2d
 from .split_batchnorm import SplitBatchNorm2d, convert_splitbn_model
 from .std_conv import StdConv2d, StdConv2dSame, ScaledStdConv2d, ScaledStdConv2dSame
 from .test_time_pool import TestTimePoolHead, apply_test_time_pool
-from .weight_init import trunc_normal_
+from .weight_init import trunc_normal_, variance_scaling_, lecun_normal_

timm/models/layers/weight_init.py

@@ -2,6 +2,8 @@ import torch
 import math
 import warnings
 
+from torch.nn.init import _calculate_fan_in_and_fan_out
+
 
 def _no_grad_trunc_normal_(tensor, mean, std, a, b):
     # Cut & paste from PyTorch official master until it's in a few official releases - RW
@@ -58,3 +60,30 @@ def trunc_normal_(tensor, mean=0., std=1., a=-2., b=2.):
         >>> nn.init.trunc_normal_(w)
     """
     return _no_grad_trunc_normal_(tensor, mean, std, a, b)
+
+
+def variance_scaling_(tensor, scale=1.0, mode='fan_in', distribution='normal'):
+    fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor)
+    if mode == 'fan_in':
+        denom = fan_in
+    elif mode == 'fan_out':
+        denom = fan_out
+    elif mode == 'fan_avg':
+        denom = (fan_in + fan_out) / 2
+
+    variance = scale / denom
+
+    if distribution == "truncated_normal":
+        # constant is stddev of standard normal truncated to (-2, 2)
+        trunc_normal_(tensor, std=math.sqrt(variance) / .87962566103423978)
+    elif distribution == "normal":
+        tensor.normal_(std=math.sqrt(variance))
+    elif distribution == "uniform":
+        bound = math.sqrt(3 * variance)
+        tensor.uniform_(-bound, bound)
+    else:
+        raise ValueError(f"invalid distribution {distribution}")
+
+
+def lecun_normal_(tensor):
+    variance_scaling_(tensor, mode='fan_in', distribution='truncated_normal')

timm/models/vision_transformer.py

@@ -28,7 +28,7 @@ import torch.nn.functional as F
 
 from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
 from .helpers import load_pretrained
-from .layers import StdConv2dSame, StdConv2d, DropPath, to_2tuple, trunc_normal_
+from .layers import StdConv2dSame, StdConv2d, DropPath, to_2tuple, trunc_normal_, lecun_normal_
 from .resnet import resnet26d, resnet50d
 from .resnetv2 import ResNetV2, create_resnetv2_stem
 from .registry import register_model
@@ -373,7 +373,7 @@ class VisionTransformer(nn.Module):
     def __init__(self, img_size=224, patch_size=None, in_chans=3, num_classes=1000, embed_dim=768, depth=12,
                  num_heads=12, mlp_ratio=4., qkv_bias=True, qk_scale=None, representation_size=None,
                  drop_rate=0., attn_drop_rate=0., drop_path_rate=0., hybrid_backbone=None, norm_layer=None,
-                 act_layer=None, weight_init=''):
+                 act_layer=None, weight_init='new_nlhb'):
         """
         Args:
             img_size (int, tuple): input image size
@@ -433,14 +433,20 @@ class VisionTransformer(nn.Module):
         # Classifier head
         self.head = nn.Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity()
 
+        self._init_weights(weight_init)
+
+    def _init_weights(self, weight_init: str):
         trunc_normal_(self.pos_embed, std=.02)
-        if weight_init != 'jax':  # leave as zeros to match JAX impl
+        if weight_init.startswith('jax'):
-            trunc_normal_(self.cls_token, std=.02)
+            init_fn = _init_weights_jax
-        for n, m in self.named_modules():
+            # leave cls token as zeros to match jax impl
-            if weight_init == 'jax':
-                _init_weights_jax(m, n)
         else:
-                _init_weights_original(m, n)
+            trunc_normal_(self.cls_token, std=.02)
+            init_fn = _init_weights_new if weight_init.startswith('new') else _init_weights_old
+        hb = -math.log(self.num_classes) if 'nlhb' in weight_init else 0.
+        init_fn = partial(init_fn, head_bias=hb)
+        for n, m in self.named_modules():
+            init_fn(m, n)
 
     @torch.jit.ignore
     def no_weight_decay(self):
@@ -475,41 +481,42 @@ class VisionTransformer(nn.Module):
         return x
 
 
-def _init_weights_original(m: nn.Module, n: str = ''):
+def _init_weights_old(m: nn.Module, n: str = '', head_bias: float = 0.):
-    if isinstance(m, (nn.Conv2d, nn.Linear)):
+    if isinstance(m, nn.Linear):
         trunc_normal_(m.weight, std=.02)
-        if isinstance(m, nn.Linear) and m.bias is not None:
+        if m.bias is not None:
-            nn.init.constant_(m.bias, 0)
+            if 'head' in n:
+                nn.init.constant_(m.bias, head_bias)
+            else:
+                nn.init.zeros_(m.bias)
     elif isinstance(m, nn.LayerNorm):
         nn.init.zeros_(m.bias)
         nn.init.ones_(m.weight)
 
 
-def _init_weights_jax(m: nn.Module, n: str):
+def _init_weights_new(m: nn.Module, n: str = '', head_bias: float = 0.):
-    """ Weight init scheme closer to the official JAX impl than my original init"""
+    if isinstance(m, (nn.Conv2d, nn.Linear)):
+        #trunc_normal_(m.weight, std=.02)
+        lecun_normal_(m.weight)
+        if m.bias is not None:
+            if 'head' in n:
+                nn.init.constant_(m.bias, head_bias)
+            else:
+                nn.init.zeros_(m.bias)
+    elif isinstance(m, nn.LayerNorm):
+        nn.init.zeros_(m.bias)
+        nn.init.ones_(m.weight)
 
-    def _fan_in(tensor):
-        dimensions = tensor.dim()
-        if dimensions < 2:
-            raise ValueError("Fan in and fan out can not be computed for tensor with fewer than 2 dimensions")
 
-        num_input_fmaps = tensor.size(1)
+def _init_weights_jax(m: nn.Module, n: str, head_bias: float = 0.):
-        receptive_field_size = 1
+    """ Attempt at weight init scheme closer to the official JAX impl than my original init"""
-        if tensor.dim() > 2:
-            receptive_field_size = tensor[0][0].numel()
-        fan_in = num_input_fmaps * receptive_field_size
-        return fan_in
-
-    def _lecun_normal(w):
-        stddev = (1.0 / _fan_in(w)) ** 0.5 / .87962566103423978
-        trunc_normal_(w, 0, stddev)
 
     if isinstance(m, nn.Linear):
         if 'head' in n:
             nn.init.zeros_(m.weight)
-            nn.init.zeros_(m.bias)
+            nn.init.constant_(m.bias, head_bias)
         elif 'pre_logits' in n:
-            _lecun_normal(m.weight)
+            lecun_normal_(m.weight)
             nn.init.zeros_(m.bias)
         else:
             nn.init.xavier_uniform_(m.weight)
@@ -519,7 +526,7 @@ def _init_weights_jax(m: nn.Module, n: str):
                 else:
                     nn.init.zeros_(m.bias)
     elif isinstance(m, nn.Conv2d):
-        _lecun_normal(m.weight)
+        lecun_normal_(m.weight)
         if m.bias is not None:
             nn.init.zeros_(m.bias)
     elif isinstance(m, nn.LayerNorm):
@@ -544,7 +551,7 @@ class DistilledVisionTransformer(VisionTransformer):
 
         trunc_normal_(self.dist_token, std=.02)
         trunc_normal_(self.pos_embed, std=.02)
-        self.head_dist.apply(_init_weights_original)
+        self.head_dist.apply(_init_weights_new)
 
     def forward_features(self, x):
         B = x.shape[0]