epsilon should be inside square root

README.md

@@ -542,7 +542,7 @@ nest = NesT(
     dim = 96,
     heads = 3,
     num_hierarchies = 3,        # number of hierarchies
-    block_repeats = (8, 4, 1),  # the number of transformer blocks at each heirarchy, starting from the bottom
+    block_repeats = (2, 2, 8),  # the number of transformer blocks at each heirarchy, starting from the bottom
     num_classes = 1000
 )
 

setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'vit-pytorch',
   packages = find_packages(exclude=['examples']),
-  version = '0.26.2',
+  version = '0.26.5',
   license='MIT',
   description = 'Vision Transformer (ViT) - Pytorch',
   author = 'Phil Wang',

vit_pytorch/crossformer.py

@@ -62,9 +62,9 @@ class LayerNorm(nn.Module):
         self.b = nn.Parameter(torch.zeros(1, dim, 1, 1))
 
     def forward(self, x):
-        std = torch.var(x, dim = 1, unbiased = False, keepdim = True).sqrt()
+        var = torch.var(x, dim = 1, unbiased = False, keepdim = True)
         mean = torch.mean(x, dim = 1, keepdim = True)
-        return (x - mean) / (std + self.eps) * self.g + self.b
+        return (x - mean) / (var + self.eps).sqrt() * self.g + self.b
 
 def FeedForward(dim, mult = 4, dropout = 0.):
     return nn.Sequential(

vit_pytorch/cvt.py

@@ -30,9 +30,9 @@ class LayerNorm(nn.Module): # layernorm, but done in the channel dimension #1
         self.b = nn.Parameter(torch.zeros(1, dim, 1, 1))
 
     def forward(self, x):
-        std = torch.var(x, dim = 1, unbiased = False, keepdim = True).sqrt()
+        var = torch.var(x, dim = 1, unbiased = False, keepdim = True)
         mean = torch.mean(x, dim = 1, keepdim = True)
-        return (x - mean) / (std + self.eps) * self.g + self.b
+        return (x - mean) / (var + self.eps).sqrt() * self.g + self.b
 
 class PreNorm(nn.Module):
     def __init__(self, dim, fn):

vit_pytorch/mae.py

@@ -14,13 +14,11 @@ class MAE(nn.Module):
         masking_ratio = 0.75,
         decoder_depth = 1,
         decoder_heads = 8,
-        decoder_dim_head = 64,
-        apply_decoder_pos_emb_all = False # whether to (re)apply decoder positional embedding to encoder unmasked tokens
+        decoder_dim_head = 64
     ):
         super().__init__()
         assert masking_ratio > 0 and masking_ratio < 1, 'masking ratio must be kept between 0 and 1'
         self.masking_ratio = masking_ratio
-        self.apply_decoder_pos_emb_all = apply_decoder_pos_emb_all
 
         # extract some hyperparameters and functions from encoder (vision transformer to be trained)
 
@@ -73,10 +71,9 @@ class MAE(nn.Module):
 
         decoder_tokens = self.enc_to_dec(encoded_tokens)
 
-        # reapply decoder position embedding to unmasked tokens, if desired
+        # reapply decoder position embedding to unmasked tokens
 
-        if self.apply_decoder_pos_emb_all:
-            decoder_tokens = decoder_tokens + self.decoder_pos_emb(unmasked_indices)
+        decoder_tokens = decoder_tokens + self.decoder_pos_emb(unmasked_indices)
 
         # repeat mask tokens for number of masked, and add the positions using the masked indices derived above
 

vit_pytorch/nest.py

@@ -20,9 +20,9 @@ class LayerNorm(nn.Module):
         self.b = nn.Parameter(torch.zeros(1, dim, 1, 1))
 
     def forward(self, x):
-        std = torch.var(x, dim = 1, unbiased = False, keepdim = True).sqrt()
+        var = torch.var(x, dim = 1, unbiased = False, keepdim = True)
         mean = torch.mean(x, dim = 1, keepdim = True)
-        return (x - mean) / (std + self.eps) * self.g + self.b
+        return (x - mean) / (var + self.eps).sqrt() * self.g + self.b
 
 class PreNorm(nn.Module):
     def __init__(self, dim, fn):
@@ -131,10 +131,11 @@ class NesT(nn.Module):
 
         seq_len = (fmap_size // blocks) ** 2   # sequence length is held constant across heirarchy
         hierarchies = list(reversed(range(num_hierarchies)))
-        mults = [2 ** i for i in hierarchies]
+        mults = [2 ** i for i in reversed(hierarchies)]
 
         layer_heads = list(map(lambda t: t * heads, mults))
         layer_dims = list(map(lambda t: t * dim, mults))
+        last_dim = layer_dims[-1]
 
         layer_dims = [*layer_dims, layer_dims[-1]]
         dim_pairs = zip(layer_dims[:-1], layer_dims[1:])
@@ -157,10 +158,11 @@ class NesT(nn.Module):
                 Aggregate(dim_in, dim_out) if not is_last else nn.Identity()
             ]))
 
+
         self.mlp_head = nn.Sequential(
-            LayerNorm(dim),
+            LayerNorm(last_dim),
             Reduce('b c h w -> b c', 'mean'),
-            nn.Linear(dim, num_classes)
+            nn.Linear(last_dim, num_classes)
         )
 
     def forward(self, img):

vit_pytorch/twins_svt.py

@@ -38,9 +38,9 @@ class LayerNorm(nn.Module):
         self.b = nn.Parameter(torch.zeros(1, dim, 1, 1))
 
     def forward(self, x):
-        std = torch.var(x, dim = 1, unbiased = False, keepdim = True).sqrt()
+        var = torch.var(x, dim = 1, unbiased = False, keepdim = True)
         mean = torch.mean(x, dim = 1, keepdim = True)
-        return (x - mean) / (std + self.eps) * self.g + self.b
+        return (x - mean) / (var + self.eps).sqrt() * self.g + self.b
 
 class PreNorm(nn.Module):
     def __init__(self, dim, fn):