add simple vit with register tokens example, cite

finish core idea of NaViT

README.md

@@ -7,6 +7,7 @@
 - [Usage](#usage)
 - [Parameters](#parameters)
 - [Simple ViT](#simple-vit)
+- [NaViT](#navit)
 - [Distillation](#distillation)
 - [Deep ViT](#deep-vit)
 - [CaiT](#cait)
@@ -139,6 +140,63 @@ img = torch.randn(1, 3, 256, 256)
 preds = v(img) # (1, 1000)
 ```
 
+## NaViT
+
+<img src="./images/navit.png" width="450px"></img>
+
+<a href="https://arxiv.org/abs/2307.06304">This paper</a> proposes to leverage the flexibility of attention and masking for variable lengthed sequences to train images of multiple resolution, packed into a single batch. They demonstrate much faster training and improved accuracies, with the only cost being extra complexity in the architecture and dataloading. They use factorized 2d positional encodings, token dropping, as well as query-key normalization.
+
+You can use it as follows
+
+```python
+import torch
+from vit_pytorch.na_vit import NaViT
+
+v = NaViT(
+    image_size = 256,
+    patch_size = 32,
+    num_classes = 1000,
+    dim = 1024,
+    depth = 6,
+    heads = 16,
+    mlp_dim = 2048,
+    dropout = 0.1,
+    emb_dropout = 0.1,
+    token_dropout_prob = 0.1  # token dropout of 10% (keep 90% of tokens)
+)
+
+# 5 images of different resolutions - List[List[Tensor]]
+
+# for now, you'll have to correctly place images in same batch element as to not exceed maximum allowed sequence length for self-attention w/ masking
+
+images = [
+    [torch.randn(3, 256, 256), torch.randn(3, 128, 128)],
+    [torch.randn(3, 128, 256), torch.randn(3, 256, 128)],
+    [torch.randn(3, 64, 256)]
+]
+
+preds = v(images) # (5, 1000) - 5, because 5 images of different resolution above
+
+```
+
+Or if you would rather that the framework auto group the images into variable lengthed sequences that do not exceed a certain max length
+
+```python
+images = [
+    torch.randn(3, 256, 256),
+    torch.randn(3, 128, 128),
+    torch.randn(3, 128, 256),
+    torch.randn(3, 256, 128),
+    torch.randn(3, 64, 256)
+]
+
+preds = v(
+    images,
+    group_images = True,
+    group_max_seq_len = 64
+) # (5, 1000)
+```
+
 ## Distillation
 
 <img src="./images/distill.png" width="300px"></img>
@@ -1934,6 +1992,14 @@ Coming from computer vision and new to transformers? Here are some resources tha
 }
 ```
 
+```bibtex
+@inproceedings{Dehghani2023PatchNP,
+    title   = {Patch n' Pack: NaViT, a Vision Transformer for any Aspect Ratio and Resolution},
+    author  = {Mostafa Dehghani and Basil Mustafa and Josip Djolonga and Jonathan Heek and Matthias Minderer and Mathilde Caron and Andreas Steiner and Joan Puigcerver and Robert Geirhos and Ibrahim M. Alabdulmohsin and Avital Oliver and Piotr Padlewski and Alexey A. Gritsenko and Mario Luvci'c and Neil Houlsby},
+    year    = {2023}
+}
+```
+
 ```bibtex
 @misc{vaswani2017attention,
     title   = {Attention Is All You Need},
@@ -1954,4 +2020,13 @@ Coming from computer vision and new to transformers? Here are some resources tha
 }
 ```
 
+```bibtex
+@inproceedings{Darcet2023VisionTN,
+    title   = {Vision Transformers Need Registers},
+    author  = {Timoth'ee Darcet and Maxime Oquab and Julien Mairal and Piotr Bojanowski},
+    year    = {2023},
+    url     = {https://api.semanticscholar.org/CorpusID:263134283}
+}
+```
+
 *I visualise a time when we will be to robots what dogs are to humans, and I’m rooting for the machines.* — Claude Shannon

setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'vit-pytorch',
   packages = find_packages(exclude=['examples']),
-  version = '1.2.1',
+  version = '1.5.0 ',
   license='MIT',
   description = 'Vision Transformer (ViT) - Pytorch',
   long_description_content_type = 'text/markdown',

vit_pytorch/ats_vit.py

@@ -110,18 +110,11 @@ class AdaptiveTokenSampling(nn.Module):
 
 # classes
 
-class PreNorm(nn.Module):
-    def __init__(self, dim, fn):
-        super().__init__()
-        self.norm = nn.LayerNorm(dim)
-        self.fn = fn
-    def forward(self, x, **kwargs):
-        return self.fn(self.norm(x), **kwargs)
-
 class FeedForward(nn.Module):
     def __init__(self, dim, hidden_dim, dropout = 0.):
         super().__init__()
         self.net = nn.Sequential(
+            nn.LayerNorm(dim),
             nn.Linear(dim, hidden_dim),
             nn.GELU(),
             nn.Dropout(dropout),
@@ -138,6 +131,7 @@ class Attention(nn.Module):
         self.heads = heads
         self.scale = dim_head ** -0.5
 
+        self.norm = nn.LayerNorm(dim)
         self.attend = nn.Softmax(dim = -1)
         self.dropout = nn.Dropout(dropout)
 
@@ -154,6 +148,7 @@ class Attention(nn.Module):
     def forward(self, x, *, mask):
         num_tokens = x.shape[1]
 
+        x = self.norm(x)
         qkv = self.to_qkv(x).chunk(3, dim = -1)
         q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = self.heads), qkv)
 
@@ -189,8 +184,8 @@ class Transformer(nn.Module):
         self.layers = nn.ModuleList([])
         for _, output_num_tokens in zip(range(depth), max_tokens_per_depth):
             self.layers.append(nn.ModuleList([
-                PreNorm(dim, Attention(dim, output_num_tokens = output_num_tokens, heads = heads, dim_head = dim_head, dropout = dropout)),
-                PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout))
+                Attention(dim, output_num_tokens = output_num_tokens, heads = heads, dim_head = dim_head, dropout = dropout),
+                FeedForward(dim, mlp_dim, dropout = dropout)
             ]))
 
     def forward(self, x):

vit_pytorch/cait.py

@@ -44,18 +44,11 @@ class LayerScale(nn.Module):
     def forward(self, x, **kwargs):
         return self.fn(x, **kwargs) * self.scale
 
-class PreNorm(nn.Module):
-    def __init__(self, dim, fn):
-        super().__init__()
-        self.norm = nn.LayerNorm(dim)
-        self.fn = fn
-    def forward(self, x, **kwargs):
-        return self.fn(self.norm(x), **kwargs)
-
 class FeedForward(nn.Module):
     def __init__(self, dim, hidden_dim, dropout = 0.):
         super().__init__()
         self.net = nn.Sequential(
+            nn.LayerNorm(dim),
             nn.Linear(dim, hidden_dim),
             nn.GELU(),
             nn.Dropout(dropout),
@@ -72,6 +65,7 @@ class Attention(nn.Module):
         self.heads = heads
         self.scale = dim_head ** -0.5
 
+        self.norm = nn.LayerNorm(dim)
         self.to_q = nn.Linear(dim, inner_dim, bias = False)
         self.to_kv = nn.Linear(dim, inner_dim * 2, bias = False)
 
@@ -89,6 +83,7 @@ class Attention(nn.Module):
     def forward(self, x, context = None):
         b, n, _, h = *x.shape, self.heads
 
+        x = self.norm(x)
         context = x if not exists(context) else torch.cat((x, context), dim = 1)
 
         qkv = (self.to_q(x), *self.to_kv(context).chunk(2, dim = -1))
@@ -115,8 +110,8 @@ class Transformer(nn.Module):
 
         for ind in range(depth):
             self.layers.append(nn.ModuleList([
-                LayerScale(dim, PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout)), depth = ind + 1),
-                LayerScale(dim, PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout)), depth = ind + 1)
+                LayerScale(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout), depth = ind + 1),
+                LayerScale(dim, FeedForward(dim, mlp_dim, dropout = dropout), depth = ind + 1)
             ]))
     def forward(self, x, context = None):
         layers = dropout_layers(self.layers, dropout = self.layer_dropout)

vit_pytorch/cross_vit.py

@@ -13,22 +13,13 @@ def exists(val):
 def default(val, d):
     return val if exists(val) else d
 
-# pre-layernorm
-
-class PreNorm(nn.Module):
-    def __init__(self, dim, fn):
-        super().__init__()
-        self.norm = nn.LayerNorm(dim)
-        self.fn = fn
-    def forward(self, x, **kwargs):
-        return self.fn(self.norm(x), **kwargs)
-
 # feedforward
 
 class FeedForward(nn.Module):
     def __init__(self, dim, hidden_dim, dropout = 0.):
         super().__init__()
         self.net = nn.Sequential(
+            nn.LayerNorm(dim),
             nn.Linear(dim, hidden_dim),
             nn.GELU(),
             nn.Dropout(dropout),
@@ -47,6 +38,7 @@ class Attention(nn.Module):
         self.heads = heads
         self.scale = dim_head ** -0.5
 
+        self.norm = nn.LayerNorm(dim)
         self.attend = nn.Softmax(dim = -1)
         self.dropout = nn.Dropout(dropout)
 
@@ -60,6 +52,7 @@ class Attention(nn.Module):
 
     def forward(self, x, context = None, kv_include_self = False):
         b, n, _, h = *x.shape, self.heads
+        x = self.norm(x)
         context = default(context, x)
 
         if kv_include_self:
@@ -86,8 +79,8 @@ class Transformer(nn.Module):
         self.norm = nn.LayerNorm(dim)
         for _ in range(depth):
             self.layers.append(nn.ModuleList([
-                PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout)),
-                PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout))
+                Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout),
+                FeedForward(dim, mlp_dim, dropout = dropout)
             ]))
 
     def forward(self, x):
@@ -121,8 +114,8 @@ class CrossTransformer(nn.Module):
         self.layers = nn.ModuleList([])
         for _ in range(depth):
             self.layers.append(nn.ModuleList([
-                ProjectInOut(sm_dim, lg_dim, PreNorm(lg_dim, Attention(lg_dim, heads = heads, dim_head = dim_head, dropout = dropout))),
-                ProjectInOut(lg_dim, sm_dim, PreNorm(sm_dim, Attention(sm_dim, heads = heads, dim_head = dim_head, dropout = dropout)))
+                ProjectInOut(sm_dim, lg_dim, Attention(lg_dim, heads = heads, dim_head = dim_head, dropout = dropout)),
+                ProjectInOut(lg_dim, sm_dim, Attention(sm_dim, heads = heads, dim_head = dim_head, dropout = dropout))
             ]))
 
     def forward(self, sm_tokens, lg_tokens):

vit_pytorch/cvt.py

@@ -34,19 +34,11 @@ class LayerNorm(nn.Module): # layernorm, but done in the channel dimension #1
         mean = torch.mean(x, dim = 1, keepdim = True)
         return (x - mean) / (var + self.eps).sqrt() * self.g + self.b
 
-class PreNorm(nn.Module):
-    def __init__(self, dim, fn):
-        super().__init__()
-        self.norm = LayerNorm(dim)
-        self.fn = fn
-    def forward(self, x, **kwargs):
-        x = self.norm(x)
-        return self.fn(x, **kwargs)
-
 class FeedForward(nn.Module):
     def __init__(self, dim, mult = 4, dropout = 0.):
         super().__init__()
         self.net = nn.Sequential(
+            LayerNorm(dim),
             nn.Conv2d(dim, dim * mult, 1),
             nn.GELU(),
             nn.Dropout(dropout),
@@ -75,6 +67,7 @@ class Attention(nn.Module):
         self.heads = heads
         self.scale = dim_head ** -0.5
 
+        self.norm = LayerNorm(dim)
         self.attend = nn.Softmax(dim = -1)
         self.dropout = nn.Dropout(dropout)
 
@@ -89,6 +82,8 @@ class Attention(nn.Module):
     def forward(self, x):
         shape = x.shape
         b, n, _, y, h = *shape, self.heads
+
+        x = self.norm(x)
         q, k, v = (self.to_q(x), *self.to_kv(x).chunk(2, dim = 1))
         q, k, v = map(lambda t: rearrange(t, 'b (h d) x y -> (b h) (x y) d', h = h), (q, k, v))
 
@@ -107,8 +102,8 @@ class Transformer(nn.Module):
         self.layers = nn.ModuleList([])
         for _ in range(depth):
             self.layers.append(nn.ModuleList([
-                PreNorm(dim, Attention(dim, proj_kernel = proj_kernel, kv_proj_stride = kv_proj_stride, heads = heads, dim_head = dim_head, dropout = dropout)),
-                PreNorm(dim, FeedForward(dim, mlp_mult, dropout = dropout))
+                Attention(dim, proj_kernel = proj_kernel, kv_proj_stride = kv_proj_stride, heads = heads, dim_head = dim_head, dropout = dropout),
+                FeedForward(dim, mlp_mult, dropout = dropout)
             ]))
     def forward(self, x):
         for attn, ff in self.layers:

vit_pytorch/deepvit.py

@@ -5,25 +5,11 @@ import torch.nn.functional as F
 from einops import rearrange, repeat
 from einops.layers.torch import Rearrange
 
-class Residual(nn.Module):
-    def __init__(self, fn):
-        super().__init__()
-        self.fn = fn
-    def forward(self, x, **kwargs):
-        return self.fn(x, **kwargs) + x
-
-class PreNorm(nn.Module):
-    def __init__(self, dim, fn):
-        super().__init__()
-        self.norm = nn.LayerNorm(dim)
-        self.fn = fn
-    def forward(self, x, **kwargs):
-        return self.fn(self.norm(x), **kwargs)
-
 class FeedForward(nn.Module):
     def __init__(self, dim, hidden_dim, dropout = 0.):
         super().__init__()
         self.net = nn.Sequential(
+            nn.LayerNorm(dim),
             nn.Linear(dim, hidden_dim),
             nn.GELU(),
             nn.Dropout(dropout),
@@ -40,6 +26,7 @@ class Attention(nn.Module):
         self.heads = heads
         self.scale = dim_head ** -0.5
 
+        self.norm = nn.LayerNorm(dim)
         self.to_qkv = nn.Linear(dim, inner_dim * 3, bias = False)
 
         self.dropout = nn.Dropout(dropout)
@@ -59,6 +46,8 @@ class Attention(nn.Module):
 
     def forward(self, x):
         b, n, _, h = *x.shape, self.heads
+        x = self.norm(x)
+
         qkv = self.to_qkv(x).chunk(3, dim = -1)
         q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = h), qkv)
 
@@ -86,13 +75,13 @@ class Transformer(nn.Module):
         self.layers = nn.ModuleList([])
         for _ in range(depth):
             self.layers.append(nn.ModuleList([
-                Residual(PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout))),
-                Residual(PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout)))
+                Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout),
+                FeedForward(dim, mlp_dim, dropout = dropout)
             ]))
     def forward(self, x):
         for attn, ff in self.layers:
-            x = attn(x)
-            x = ff(x)
+            x = attn(x) + x
+            x = ff(x) + x
         return x
 
 class DeepViT(nn.Module):

vit_pytorch/local_vit.py

@@ -26,16 +26,6 @@ class ExcludeCLS(nn.Module):
         x = self.fn(x, **kwargs)
         return torch.cat((cls_token, x), dim = 1)
 
-# prenorm
-
-class PreNorm(nn.Module):
-    def __init__(self, dim, fn):
-        super().__init__()
-        self.norm = nn.LayerNorm(dim)
-        self.fn = fn
-    def forward(self, x, **kwargs):
-        return self.fn(self.norm(x), **kwargs)
-
 # feed forward related classes
 
 class DepthWiseConv2d(nn.Module):
@@ -52,6 +42,7 @@ class FeedForward(nn.Module):
     def __init__(self, dim, hidden_dim, dropout = 0.):
         super().__init__()
         self.net = nn.Sequential(
+            nn.LayerNorm(dim),
             nn.Conv2d(dim, hidden_dim, 1),
             nn.Hardswish(),
             DepthWiseConv2d(hidden_dim, hidden_dim, 3, padding = 1),
@@ -77,6 +68,7 @@ class Attention(nn.Module):
         self.heads = heads
         self.scale = dim_head ** -0.5
 
+        self.norm = nn.LayerNorm(dim)
         self.attend = nn.Softmax(dim = -1)
         self.dropout = nn.Dropout(dropout)
         self.to_qkv = nn.Linear(dim, inner_dim * 3, bias = False)
@@ -88,6 +80,8 @@ class Attention(nn.Module):
 
     def forward(self, x):
         b, n, _, h = *x.shape, self.heads
+
+        x = self.norm(x)
         qkv = self.to_qkv(x).chunk(3, dim = -1)
         q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = h), qkv)
 
@@ -106,8 +100,8 @@ class Transformer(nn.Module):
         self.layers = nn.ModuleList([])
         for _ in range(depth):
             self.layers.append(nn.ModuleList([
-                Residual(PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout))),
-                ExcludeCLS(Residual(PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout))))
+                Residual(Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout)),
+                ExcludeCLS(Residual(FeedForward(dim, mlp_dim, dropout = dropout)))
             ]))
     def forward(self, x):
         for attn, ff in self.layers:

vit_pytorch/mae.py

@@ -49,7 +49,10 @@ class MAE(nn.Module):
         # patch to encoder tokens and add positions
 
         tokens = self.patch_to_emb(patches)
-        tokens = tokens + self.encoder.pos_embedding[:, 1:(num_patches + 1)]
+        if self.encoder.pool == "cls":
+            tokens += self.encoder.pos_embedding[:, 1:(num_patches + 1)]
+        elif self.encoder.pool == "mean":
+            tokens += self.encoder.pos_embedding.to(device, dtype=tokens.dtype) 
 
         # calculate of patches needed to be masked, and get random indices, dividing it up for mask vs unmasked
 

vit_pytorch/max_vit.py

@@ -19,20 +19,20 @@ def cast_tuple(val, length = 1):
 
 # helper classes
 
-class PreNormResidual(nn.Module):
+class Residual(nn.Module):
     def __init__(self, dim, fn):
         super().__init__()
-        self.norm = nn.LayerNorm(dim)
         self.fn = fn
 
     def forward(self, x):
-        return self.fn(self.norm(x)) + x
+        return self.fn(x) + x
 
 class FeedForward(nn.Module):
     def __init__(self, dim, mult = 4, dropout = 0.):
         super().__init__()
         inner_dim = int(dim * mult)
         self.net = nn.Sequential(
+            nn.LayerNorm(dim),
             nn.Linear(dim, inner_dim),
             nn.GELU(),
             nn.Dropout(dropout),
@@ -132,6 +132,7 @@ class Attention(nn.Module):
         self.heads = dim // dim_head
         self.scale = dim_head ** -0.5
 
+        self.norm = nn.LayerNorm(dim)
         self.to_qkv = nn.Linear(dim, dim * 3, bias = False)
 
         self.attend = nn.Sequential(
@@ -160,6 +161,8 @@ class Attention(nn.Module):
     def forward(self, x):
         batch, height, width, window_height, window_width, _, device, h = *x.shape, x.device, self.heads
 
+        x = self.norm(x)
+
         # flatten
 
         x = rearrange(x, 'b x y w1 w2 d -> (b x y) (w1 w2) d')
@@ -259,13 +262,13 @@ class MaxViT(nn.Module):
                         shrinkage_rate = mbconv_shrinkage_rate
                     ),
                     Rearrange('b d (x w1) (y w2) -> b x y w1 w2 d', w1 = w, w2 = w),  # block-like attention
-                    PreNormResidual(layer_dim, Attention(dim = layer_dim, dim_head = dim_head, dropout = dropout, window_size = w)),
-                    PreNormResidual(layer_dim, FeedForward(dim = layer_dim, dropout = dropout)),
+                    Residual(layer_dim, Attention(dim = layer_dim, dim_head = dim_head, dropout = dropout, window_size = w)),
+                    Residual(layer_dim, FeedForward(dim = layer_dim, dropout = dropout)),
                     Rearrange('b x y w1 w2 d -> b d (x w1) (y w2)'),
 
                     Rearrange('b d (w1 x) (w2 y) -> b x y w1 w2 d', w1 = w, w2 = w),  # grid-like attention
-                    PreNormResidual(layer_dim, Attention(dim = layer_dim, dim_head = dim_head, dropout = dropout, window_size = w)),
-                    PreNormResidual(layer_dim, FeedForward(dim = layer_dim, dropout = dropout)),
+                    Residual(layer_dim, Attention(dim = layer_dim, dim_head = dim_head, dropout = dropout, window_size = w)),
+                    Residual(layer_dim, FeedForward(dim = layer_dim, dropout = dropout)),
                     Rearrange('b x y w1 w2 d -> b d (w1 x) (w2 y)'),
                 )
 

vit_pytorch/mobile_vit.py

@@ -22,20 +22,11 @@ def conv_nxn_bn(inp, oup, kernel_size=3, stride=1):
 
 # classes
 
-class PreNorm(nn.Module):
-    def __init__(self, dim, fn):
-        super().__init__()
-        self.norm = nn.LayerNorm(dim)
-        self.fn = fn
-
-    def forward(self, x, **kwargs):
-        return self.fn(self.norm(x), **kwargs)
-
-
 class FeedForward(nn.Module):
     def __init__(self, dim, hidden_dim, dropout=0.):
         super().__init__()
         self.net = nn.Sequential(
+            nn.LayerNorm(dim),
             nn.Linear(dim, hidden_dim),
             nn.SiLU(),
             nn.Dropout(dropout),
@@ -53,6 +44,7 @@ class Attention(nn.Module):
         self.heads = heads
         self.scale = dim_head ** -0.5
 
+        self.norm = nn.LayerNorm(dim)
         self.attend = nn.Softmax(dim=-1)
         self.dropout = nn.Dropout(dropout)
 
@@ -64,9 +56,10 @@ class Attention(nn.Module):
         )
 
     def forward(self, x):
+        x = self.norm(x)
         qkv = self.to_qkv(x).chunk(3, dim=-1)
-        q, k, v = map(lambda t: rearrange(
-            t, 'b p n (h d) -> b p h n d', h=self.heads), qkv)
+
+        q, k, v = map(lambda t: rearrange(t, 'b p n (h d) -> b p h n d', h=self.heads), qkv)
 
         dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale
 
@@ -88,8 +81,8 @@ class Transformer(nn.Module):
         self.layers = nn.ModuleList([])
         for _ in range(depth):
             self.layers.append(nn.ModuleList([
-                PreNorm(dim, Attention(dim, heads, dim_head, dropout)),
-                PreNorm(dim, FeedForward(dim, mlp_dim, dropout))
+                Attention(dim, heads, dim_head, dropout),
+                FeedForward(dim, mlp_dim, dropout)
             ]))
 
     def forward(self, x):
@@ -167,11 +160,9 @@ class MobileViTBlock(nn.Module):
 
         # Global representations
         _, _, h, w = x.shape
-        x = rearrange(x, 'b d (h ph) (w pw) -> b (ph pw) (h w) d',
-                      ph=self.ph, pw=self.pw)
-        x = self.transformer(x)
-        x = rearrange(x, 'b (ph pw) (h w) d -> b d (h ph) (w pw)',
-                      h=h//self.ph, w=w//self.pw, ph=self.ph, pw=self.pw)
+        x = rearrange(x, 'b d (h ph) (w pw) -> b (ph pw) (h w) d', ph=self.ph, pw=self.pw)
+        x = self.transformer(x)        
+        x = rearrange(x, 'b (ph pw) (h w) d -> b d (h ph) (w pw)', h=h//self.ph, w=w//self.pw, ph=self.ph, pw=self.pw)
 
         # Fusion
         x = self.conv3(x)

vit_pytorch/mpp.py

@@ -96,6 +96,9 @@ class MPP(nn.Module):
         self.loss = MPPLoss(patch_size, channels, output_channel_bits,
                             max_pixel_val, mean, std)
 
+        # extract patching function
+        self.patch_to_emb = nn.Sequential(transformer.to_patch_embedding[1:])
+
         # output transformation
         self.to_bits = nn.Linear(dim, 2**(output_channel_bits * channels))
 
@@ -151,7 +154,7 @@ class MPP(nn.Module):
         masked_input[bool_mask_replace] = self.mask_token
 
         # linear embedding of patches
-        masked_input = transformer.to_patch_embedding[-1](masked_input)
+        masked_input = self.patch_to_emb(masked_input)
 
         # add cls token to input sequence
         b, n, _ = masked_input.shape

vit_pytorch/na_vit.py

@@ -0,0 +1,389 @@
+from functools import partial
+from typing import List, Union
+
+import torch
+import torch.nn.functional as F
+from torch import nn, Tensor
+from torch.nn.utils.rnn import pad_sequence as orig_pad_sequence
+
+from einops import rearrange, repeat
+from einops.layers.torch import Rearrange
+
+# helpers
+
+def exists(val):
+    return val is not None
+
+def default(val, d):
+    return val if exists(val) else d
+
+def always(val):
+    return lambda *args: val
+
+def pair(t):
+    return t if isinstance(t, tuple) else (t, t)
+
+def divisible_by(numer, denom):
+    return (numer % denom) == 0
+
+# auto grouping images
+
+def group_images_by_max_seq_len(
+    images: List[Tensor],
+    patch_size: int,
+    calc_token_dropout = None,
+    max_seq_len = 2048
+
+) -> List[List[Tensor]]:
+
+    calc_token_dropout = default(calc_token_dropout, always(0.))
+
+    groups = []
+    group = []
+    seq_len = 0
+
+    if isinstance(calc_token_dropout, (float, int)):
+        calc_token_dropout = always(calc_token_dropout)
+
+    for image in images:
+        assert isinstance(image, Tensor)
+
+        image_dims = image.shape[-2:]
+        ph, pw = map(lambda t: t // patch_size, image_dims)
+
+        image_seq_len = (ph * pw)
+        image_seq_len = int(image_seq_len * (1 - calc_token_dropout(*image_dims)))
+
+        assert image_seq_len <= max_seq_len, f'image with dimensions {image_dims} exceeds maximum sequence length'
+
+        if (seq_len + image_seq_len) > max_seq_len:
+            groups.append(group)
+            group = []
+            seq_len = 0
+
+        group.append(image)
+        seq_len += image_seq_len
+
+    if len(group) > 0:
+        groups.append(group)
+
+    return groups
+
+# normalization
+# they use layernorm without bias, something that pytorch does not offer
+
+class LayerNorm(nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.gamma = nn.Parameter(torch.ones(dim))
+        self.register_buffer('beta', torch.zeros(dim))
+
+    def forward(self, x):
+        return F.layer_norm(x, x.shape[-1:], self.gamma, self.beta)
+
+# they use a query-key normalization that is equivalent to rms norm (no mean-centering, learned gamma), from vit 22B paper
+
+class RMSNorm(nn.Module):
+    def __init__(self, heads, dim):
+        super().__init__()
+        self.scale = dim ** 0.5
+        self.gamma = nn.Parameter(torch.ones(heads, 1, dim))
+
+    def forward(self, x):
+        normed = F.normalize(x, dim = -1)
+        return normed * self.scale * self.gamma
+
+# feedforward
+
+def FeedForward(dim, hidden_dim, dropout = 0.):
+    return nn.Sequential(
+        LayerNorm(dim),
+        nn.Linear(dim, hidden_dim),
+        nn.GELU(),
+        nn.Dropout(dropout),
+        nn.Linear(hidden_dim, dim),
+        nn.Dropout(dropout)
+    )
+
+class Attention(nn.Module):
+    def __init__(self, dim, heads = 8, dim_head = 64, dropout = 0.):
+        super().__init__()
+        inner_dim = dim_head *  heads
+        self.heads = heads
+        self.norm = LayerNorm(dim)
+
+        self.q_norm = RMSNorm(heads, dim_head)
+        self.k_norm = RMSNorm(heads, dim_head)
+
+        self.attend = nn.Softmax(dim = -1)
+        self.dropout = nn.Dropout(dropout)
+
+        self.to_q = nn.Linear(dim, inner_dim, bias = False)
+        self.to_kv = nn.Linear(dim, inner_dim * 2, bias = False)
+
+        self.to_out = nn.Sequential(
+            nn.Linear(inner_dim, dim, bias = False),
+            nn.Dropout(dropout)
+        )
+
+    def forward(
+        self,
+        x,
+        context = None,
+        mask = None,
+        attn_mask = None
+    ):
+        x = self.norm(x)
+        kv_input = default(context, x)
+
+        qkv = (self.to_q(x), *self.to_kv(kv_input).chunk(2, dim = -1))
+
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = self.heads), qkv)
+
+        q = self.q_norm(q)
+        k = self.k_norm(k)
+
+        dots = torch.matmul(q, k.transpose(-1, -2))
+
+        if exists(mask):
+            mask = rearrange(mask, 'b j -> b 1 1 j')
+            dots = dots.masked_fill(~mask, -torch.finfo(dots.dtype).max)
+
+        if exists(attn_mask):
+            dots = dots.masked_fill(~attn_mask, -torch.finfo(dots.dtype).max)
+
+        attn = self.attend(dots)
+        attn = self.dropout(attn)
+
+        out = torch.matmul(attn, v)
+        out = rearrange(out, 'b h n d -> b n (h d)')
+        return self.to_out(out)
+
+class Transformer(nn.Module):
+    def __init__(self, dim, depth, heads, dim_head, mlp_dim, dropout = 0.):
+        super().__init__()
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(nn.ModuleList([
+                Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout),
+                FeedForward(dim, mlp_dim, dropout = dropout)
+            ]))
+
+        self.norm = LayerNorm(dim)
+
+    def forward(
+        self,
+        x,
+        mask = None,
+        attn_mask = None
+    ):
+        for attn, ff in self.layers:
+            x = attn(x, mask = mask, attn_mask = attn_mask) + x
+            x = ff(x) + x
+
+        return self.norm(x)
+
+class NaViT(nn.Module):
+    def __init__(self, *, image_size, patch_size, num_classes, dim, depth, heads, mlp_dim, channels = 3, dim_head = 64, dropout = 0., emb_dropout = 0., token_dropout_prob = None):
+        super().__init__()
+        image_height, image_width = pair(image_size)
+
+        # what percent of tokens to dropout
+        # if int or float given, then assume constant dropout prob
+        # otherwise accept a callback that in turn calculates dropout prob from height and width
+
+        self.calc_token_dropout = None
+
+        if callable(token_dropout_prob):
+            self.calc_token_dropout = token_dropout_prob
+
+        elif isinstance(token_dropout_prob, (float, int)):
+            assert 0. < token_dropout_prob < 1.
+            token_dropout_prob = float(token_dropout_prob)
+            self.calc_token_dropout = lambda height, width: token_dropout_prob
+
+        # calculate patching related stuff
+
+        assert divisible_by(image_height, patch_size) and divisible_by(image_width, patch_size), 'Image dimensions must be divisible by the patch size.'
+
+        patch_height_dim, patch_width_dim = (image_height // patch_size), (image_width // patch_size)
+        patch_dim = channels * (patch_size ** 2)
+
+        self.channels = channels
+        self.patch_size = patch_size
+
+        self.to_patch_embedding = nn.Sequential(
+            LayerNorm(patch_dim),
+            nn.Linear(patch_dim, dim),
+            LayerNorm(dim),
+        )
+
+        self.pos_embed_height = nn.Parameter(torch.randn(patch_height_dim, dim))
+        self.pos_embed_width = nn.Parameter(torch.randn(patch_width_dim, dim))
+
+        self.dropout = nn.Dropout(emb_dropout)
+
+        self.transformer = Transformer(dim, depth, heads, dim_head, mlp_dim, dropout)
+
+        # final attention pooling queries
+
+        self.attn_pool_queries = nn.Parameter(torch.randn(dim))
+        self.attn_pool = Attention(dim = dim, dim_head = dim_head, heads = heads)
+
+        # output to logits
+
+        self.to_latent = nn.Identity()
+
+        self.mlp_head = nn.Sequential(
+            LayerNorm(dim),
+            nn.Linear(dim, num_classes, bias = False)
+        )
+
+    @property
+    def device(self):
+        return next(self.parameters()).device
+
+    def forward(
+        self,
+        batched_images: Union[List[Tensor], List[List[Tensor]]], # assume different resolution images already grouped correctly
+        group_images = False,
+        group_max_seq_len = 2048
+    ):
+        p, c, device, has_token_dropout = self.patch_size, self.channels, self.device, exists(self.calc_token_dropout)
+
+        arange = partial(torch.arange, device = device)
+        pad_sequence = partial(orig_pad_sequence, batch_first = True)
+
+        # auto pack if specified
+
+        if group_images:
+            batched_images = group_images_by_max_seq_len(
+                batched_images,
+                patch_size = self.patch_size,
+                calc_token_dropout = self.calc_token_dropout,
+                max_seq_len = group_max_seq_len
+            )
+
+        # process images into variable lengthed sequences with attention mask
+
+        num_images = []
+        batched_sequences = []
+        batched_positions = []
+        batched_image_ids = []
+
+        for images in batched_images:
+            num_images.append(len(images))
+
+            sequences = []
+            positions = []
+            image_ids = torch.empty((0,), device = device, dtype = torch.long)
+
+            for image_id, image in enumerate(images):
+                assert image.ndim ==3 and image.shape[0] == c
+                image_dims = image.shape[-2:]
+                assert all([divisible_by(dim, p) for dim in image_dims]), f'height and width {image_dims} of images must be divisible by patch size {p}'
+
+                ph, pw = map(lambda dim: dim // p, image_dims)
+
+                pos = torch.stack(torch.meshgrid((
+                    arange(ph),
+                    arange(pw)
+                ), indexing = 'ij'), dim = -1)
+
+                pos = rearrange(pos, 'h w c -> (h w) c')
+                seq = rearrange(image, 'c (h p1) (w p2) -> (h w) (c p1 p2)', p1 = p, p2 = p)
+
+                seq_len = seq.shape[-2]
+
+                if has_token_dropout:
+                    token_dropout = self.calc_token_dropout(*image_dims)
+                    num_keep = max(1, int(seq_len * (1 - token_dropout)))
+                    keep_indices = torch.randn((seq_len,), device = device).topk(num_keep, dim = -1).indices
+
+                    seq = seq[keep_indices]
+                    pos = pos[keep_indices]
+
+                image_ids = F.pad(image_ids, (0, seq.shape[-2]), value = image_id)
+                sequences.append(seq)
+                positions.append(pos)
+
+            batched_image_ids.append(image_ids)
+            batched_sequences.append(torch.cat(sequences, dim = 0))
+            batched_positions.append(torch.cat(positions, dim = 0))
+
+        # derive key padding mask
+
+        lengths = torch.tensor([seq.shape[-2] for seq in batched_sequences], device = device, dtype = torch.long)
+        max_length = arange(lengths.amax().item())
+        key_pad_mask = rearrange(lengths, 'b -> b 1') <= rearrange(max_length, 'n -> 1 n')
+
+        # derive attention mask, and combine with key padding mask from above
+
+        batched_image_ids = pad_sequence(batched_image_ids)
+        attn_mask = rearrange(batched_image_ids, 'b i -> b 1 i 1') == rearrange(batched_image_ids, 'b j -> b 1 1 j')
+        attn_mask = attn_mask & rearrange(key_pad_mask, 'b j -> b 1 1 j')
+
+        # combine patched images as well as the patched width / height positions for 2d positional embedding
+
+        patches = pad_sequence(batched_sequences)
+        patch_positions = pad_sequence(batched_positions)
+
+        # need to know how many images for final attention pooling
+
+        num_images = torch.tensor(num_images, device = device, dtype = torch.long)        
+
+        # to patches
+
+        x = self.to_patch_embedding(patches)        
+
+        # factorized 2d absolute positional embedding
+
+        h_indices, w_indices = patch_positions.unbind(dim = -1)
+
+        h_pos = self.pos_embed_height[h_indices]
+        w_pos = self.pos_embed_width[w_indices]
+
+        x = x + h_pos + w_pos
+
+        # embed dropout
+
+        x = self.dropout(x)
+
+        # attention
+
+        x = self.transformer(x, attn_mask = attn_mask)
+
+        # do attention pooling at the end
+
+        max_queries = num_images.amax().item()
+
+        queries = repeat(self.attn_pool_queries, 'd -> b n d', n = max_queries, b = x.shape[0])
+
+        # attention pool mask
+
+        image_id_arange = arange(max_queries)
+
+        attn_pool_mask = rearrange(image_id_arange, 'i -> i 1') == rearrange(batched_image_ids, 'b j -> b 1 j')
+
+        attn_pool_mask = attn_pool_mask & rearrange(key_pad_mask, 'b j -> b 1 j')
+
+        attn_pool_mask = rearrange(attn_pool_mask, 'b i j -> b 1 i j')
+
+        # attention pool
+
+        x = self.attn_pool(queries, context = x, attn_mask = attn_pool_mask) + queries
+
+        x = rearrange(x, 'b n d -> (b n) d')
+
+        # each batch element may not have same amount of images
+
+        is_images = image_id_arange < rearrange(num_images, 'b -> b 1')
+        is_images = rearrange(is_images, 'b n -> (b n)')
+
+        x = x[is_images]
+
+        # project out to logits
+
+        x = self.to_latent(x)
+
+        return self.mlp_head(x)

vit_pytorch/nest.py

@@ -24,19 +24,11 @@ class LayerNorm(nn.Module):
         mean = torch.mean(x, dim = 1, keepdim = True)
         return (x - mean) / (var + self.eps).sqrt() * self.g + self.b
 
-class PreNorm(nn.Module):
-    def __init__(self, dim, fn):
-        super().__init__()
-        self.norm = LayerNorm(dim)
-        self.fn = fn
-
-    def forward(self, x, **kwargs):
-        return self.fn(self.norm(x), **kwargs)
-
 class FeedForward(nn.Module):
     def __init__(self, dim, mlp_mult = 4, dropout = 0.):
         super().__init__()
         self.net = nn.Sequential(
+            LayerNorm(dim),
             nn.Conv2d(dim, dim * mlp_mult, 1),
             nn.GELU(),
             nn.Dropout(dropout),
@@ -54,6 +46,7 @@ class Attention(nn.Module):
         self.heads = heads
         self.scale = dim_head ** -0.5
 
+        self.norm = LayerNorm(dim)
         self.attend = nn.Softmax(dim = -1)
         self.dropout = nn.Dropout(dropout)
         self.to_qkv = nn.Conv2d(dim, inner_dim * 3, 1, bias = False)
@@ -66,6 +59,8 @@ class Attention(nn.Module):
     def forward(self, x):
         b, c, h, w, heads = *x.shape, self.heads
 
+        x = self.norm(x)
+
         qkv = self.to_qkv(x).chunk(3, dim = 1)
         q, k, v = map(lambda t: rearrange(t, 'b (h d) x y -> b h (x y) d', h = heads), qkv)
 
@@ -93,8 +88,8 @@ class Transformer(nn.Module):
 
         for _ in range(depth):
             self.layers.append(nn.ModuleList([
-                PreNorm(dim, Attention(dim, heads = heads, dropout = dropout)),
-                PreNorm(dim, FeedForward(dim, mlp_mult, dropout = dropout))
+                Attention(dim, heads = heads, dropout = dropout),
+                FeedForward(dim, mlp_mult, dropout = dropout)
             ]))
     def forward(self, x):
         *_, h, w = x.shape

vit_pytorch/parallel_vit.py

@@ -19,18 +19,11 @@ class Parallel(nn.Module):
     def forward(self, x):
         return sum([fn(x) for fn in self.fns])
 
-class PreNorm(nn.Module):
-    def __init__(self, dim, fn):
-        super().__init__()
-        self.norm = nn.LayerNorm(dim)
-        self.fn = fn
-    def forward(self, x, **kwargs):
-        return self.fn(self.norm(x), **kwargs)
-
 class FeedForward(nn.Module):
     def __init__(self, dim, hidden_dim, dropout = 0.):
         super().__init__()
         self.net = nn.Sequential(
+            nn.LayerNorm(dim),
             nn.Linear(dim, hidden_dim),
             nn.GELU(),
             nn.Dropout(dropout),
@@ -49,6 +42,7 @@ class Attention(nn.Module):
         self.heads = heads
         self.scale = dim_head ** -0.5
 
+        self.norm = nn.LayerNorm(dim)
         self.attend = nn.Softmax(dim = -1)
         self.dropout = nn.Dropout(dropout)
 
@@ -60,6 +54,7 @@ class Attention(nn.Module):
         ) if project_out else nn.Identity()
 
     def forward(self, x):
+        x = self.norm(x)
         qkv = self.to_qkv(x).chunk(3, dim = -1)
         q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = self.heads), qkv)
 
@@ -77,8 +72,8 @@ class Transformer(nn.Module):
         super().__init__()
         self.layers = nn.ModuleList([])
 
-        attn_block = lambda: PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout))
-        ff_block = lambda: PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout))        
+        attn_block = lambda: Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout)
+        ff_block = lambda: FeedForward(dim, mlp_dim, dropout = dropout)
 
         for _ in range(depth):
             self.layers.append(nn.ModuleList([

vit_pytorch/pit.py

@@ -17,18 +17,11 @@ def conv_output_size(image_size, kernel_size, stride, padding = 0):
 
 # classes
 
-class PreNorm(nn.Module):
-    def __init__(self, dim, fn):
-        super().__init__()
-        self.norm = nn.LayerNorm(dim)
-        self.fn = fn
-    def forward(self, x, **kwargs):
-        return self.fn(self.norm(x), **kwargs)
-
 class FeedForward(nn.Module):
     def __init__(self, dim, hidden_dim, dropout = 0.):
         super().__init__()
         self.net = nn.Sequential(
+            nn.LayerNorm(dim),
             nn.Linear(dim, hidden_dim),
             nn.GELU(),
             nn.Dropout(dropout),
@@ -47,6 +40,7 @@ class Attention(nn.Module):
         self.heads = heads
         self.scale = dim_head ** -0.5
 
+        self.norm = nn.LayerNorm(dim)
         self.attend = nn.Softmax(dim = -1)
         self.dropout = nn.Dropout(dropout)
         self.to_qkv = nn.Linear(dim, inner_dim * 3, bias = False)
@@ -58,6 +52,8 @@ class Attention(nn.Module):
 
     def forward(self, x):
         b, n, _, h = *x.shape, self.heads
+
+        x = self.norm(x)
         qkv = self.to_qkv(x).chunk(3, dim = -1)
         q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = h), qkv)
 
@@ -76,8 +72,8 @@ class Transformer(nn.Module):
         self.layers = nn.ModuleList([])
         for _ in range(depth):
             self.layers.append(nn.ModuleList([
-                PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout)),
-                PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout))
+                Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout),
+                FeedForward(dim, mlp_dim, dropout = dropout)
             ]))
     def forward(self, x):
         for attn, ff in self.layers:

vit_pytorch/rvt.py

@@ -55,14 +55,6 @@ class DepthWiseConv2d(nn.Module):
 
 # helper classes
 
-class PreNorm(nn.Module):
-    def __init__(self, dim, fn):
-        super().__init__()
-        self.norm = nn.LayerNorm(dim)
-        self.fn = fn
-    def forward(self, x, **kwargs):
-        return self.fn(self.norm(x), **kwargs)
-
 class SpatialConv(nn.Module):
     def __init__(self, dim_in, dim_out, kernel, bias = False):
         super().__init__()
@@ -86,6 +78,7 @@ class FeedForward(nn.Module):
     def __init__(self, dim, hidden_dim, dropout = 0., use_glu = True):
         super().__init__()
         self.net = nn.Sequential(
+            nn.LayerNorm(dim),
             nn.Linear(dim, hidden_dim * 2 if use_glu else hidden_dim),
             GEGLU() if use_glu else nn.GELU(),
             nn.Dropout(dropout),
@@ -103,6 +96,7 @@ class Attention(nn.Module):
         self.heads = heads
         self.scale = dim_head ** -0.5
 
+        self.norm = nn.LayerNorm(dim)
         self.attend = nn.Softmax(dim = -1)
         self.dropout = nn.Dropout(dropout)
 
@@ -121,6 +115,9 @@ class Attention(nn.Module):
         b, n, _, h = *x.shape, self.heads
 
         to_q_kwargs = {'fmap_dims': fmap_dims} if self.use_ds_conv else {}
+
+        x = self.norm(x)
+
         q = self.to_q(x, **to_q_kwargs)
 
         qkv = (q, *self.to_kv(x).chunk(2, dim = -1))
@@ -162,8 +159,8 @@ class Transformer(nn.Module):
         self.pos_emb = AxialRotaryEmbedding(dim_head, max_freq = image_size)
         for _ in range(depth):
             self.layers.append(nn.ModuleList([
-                PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout, use_rotary = use_rotary, use_ds_conv = use_ds_conv)),
-                PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout, use_glu = use_glu))
+                Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout, use_rotary = use_rotary, use_ds_conv = use_ds_conv),
+                FeedForward(dim, mlp_dim, dropout = dropout, use_glu = use_glu)
             ]))
     def forward(self, x, fmap_dims):
         pos_emb = self.pos_emb(x[:, 1:])

vit_pytorch/scalable_vit.py

@@ -33,15 +33,6 @@ class ChanLayerNorm(nn.Module):
         mean = torch.mean(x, dim = 1, keepdim = True)
         return (x - mean) / (var + self.eps).sqrt() * self.g + self.b
 
-class PreNorm(nn.Module):
-    def __init__(self, dim, fn):
-        super().__init__()
-        self.norm = ChanLayerNorm(dim)
-        self.fn = fn
-
-    def forward(self, x):
-        return self.fn(self.norm(x))
-
 class Downsample(nn.Module):
     def __init__(self, dim_in, dim_out):
         super().__init__()
@@ -65,6 +56,7 @@ class FeedForward(nn.Module):
         super().__init__()
         inner_dim = dim * expansion_factor
         self.net = nn.Sequential(
+            ChanLayerNorm(dim),
             nn.Conv2d(dim, inner_dim, 1),
             nn.GELU(),
             nn.Dropout(dropout),
@@ -92,6 +84,7 @@ class ScalableSelfAttention(nn.Module):
         self.attend = nn.Softmax(dim = -1)
         self.dropout = nn.Dropout(dropout)
 
+        self.norm = ChanLayerNorm(dim)
         self.to_q = nn.Conv2d(dim, dim_key * heads, 1, bias = False)
         self.to_k = nn.Conv2d(dim, dim_key * heads, reduction_factor, stride = reduction_factor, bias = False)
         self.to_v = nn.Conv2d(dim, dim_value * heads, reduction_factor, stride = reduction_factor, bias = False)
@@ -104,6 +97,8 @@ class ScalableSelfAttention(nn.Module):
     def forward(self, x):
         height, width, heads = *x.shape[-2:], self.heads
 
+        x = self.norm(x)
+
         q, k, v = self.to_q(x), self.to_k(x), self.to_v(x)
 
         # split out heads
@@ -145,6 +140,7 @@ class InteractiveWindowedSelfAttention(nn.Module):
         self.attend = nn.Softmax(dim = -1)
         self.dropout = nn.Dropout(dropout)
 
+        self.norm = ChanLayerNorm(dim)
         self.local_interactive_module = nn.Conv2d(dim_value * heads, dim_value * heads, 3, padding = 1)
 
         self.to_q = nn.Conv2d(dim, dim_key * heads, 1, bias = False)
@@ -159,6 +155,8 @@ class InteractiveWindowedSelfAttention(nn.Module):
     def forward(self, x):
         height, width, heads, wsz = *x.shape[-2:], self.heads, self.window_size
 
+        x = self.norm(x)
+
         wsz_h, wsz_w = default(wsz, height), default(wsz, width)
         assert (height % wsz_h) == 0 and (width % wsz_w) == 0, f'height ({height}) or width ({width}) of feature map is not divisible by the window size ({wsz_h}, {wsz_w})'
 
@@ -217,11 +215,11 @@ class Transformer(nn.Module):
             is_first = ind == 0
 
             self.layers.append(nn.ModuleList([
-                PreNorm(dim, ScalableSelfAttention(dim, heads = heads, dim_key = ssa_dim_key, dim_value = ssa_dim_value, reduction_factor = ssa_reduction_factor, dropout = dropout)),
-                PreNorm(dim, FeedForward(dim, expansion_factor = ff_expansion_factor, dropout = dropout)),
+                ScalableSelfAttention(dim, heads = heads, dim_key = ssa_dim_key, dim_value = ssa_dim_value, reduction_factor = ssa_reduction_factor, dropout = dropout),
+                FeedForward(dim, expansion_factor = ff_expansion_factor, dropout = dropout),
                 PEG(dim) if is_first else None,
-                PreNorm(dim, FeedForward(dim, expansion_factor = ff_expansion_factor, dropout = dropout)),
-                PreNorm(dim, InteractiveWindowedSelfAttention(dim, heads = heads, dim_key = iwsa_dim_key, dim_value = iwsa_dim_value, window_size = iwsa_window_size, dropout = dropout))
+                FeedForward(dim, expansion_factor = ff_expansion_factor, dropout = dropout),
+                InteractiveWindowedSelfAttention(dim, heads = heads, dim_key = iwsa_dim_key, dim_value = iwsa_dim_value, window_size = iwsa_window_size, dropout = dropout)
             ]))
 
         self.norm = ChanLayerNorm(dim) if norm_output else nn.Identity()

vit_pytorch/sep_vit.py

@@ -25,15 +25,6 @@ class ChanLayerNorm(nn.Module):
         mean = torch.mean(x, dim = 1, keepdim = True)
         return (x - mean) / (var + self.eps).sqrt() * self.g + self.b
 
-class PreNorm(nn.Module):
-    def __init__(self, dim, fn):
-        super().__init__()
-        self.norm = ChanLayerNorm(dim)
-        self.fn = fn
-
-    def forward(self, x):
-        return self.fn(self.norm(x))
-
 class OverlappingPatchEmbed(nn.Module):
     def __init__(self, dim_in, dim_out, stride = 2):
         super().__init__()
@@ -59,6 +50,7 @@ class FeedForward(nn.Module):
         super().__init__()
         inner_dim = int(dim * mult)
         self.net = nn.Sequential(
+            ChanLayerNorm(dim),
             nn.Conv2d(dim, inner_dim, 1),
             nn.GELU(),
             nn.Dropout(dropout),
@@ -85,6 +77,8 @@ class DSSA(nn.Module):
         self.window_size = window_size
         inner_dim = dim_head * heads
 
+        self.norm = ChanLayerNorm(dim)
+
         self.attend = nn.Sequential(
             nn.Softmax(dim = -1),
             nn.Dropout(dropout)
@@ -138,6 +132,8 @@ class DSSA(nn.Module):
         assert (height % wsz) == 0 and (width % wsz) == 0, f'height {height} and width {width} must be divisible by window size {wsz}'
         num_windows = (height // wsz) * (width // wsz)
 
+        x = self.norm(x)
+
         # fold in windows for "depthwise" attention - not sure why it is named depthwise when it is just "windowed" attention
 
         x = rearrange(x, 'b c (h w1) (w w2) -> (b h w) c (w1 w2)', w1 = wsz, w2 = wsz)
@@ -225,8 +221,8 @@ class Transformer(nn.Module):
 
         for ind in range(depth):
             self.layers.append(nn.ModuleList([
-                PreNorm(dim, DSSA(dim, heads = heads, dim_head = dim_head, dropout = dropout)),
-                PreNorm(dim, FeedForward(dim, mult = ff_mult, dropout = dropout)),
+                DSSA(dim, heads = heads, dim_head = dim_head, dropout = dropout),
+                FeedForward(dim, mult = ff_mult, dropout = dropout),
             ]))
 
         self.norm = ChanLayerNorm(dim) if norm_output else nn.Identity()

vit_pytorch/simple_vit.py

@@ -9,17 +9,15 @@ from einops.layers.torch import Rearrange
 def pair(t):
     return t if isinstance(t, tuple) else (t, t)
 
-def posemb_sincos_2d(patches, temperature = 10000, dtype = torch.float32):
-    _, h, w, dim, device, dtype = *patches.shape, patches.device, patches.dtype
-
-    y, x = torch.meshgrid(torch.arange(h, device = device), torch.arange(w, device = device), indexing = 'ij')
-    assert (dim % 4) == 0, 'feature dimension must be multiple of 4 for sincos emb'
-    omega = torch.arange(dim // 4, device = device) / (dim // 4 - 1)
-    omega = 1. / (temperature ** omega)
+def posemb_sincos_2d(h, w, dim, temperature: int = 10000, dtype = torch.float32):
+    y, x = torch.meshgrid(torch.arange(h), torch.arange(w), indexing="ij")
+    assert (dim % 4) == 0, "feature dimension must be multiple of 4 for sincos emb"
+    omega = torch.arange(dim // 4) / (dim // 4 - 1)
+    omega = 1.0 / (temperature ** omega)
 
     y = y.flatten()[:, None] * omega[None, :]
-    x = x.flatten()[:, None] * omega[None, :] 
-    pe = torch.cat((x.sin(), x.cos(), y.sin(), y.cos()), dim = 1)
+    x = x.flatten()[:, None] * omega[None, :]
+    pe = torch.cat((x.sin(), x.cos(), y.sin(), y.cos()), dim=1)
     return pe.type(dtype)
 
 # classes
@@ -66,6 +64,7 @@ class Attention(nn.Module):
 class Transformer(nn.Module):
     def __init__(self, dim, depth, heads, dim_head, mlp_dim):
         super().__init__()
+        self.norm = nn.LayerNorm(dim)
         self.layers = nn.ModuleList([])
         for _ in range(depth):
             self.layers.append(nn.ModuleList([
@@ -76,7 +75,7 @@ class Transformer(nn.Module):
         for attn, ff in self.layers:
             x = attn(x) + x
             x = ff(x) + x
-        return x
+        return self.norm(x)
 
 class SimpleViT(nn.Module):
     def __init__(self, *, image_size, patch_size, num_classes, dim, depth, heads, mlp_dim, channels = 3, dim_head = 64):
@@ -86,30 +85,33 @@ class SimpleViT(nn.Module):
 
         assert image_height % patch_height == 0 and image_width % patch_width == 0, 'Image dimensions must be divisible by the patch size.'
 
-        num_patches = (image_height // patch_height) * (image_width // patch_width)
         patch_dim = channels * patch_height * patch_width
 
         self.to_patch_embedding = nn.Sequential(
-            Rearrange('b c (h p1) (w p2) -> b h w (p1 p2 c)', p1 = patch_height, p2 = patch_width),
+            Rearrange("b c (h p1) (w p2) -> b (h w) (p1 p2 c)", p1 = patch_height, p2 = patch_width),
             nn.LayerNorm(patch_dim),
             nn.Linear(patch_dim, dim),
             nn.LayerNorm(dim),
         )
 
+        self.pos_embedding = posemb_sincos_2d(
+            h = image_height // patch_height,
+            w = image_width // patch_width,
+            dim = dim,
+        ) 
+
         self.transformer = Transformer(dim, depth, heads, dim_head, mlp_dim)
 
+        self.pool = "mean"
         self.to_latent = nn.Identity()
-        self.linear_head = nn.Sequential(
-            nn.LayerNorm(dim),
-            nn.Linear(dim, num_classes)
-        )
+
+        self.linear_head = nn.Linear(dim, num_classes)
 
     def forward(self, img):
-        *_, h, w, dtype = *img.shape, img.dtype
+        device = img.device
 
         x = self.to_patch_embedding(img)
-        pe = posemb_sincos_2d(x)
-        x = rearrange(x, 'b ... d -> b (...) d') + pe
+        x += self.pos_embedding.to(device, dtype=x.dtype)
 
         x = self.transformer(x)
         x = x.mean(dim = 1)

vit_pytorch/simple_vit_1d.py

@@ -62,6 +62,7 @@ class Attention(nn.Module):
 class Transformer(nn.Module):
     def __init__(self, dim, depth, heads, dim_head, mlp_dim):
         super().__init__()
+        self.norm = nn.LayerNorm(dim)
         self.layers = nn.ModuleList([])
         for _ in range(depth):
             self.layers.append(nn.ModuleList([
@@ -72,7 +73,7 @@ class Transformer(nn.Module):
         for attn, ff in self.layers:
             x = attn(x) + x
             x = ff(x) + x
-        return x
+        return self.norm(x)
 
 class SimpleViT(nn.Module):
     def __init__(self, *, seq_len, patch_size, num_classes, dim, depth, heads, mlp_dim, channels = 3, dim_head = 64):
@@ -93,10 +94,7 @@ class SimpleViT(nn.Module):
         self.transformer = Transformer(dim, depth, heads, dim_head, mlp_dim)
 
         self.to_latent = nn.Identity()
-        self.linear_head = nn.Sequential(
-            nn.LayerNorm(dim),
-            nn.Linear(dim, num_classes)
-        )
+        self.linear_head = nn.Linear(dim, num_classes)
 
     def forward(self, series):
         *_, n, dtype = *series.shape, series.dtype

vit_pytorch/simple_vit_3d.py

@@ -77,6 +77,7 @@ class Attention(nn.Module):
 class Transformer(nn.Module):
     def __init__(self, dim, depth, heads, dim_head, mlp_dim):
         super().__init__()
+        self.norm = nn.LayerNorm(dim)
         self.layers = nn.ModuleList([])
         for _ in range(depth):
             self.layers.append(nn.ModuleList([
@@ -87,7 +88,7 @@ class Transformer(nn.Module):
         for attn, ff in self.layers:
             x = attn(x) + x
             x = ff(x) + x
-        return x
+        return self.norm(x)
 
 class SimpleViT(nn.Module):
     def __init__(self, *, image_size, image_patch_size, frames, frame_patch_size, num_classes, dim, depth, heads, mlp_dim, channels = 3, dim_head = 64):
@@ -111,10 +112,7 @@ class SimpleViT(nn.Module):
         self.transformer = Transformer(dim, depth, heads, dim_head, mlp_dim)
 
         self.to_latent = nn.Identity()
-        self.linear_head = nn.Sequential(
-            nn.LayerNorm(dim),
-            nn.Linear(dim, num_classes)
-        )
+        self.linear_head = nn.Linear(dim, num_classes)
 
     def forward(self, video):
         *_, h, w, dtype = *video.shape, video.dtype

vit_pytorch/simple_vit_with_patch_dropout.py

@@ -87,6 +87,7 @@ class Attention(nn.Module):
 class Transformer(nn.Module):
     def __init__(self, dim, depth, heads, dim_head, mlp_dim):
         super().__init__()
+        self.norm = nn.LayerNorm(dim)
         self.layers = nn.ModuleList([])
         for _ in range(depth):
             self.layers.append(nn.ModuleList([
@@ -97,7 +98,7 @@ class Transformer(nn.Module):
         for attn, ff in self.layers:
             x = attn(x) + x
             x = ff(x) + x
-        return x
+        return self.norm(x)
 
 class SimpleViT(nn.Module):
     def __init__(self, *, image_size, patch_size, num_classes, dim, depth, heads, mlp_dim, channels = 3, dim_head = 64, patch_dropout = 0.5):
@@ -122,10 +123,7 @@ class SimpleViT(nn.Module):
         self.transformer = Transformer(dim, depth, heads, dim_head, mlp_dim)
 
         self.to_latent = nn.Identity()
-        self.linear_head = nn.Sequential(
-            nn.LayerNorm(dim),
-            nn.Linear(dim, num_classes)
-        )
+        self.linear_head = nn.Linear(dim, num_classes)
 
     def forward(self, img):
         *_, h, w, dtype = *img.shape, img.dtype

vit_pytorch/simple_vit_with_qk_norm.py

@@ -0,0 +1,141 @@
+import torch
+from torch import nn
+import torch.nn.functional as F
+
+from einops import rearrange
+from einops.layers.torch import Rearrange
+
+# helpers
+
+def pair(t):
+    return t if isinstance(t, tuple) else (t, t)
+
+def posemb_sincos_2d(h, w, dim, temperature: int = 10000, dtype = torch.float32):
+    y, x = torch.meshgrid(torch.arange(h), torch.arange(w), indexing="ij")
+    assert (dim % 4) == 0, "feature dimension must be multiple of 4 for sincos emb"
+    omega = torch.arange(dim // 4) / (dim // 4 - 1)
+    omega = 1.0 / (temperature ** omega)
+
+    y = y.flatten()[:, None] * omega[None, :]
+    x = x.flatten()[:, None] * omega[None, :]
+    pe = torch.cat((x.sin(), x.cos(), y.sin(), y.cos()), dim=1)
+    return pe.type(dtype)
+
+# they use a query-key normalization that is equivalent to rms norm (no mean-centering, learned gamma), from vit 22B paper
+
+# in latest tweet, seem to claim more stable training at higher learning rates
+# unsure if this has taken off within Brain, or it has some hidden drawback
+
+class RMSNorm(nn.Module):
+    def __init__(self, heads, dim):
+        super().__init__()
+        self.scale = dim ** 0.5
+        self.gamma = nn.Parameter(torch.ones(heads, 1, dim) / self.scale)
+
+    def forward(self, x):
+        normed = F.normalize(x, dim = -1)
+        return normed * self.scale * self.gamma
+
+# classes
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, hidden_dim):
+        super().__init__()
+        self.net = nn.Sequential(
+            nn.LayerNorm(dim),
+            nn.Linear(dim, hidden_dim),
+            nn.GELU(),
+            nn.Linear(hidden_dim, dim),
+        )
+    def forward(self, x):
+        return self.net(x)
+
+class Attention(nn.Module):
+    def __init__(self, dim, heads = 8, dim_head = 64):
+        super().__init__()
+        inner_dim = dim_head *  heads
+        self.heads = heads
+        self.norm = nn.LayerNorm(dim)
+
+        self.attend = nn.Softmax(dim = -1)
+
+        self.q_norm = RMSNorm(heads, dim_head)
+        self.k_norm = RMSNorm(heads, dim_head)
+
+        self.to_qkv = nn.Linear(dim, inner_dim * 3, bias = False)
+        self.to_out = nn.Linear(inner_dim, dim, bias = False)
+
+    def forward(self, x):
+        x = self.norm(x)
+
+        qkv = self.to_qkv(x).chunk(3, dim = -1)
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = self.heads), qkv)
+
+        q = self.q_norm(q)
+        k = self.k_norm(k)
+
+        dots = torch.matmul(q, k.transpose(-1, -2))
+
+        attn = self.attend(dots)
+
+        out = torch.matmul(attn, v)
+        out = rearrange(out, 'b h n d -> b n (h d)')
+        return self.to_out(out)
+
+class Transformer(nn.Module):
+    def __init__(self, dim, depth, heads, dim_head, mlp_dim):
+        super().__init__()
+        self.norm = nn.LayerNorm(dim)
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(nn.ModuleList([
+                Attention(dim, heads = heads, dim_head = dim_head),
+                FeedForward(dim, mlp_dim)
+            ]))
+    def forward(self, x):
+        for attn, ff in self.layers:
+            x = attn(x) + x
+            x = ff(x) + x
+        return self.norm(x)
+
+class SimpleViT(nn.Module):
+    def __init__(self, *, image_size, patch_size, num_classes, dim, depth, heads, mlp_dim, channels = 3, dim_head = 64):
+        super().__init__()
+        image_height, image_width = pair(image_size)
+        patch_height, patch_width = pair(patch_size)
+
+        assert image_height % patch_height == 0 and image_width % patch_width == 0, 'Image dimensions must be divisible by the patch size.'
+
+        patch_dim = channels * patch_height * patch_width
+
+        self.to_patch_embedding = nn.Sequential(
+            Rearrange("b c (h p1) (w p2) -> b (h w) (p1 p2 c)", p1 = patch_height, p2 = patch_width),
+            nn.LayerNorm(patch_dim),
+            nn.Linear(patch_dim, dim),
+            nn.LayerNorm(dim),
+        )
+
+        self.pos_embedding = posemb_sincos_2d(
+            h = image_height // patch_height,
+            w = image_width // patch_width,
+            dim = dim,
+        ) 
+
+        self.transformer = Transformer(dim, depth, heads, dim_head, mlp_dim)
+
+        self.pool = "mean"
+        self.to_latent = nn.Identity()
+
+        self.linear_head = nn.LayerNorm(dim)
+
+    def forward(self, img):
+        device = img.device
+
+        x = self.to_patch_embedding(img)
+        x += self.pos_embedding.to(device, dtype=x.dtype)
+
+        x = self.transformer(x)
+        x = x.mean(dim = 1)
+
+        x = self.to_latent(x)
+        return self.linear_head(x)

vit_pytorch/simple_vit_with_register_tokens.py

@@ -0,0 +1,129 @@
+import torch
+from torch import nn
+
+from einops import rearrange, repeat, pack, unpack
+from einops.layers.torch import Rearrange
+
+# helpers
+
+def pair(t):
+    return t if isinstance(t, tuple) else (t, t)
+
+def posemb_sincos_2d(h, w, dim, temperature: int = 10000, dtype = torch.float32):
+    y, x = torch.meshgrid(torch.arange(h), torch.arange(w), indexing="ij")
+    assert (dim % 4) == 0, "feature dimension must be multiple of 4 for sincos emb"
+    omega = torch.arange(dim // 4) / (dim // 4 - 1)
+    omega = 1.0 / (temperature ** omega)
+
+    y = y.flatten()[:, None] * omega[None, :]
+    x = x.flatten()[:, None] * omega[None, :]
+    pe = torch.cat((x.sin(), x.cos(), y.sin(), y.cos()), dim=1)
+    return pe.type(dtype)
+
+# classes
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, hidden_dim):
+        super().__init__()
+        self.net = nn.Sequential(
+            nn.LayerNorm(dim),
+            nn.Linear(dim, hidden_dim),
+            nn.GELU(),
+            nn.Linear(hidden_dim, dim),
+        )
+    def forward(self, x):
+        return self.net(x)
+
+class Attention(nn.Module):
+    def __init__(self, dim, heads = 8, dim_head = 64):
+        super().__init__()
+        inner_dim = dim_head *  heads
+        self.heads = heads
+        self.scale = dim_head ** -0.5
+        self.norm = nn.LayerNorm(dim)
+
+        self.attend = nn.Softmax(dim = -1)
+
+        self.to_qkv = nn.Linear(dim, inner_dim * 3, bias = False)
+        self.to_out = nn.Linear(inner_dim, dim, bias = False)
+
+    def forward(self, x):
+        x = self.norm(x)
+
+        qkv = self.to_qkv(x).chunk(3, dim = -1)
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = self.heads), qkv)
+
+        dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale
+
+        attn = self.attend(dots)
+
+        out = torch.matmul(attn, v)
+        out = rearrange(out, 'b h n d -> b n (h d)')
+        return self.to_out(out)
+
+class Transformer(nn.Module):
+    def __init__(self, dim, depth, heads, dim_head, mlp_dim):
+        super().__init__()
+        self.norm = nn.LayerNorm(dim)
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(nn.ModuleList([
+                Attention(dim, heads = heads, dim_head = dim_head),
+                FeedForward(dim, mlp_dim)
+            ]))
+    def forward(self, x):
+        for attn, ff in self.layers:
+            x = attn(x) + x
+            x = ff(x) + x
+        return self.norm(x)
+
+class SimpleViT(nn.Module):
+    def __init__(self, *, image_size, patch_size, num_classes, dim, depth, heads, mlp_dim, num_register_tokens = 4, channels = 3, dim_head = 64):
+        super().__init__()
+        image_height, image_width = pair(image_size)
+        patch_height, patch_width = pair(patch_size)
+
+        assert image_height % patch_height == 0 and image_width % patch_width == 0, 'Image dimensions must be divisible by the patch size.'
+
+        patch_dim = channels * patch_height * patch_width
+
+        self.to_patch_embedding = nn.Sequential(
+            Rearrange("b c (h p1) (w p2) -> b (h w) (p1 p2 c)", p1 = patch_height, p2 = patch_width),
+            nn.LayerNorm(patch_dim),
+            nn.Linear(patch_dim, dim),
+            nn.LayerNorm(dim),
+        )
+
+        self.register_tokens = nn.Parameter(torch.randn(num_register_tokens, dim))
+
+        self.pos_embedding = posemb_sincos_2d(
+            h = image_height // patch_height,
+            w = image_width // patch_width,
+            dim = dim,
+        ) 
+
+        self.transformer = Transformer(dim, depth, heads, dim_head, mlp_dim)
+
+        self.pool = "mean"
+        self.to_latent = nn.Identity()
+
+        self.linear_head = nn.Linear(dim, num_classes)
+
+    def forward(self, img):
+        batch, device = img.shape[0], img.device
+
+        x = self.to_patch_embedding(img)
+        x += self.pos_embedding.to(device, dtype=x.dtype)
+
+        r = repeat(self.register_tokens, 'n d -> b n d', b = batch)
+
+        x, ps = pack([x, r], 'b * d')
+
+        x = self.transformer(x)
+
+        x, _ = unpack(x, ps, 'b * d')
+
+        x = x.mean(dim = 1)
+
+        x = self.to_latent(x)
+        return self.linear_head(x)

vit_pytorch/twins_svt.py

@@ -42,20 +42,11 @@ class LayerNorm(nn.Module):
         mean = torch.mean(x, dim = 1, keepdim = True)
         return (x - mean) / (var + self.eps).sqrt() * self.g + self.b
 
-class PreNorm(nn.Module):
-    def __init__(self, dim, fn):
-        super().__init__()
-        self.norm = LayerNorm(dim)
-        self.fn = fn
-
-    def forward(self, x, **kwargs):
-        x = self.norm(x)
-        return self.fn(x, **kwargs)
-
 class FeedForward(nn.Module):
     def __init__(self, dim, mult = 4, dropout = 0.):
         super().__init__()
         self.net = nn.Sequential(
+            LayerNorm(dim),
             nn.Conv2d(dim, dim * mult, 1),
             nn.GELU(),
             nn.Dropout(dropout),
@@ -99,6 +90,7 @@ class LocalAttention(nn.Module):
         self.heads = heads
         self.scale = dim_head ** -0.5
 
+        self.norm = LayerNorm(dim)
         self.to_q = nn.Conv2d(dim, inner_dim, 1, bias = False)
         self.to_kv = nn.Conv2d(dim, inner_dim * 2, 1, bias = False)
 
@@ -108,6 +100,8 @@ class LocalAttention(nn.Module):
         )
 
     def forward(self, fmap):
+        fmap = self.norm(fmap)
+
         shape, p = fmap.shape, self.patch_size
         b, n, x, y, h = *shape, self.heads
         x, y = map(lambda t: t // p, (x, y))
@@ -132,6 +126,8 @@ class GlobalAttention(nn.Module):
         self.heads = heads
         self.scale = dim_head ** -0.5
 
+        self.norm = LayerNorm(dim)
+
         self.to_q = nn.Conv2d(dim, inner_dim, 1, bias = False)
         self.to_kv = nn.Conv2d(dim, inner_dim * 2, k, stride = k, bias = False)
 
@@ -143,6 +139,8 @@ class GlobalAttention(nn.Module):
         )
 
     def forward(self, x):
+        x = self.norm(x)
+
         shape = x.shape
         b, n, _, y, h = *shape, self.heads
         q, k, v = (self.to_q(x), *self.to_kv(x).chunk(2, dim = 1))
@@ -164,10 +162,10 @@ class Transformer(nn.Module):
         self.layers = nn.ModuleList([])
         for _ in range(depth):
             self.layers.append(nn.ModuleList([
-                Residual(PreNorm(dim, LocalAttention(dim, heads = heads, dim_head = dim_head, dropout = dropout, patch_size = local_patch_size))) if has_local else nn.Identity(),
-                Residual(PreNorm(dim, FeedForward(dim, mlp_mult, dropout = dropout))) if has_local else nn.Identity(),
-                Residual(PreNorm(dim, GlobalAttention(dim, heads = heads, dim_head = dim_head, dropout = dropout, k = global_k))),
-                Residual(PreNorm(dim, FeedForward(dim, mlp_mult, dropout = dropout)))
+                Residual(LocalAttention(dim, heads = heads, dim_head = dim_head, dropout = dropout, patch_size = local_patch_size)) if has_local else nn.Identity(),
+                Residual(FeedForward(dim, mlp_mult, dropout = dropout)) if has_local else nn.Identity(),
+                Residual(GlobalAttention(dim, heads = heads, dim_head = dim_head, dropout = dropout, k = global_k)),
+                Residual(FeedForward(dim, mlp_mult, dropout = dropout))
             ]))
     def forward(self, x):
         for local_attn, ff1, global_attn, ff2 in self.layers:

vit_pytorch/vit.py

@@ -11,24 +11,18 @@ def pair(t):
 
 # classes
 
-class PreNorm(nn.Module):
-    def __init__(self, dim, fn):
-        super().__init__()
-        self.norm = nn.LayerNorm(dim)
-        self.fn = fn
-    def forward(self, x, **kwargs):
-        return self.fn(self.norm(x), **kwargs)
-
 class FeedForward(nn.Module):
     def __init__(self, dim, hidden_dim, dropout = 0.):
         super().__init__()
         self.net = nn.Sequential(
+            nn.LayerNorm(dim),
             nn.Linear(dim, hidden_dim),
             nn.GELU(),
             nn.Dropout(dropout),
             nn.Linear(hidden_dim, dim),
             nn.Dropout(dropout)
         )
+
     def forward(self, x):
         return self.net(x)
 
@@ -41,6 +35,8 @@ class Attention(nn.Module):
         self.heads = heads
         self.scale = dim_head ** -0.5
 
+        self.norm = nn.LayerNorm(dim)
+
         self.attend = nn.Softmax(dim = -1)
         self.dropout = nn.Dropout(dropout)
 
@@ -52,6 +48,8 @@ class Attention(nn.Module):
         ) if project_out else nn.Identity()
 
     def forward(self, x):
+        x = self.norm(x)
+
         qkv = self.to_qkv(x).chunk(3, dim = -1)
         q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = self.heads), qkv)
 
@@ -67,17 +65,20 @@ class Attention(nn.Module):
 class Transformer(nn.Module):
     def __init__(self, dim, depth, heads, dim_head, mlp_dim, dropout = 0.):
         super().__init__()
+        self.norm = nn.LayerNorm(dim)
         self.layers = nn.ModuleList([])
         for _ in range(depth):
             self.layers.append(nn.ModuleList([
-                PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout)),
-                PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout))
+                Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout),
+                FeedForward(dim, mlp_dim, dropout = dropout)
             ]))
+
     def forward(self, x):
         for attn, ff in self.layers:
             x = attn(x) + x
             x = ff(x) + x
-        return x
+
+        return self.norm(x)
 
 class ViT(nn.Module):
     def __init__(self, *, image_size, patch_size, num_classes, dim, depth, heads, mlp_dim, pool = 'cls', channels = 3, dim_head = 64, dropout = 0., emb_dropout = 0.):
@@ -107,10 +108,7 @@ class ViT(nn.Module):
         self.pool = pool
         self.to_latent = nn.Identity()
 
-        self.mlp_head = nn.Sequential(
-            nn.LayerNorm(dim),
-            nn.Linear(dim, num_classes)
-        )
+        self.mlp_head = nn.Linear(dim, num_classes)
 
     def forward(self, img):
         x = self.to_patch_embedding(img)

vit_pytorch/vit_1d.py

@@ -6,18 +6,11 @@ from einops.layers.torch import Rearrange
 
 # classes
 
-class PreNorm(nn.Module):
-    def __init__(self, dim, fn):
-        super().__init__()
-        self.norm = nn.LayerNorm(dim)
-        self.fn = fn
-    def forward(self, x, **kwargs):
-        return self.fn(self.norm(x), **kwargs)
-
 class FeedForward(nn.Module):
     def __init__(self, dim, hidden_dim, dropout = 0.):
         super().__init__()
         self.net = nn.Sequential(
+            nn.Layernorm(dim),
             nn.Linear(dim, hidden_dim),
             nn.GELU(),
             nn.Dropout(dropout),
@@ -36,6 +29,7 @@ class Attention(nn.Module):
         self.heads = heads
         self.scale = dim_head ** -0.5
 
+        self.norm = nn.LayerNorm(dim)
         self.attend = nn.Softmax(dim = -1)
         self.dropout = nn.Dropout(dropout)
 
@@ -47,6 +41,7 @@ class Attention(nn.Module):
         ) if project_out else nn.Identity()
 
     def forward(self, x):
+        x = self.norm(x)
         qkv = self.to_qkv(x).chunk(3, dim = -1)
         q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = self.heads), qkv)
 
@@ -65,8 +60,8 @@ class Transformer(nn.Module):
         self.layers = nn.ModuleList([])
         for _ in range(depth):
             self.layers.append(nn.ModuleList([
-                PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout)),
-                PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout))
+                Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout),
+                FeedForward(dim, mlp_dim, dropout = dropout)
             ]))
     def forward(self, x):
         for attn, ff in self.layers:

vit_pytorch/vit_3d.py

@@ -11,18 +11,11 @@ def pair(t):
 
 # classes
 
-class PreNorm(nn.Module):
-    def __init__(self, dim, fn):
-        super().__init__()
-        self.norm = nn.LayerNorm(dim)
-        self.fn = fn
-    def forward(self, x, **kwargs):
-        return self.fn(self.norm(x), **kwargs)
-
 class FeedForward(nn.Module):
     def __init__(self, dim, hidden_dim, dropout = 0.):
         super().__init__()
         self.net = nn.Sequential(
+            nn.LayerNorm(dim),
             nn.Linear(dim, hidden_dim),
             nn.GELU(),
             nn.Dropout(dropout),
@@ -41,6 +34,7 @@ class Attention(nn.Module):
         self.heads = heads
         self.scale = dim_head ** -0.5
 
+        self.norm = nn.LayerNorm(dim)
         self.attend = nn.Softmax(dim = -1)
         self.dropout = nn.Dropout(dropout)
 
@@ -52,6 +46,7 @@ class Attention(nn.Module):
         ) if project_out else nn.Identity()
 
     def forward(self, x):
+        x = self.norm(x)
         qkv = self.to_qkv(x).chunk(3, dim = -1)
         q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = self.heads), qkv)
 
@@ -70,8 +65,8 @@ class Transformer(nn.Module):
         self.layers = nn.ModuleList([])
         for _ in range(depth):
             self.layers.append(nn.ModuleList([
-                PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout)),
-                PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout))
+                Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout),
+                FeedForward(dim, mlp_dim, dropout = dropout)
             ]))
     def forward(self, x):
         for attn, ff in self.layers:

vit_pytorch/vit_for_small_dataset.py

@@ -13,18 +13,11 @@ def pair(t):
 
 # classes
 
-class PreNorm(nn.Module):
-    def __init__(self, dim, fn):
-        super().__init__()
-        self.norm = nn.LayerNorm(dim)
-        self.fn = fn
-    def forward(self, x, **kwargs):
-        return self.fn(self.norm(x), **kwargs)
-
 class FeedForward(nn.Module):
     def __init__(self, dim, hidden_dim, dropout = 0.):
         super().__init__()
         self.net = nn.Sequential(
+            nn.LayerNorm(dim),
             nn.Linear(dim, hidden_dim),
             nn.GELU(),
             nn.Dropout(dropout),
@@ -41,6 +34,7 @@ class LSA(nn.Module):
         self.heads = heads
         self.temperature = nn.Parameter(torch.log(torch.tensor(dim_head ** -0.5)))
 
+        self.norm = nn.LayerNorm(dim)
         self.attend = nn.Softmax(dim = -1)
         self.dropout = nn.Dropout(dropout)
 
@@ -52,6 +46,7 @@ class LSA(nn.Module):
         )
 
     def forward(self, x):
+        x = self.norm(x)
         qkv = self.to_qkv(x).chunk(3, dim = -1)
         q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = self.heads), qkv)
 
@@ -74,8 +69,8 @@ class Transformer(nn.Module):
         self.layers = nn.ModuleList([])
         for _ in range(depth):
             self.layers.append(nn.ModuleList([
-                PreNorm(dim, LSA(dim, heads = heads, dim_head = dim_head, dropout = dropout)),
-                PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout))
+                LSA(dim, heads = heads, dim_head = dim_head, dropout = dropout),
+                FeedForward(dim, mlp_dim, dropout = dropout)
             ]))
     def forward(self, x):
         for attn, ff in self.layers:

vit_pytorch/vit_with_patch_dropout.py

@@ -30,18 +30,11 @@ class PatchDropout(nn.Module):
 
         return x[batch_indices, patch_indices_keep]
 
-class PreNorm(nn.Module):
-    def __init__(self, dim, fn):
-        super().__init__()
-        self.norm = nn.LayerNorm(dim)
-        self.fn = fn
-    def forward(self, x, **kwargs):
-        return self.fn(self.norm(x), **kwargs)
-
 class FeedForward(nn.Module):
     def __init__(self, dim, hidden_dim, dropout = 0.):
         super().__init__()
         self.net = nn.Sequential(
+            nn.LayerNorm(dim),
             nn.Linear(dim, hidden_dim),
             nn.GELU(),
             nn.Dropout(dropout),
@@ -60,6 +53,7 @@ class Attention(nn.Module):
         self.heads = heads
         self.scale = dim_head ** -0.5
 
+        self.norm = nn.LayerNorm(dim)
         self.attend = nn.Softmax(dim = -1)
         self.dropout = nn.Dropout(dropout)
 
@@ -71,6 +65,7 @@ class Attention(nn.Module):
         ) if project_out else nn.Identity()
 
     def forward(self, x):
+        x = self.norm(x)
         qkv = self.to_qkv(x).chunk(3, dim = -1)
         q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = self.heads), qkv)
 
@@ -89,8 +84,8 @@ class Transformer(nn.Module):
         self.layers = nn.ModuleList([])
         for _ in range(depth):
             self.layers.append(nn.ModuleList([
-                PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout)),
-                PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout))
+                Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout),
+                FeedForward(dim, mlp_dim, dropout = dropout)
             ]))
     def forward(self, x):
         for attn, ff in self.layers:

vit_pytorch/vit_with_patch_merger.py

@@ -32,18 +32,11 @@ class PatchMerger(nn.Module):
 
 # classes
 
-class PreNorm(nn.Module):
-    def __init__(self, dim, fn):
-        super().__init__()
-        self.norm = nn.LayerNorm(dim)
-        self.fn = fn
-    def forward(self, x, **kwargs):
-        return self.fn(self.norm(x), **kwargs)
-
 class FeedForward(nn.Module):
     def __init__(self, dim, hidden_dim, dropout = 0.):
         super().__init__()
         self.net = nn.Sequential(
+            nn.LayerNorm(dim),
             nn.Linear(dim, hidden_dim),
             nn.GELU(),
             nn.Dropout(dropout),
@@ -62,6 +55,7 @@ class Attention(nn.Module):
         self.heads = heads
         self.scale = dim_head ** -0.5
 
+        self.norm = nn.LayerNorm(dim)
         self.attend = nn.Softmax(dim = -1)
         self.dropout = nn.Dropout(dropout)
 
@@ -73,6 +67,7 @@ class Attention(nn.Module):
         ) if project_out else nn.Identity()
 
     def forward(self, x):
+        x = self.norm(x)
         qkv = self.to_qkv(x).chunk(3, dim = -1)
         q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = self.heads), qkv)
 
@@ -88,6 +83,7 @@ class Attention(nn.Module):
 class Transformer(nn.Module):
     def __init__(self, dim, depth, heads, dim_head, mlp_dim, dropout = 0., patch_merge_layer = None, patch_merge_num_tokens = 8):
         super().__init__()
+        self.norm = nn.LayerNorm(dim)
         self.layers = nn.ModuleList([])
 
         self.patch_merge_layer_index = default(patch_merge_layer, depth // 2) - 1 # default to mid-way through transformer, as shown in paper
@@ -95,8 +91,8 @@ class Transformer(nn.Module):
 
         for _ in range(depth):
             self.layers.append(nn.ModuleList([
-                PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout)),
-                PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout))
+                Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout),
+                FeedForward(dim, mlp_dim, dropout = dropout)
             ]))
     def forward(self, x):
         for index, (attn, ff) in enumerate(self.layers):
@@ -106,7 +102,7 @@ class Transformer(nn.Module):
             if index == self.patch_merge_layer_index:
                 x = self.patch_merger(x)
 
-        return x
+        return self.norm(x)
 
 class ViT(nn.Module):
     def __init__(self, *, image_size, patch_size, num_classes, dim, depth, heads, mlp_dim, patch_merge_layer = None, patch_merge_num_tokens = 8, channels = 3, dim_head = 64, dropout = 0., emb_dropout = 0.):
@@ -133,7 +129,6 @@ class ViT(nn.Module):
 
         self.mlp_head = nn.Sequential(
             Reduce('b n d -> b d', 'mean'),
-            nn.LayerNorm(dim),
             nn.Linear(dim, num_classes)
         )
 

vit_pytorch/vivit.py

@@ -14,18 +14,11 @@ def pair(t):
 
 # classes
 
-class PreNorm(nn.Module):
-    def __init__(self, dim, fn):
-        super().__init__()
-        self.norm = nn.LayerNorm(dim)
-        self.fn = fn
-    def forward(self, x, **kwargs):
-        return self.fn(self.norm(x), **kwargs)
-
 class FeedForward(nn.Module):
     def __init__(self, dim, hidden_dim, dropout = 0.):
         super().__init__()
         self.net = nn.Sequential(
+            nn.LayerNorm(dim),
             nn.Linear(dim, hidden_dim),
             nn.GELU(),
             nn.Dropout(dropout),
@@ -44,6 +37,7 @@ class Attention(nn.Module):
         self.heads = heads
         self.scale = dim_head ** -0.5
 
+        self.norm = nn.LayerNorm(dim)
         self.attend = nn.Softmax(dim = -1)
         self.dropout = nn.Dropout(dropout)
 
@@ -55,6 +49,7 @@ class Attention(nn.Module):
         ) if project_out else nn.Identity()
 
     def forward(self, x):
+        x = self.norm(x)
         qkv = self.to_qkv(x).chunk(3, dim = -1)
         q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = self.heads), qkv)
 
@@ -70,17 +65,18 @@ class Attention(nn.Module):
 class Transformer(nn.Module):
     def __init__(self, dim, depth, heads, dim_head, mlp_dim, dropout = 0.):
         super().__init__()
+        self.norm = nn.LayerNorm(dim)
         self.layers = nn.ModuleList([])
         for _ in range(depth):
             self.layers.append(nn.ModuleList([
-                PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout)),
-                PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout))
+                Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout),
+                FeedForward(dim, mlp_dim, dropout = dropout)
             ]))
     def forward(self, x):
         for attn, ff in self.layers:
             x = attn(x) + x
             x = ff(x) + x
-        return x
+        return self.norm(x)
 
 class ViT(nn.Module):
     def __init__(
@@ -137,16 +133,13 @@ class ViT(nn.Module):
         self.pool = pool
         self.to_latent = nn.Identity()
 
-        self.mlp_head = nn.Sequential(
-            nn.LayerNorm(dim),
-            nn.Linear(dim, num_classes)
-        )
+        self.mlp_head = nn.Linear(dim, num_classes)
 
     def forward(self, video):
         x = self.to_patch_embedding(video)
         b, f, n, _ = x.shape
 
-        x = x + self.pos_embedding
+        x = x + self.pos_embedding[:, :f, :n]
 
         if exists(self.spatial_cls_token):
             spatial_cls_tokens = repeat(self.spatial_cls_token, '1 1 d -> b f 1 d', b = b, f = f)