norm not needed when reusing attention in lookvit

Co-authored-by: Artem Lukin <artyom.lukin98@gmail.com>

README.md

@@ -777,7 +777,7 @@ pred = mbvit_xs(img) # (1, 1000)
 
 <img src="./images/xcit.png" width="400px"></img>
 
-This <a href="https://arxiv.org/abs/2106.09681">paper</a> introduces the cross correlation attention (abbreviated XCA). One can think of it as doing attention across the features dimension rather than the spatial one (another perspective would be a dynamic 1x1 convolution, the kernel being attention map defined by spatial correlations).
+This <a href="https://arxiv.org/abs/2106.09681">paper</a> introduces the cross covariance attention (abbreviated XCA). One can think of it as doing attention across the features dimension rather than the spatial one (another perspective would be a dynamic 1x1 convolution, the kernel being attention map defined by spatial correlations).
 
 Technically, this amounts to simply transposing the query, key, values before executing cosine similarity attention with learned temperature.
 
@@ -2072,4 +2072,20 @@ Coming from computer vision and new to transformers? Here are some resources tha
 }
 ```
 
+```bibtex
+@inproceedings{Koner2024LookupViTCV,
+    title   = {LookupViT: Compressing visual information to a limited number of tokens},
+    author  = {Rajat Koner and Gagan Jain and Prateek Jain and Volker Tresp and Sujoy Paul},
+    year    = {2024},
+    url     = {https://api.semanticscholar.org/CorpusID:271244592}
+}
+```
+
+```bibtex
+@misc{Rubin2024,
+    author  = {Ohad Rubin},
+    url     = {https://medium.com/@ohadrubin/exploring-weight-decay-in-layer-normalization-challenges-and-a-reparameterization-solution-ad4d12c24950}
+}
+```
+
 *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

@@ -1,11 +1,15 @@
 from setuptools import setup, find_packages
 
+with open('README.md') as f:
+    long_description = f.read()
+
 setup(
   name = 'vit-pytorch',
   packages = find_packages(exclude=['examples']),
-  version = '1.6.0',
+  version = '1.7.2',
   license='MIT',
   description = 'Vision Transformer (ViT) - Pytorch',
+  long_description=long_description,
   long_description_content_type = 'text/markdown',
   author = 'Phil Wang',
   author_email = 'lucidrains@gmail.com',

vit_pytorch/__init__.py

@@ -1,10 +1,3 @@
-import torch
-from packaging import version
-
-if version.parse(torch.__version__) >= version.parse('2.0.0'):
-    from einops._torch_specific import allow_ops_in_compiled_graph
-    allow_ops_in_compiled_graph()
-
 from vit_pytorch.vit import ViT
 from vit_pytorch.simple_vit import SimpleViT
 

vit_pytorch/cross_vit.py

@@ -170,12 +170,13 @@ class ImageEmbedder(nn.Module):
         dim,
         image_size,
         patch_size,
-        dropout = 0.
+        dropout = 0.,
+        channels = 3
     ):
         super().__init__()
         assert image_size % patch_size == 0, 'Image dimensions must be divisible by the patch size.'
         num_patches = (image_size // patch_size) ** 2
-        patch_dim = 3 * patch_size ** 2
+        patch_dim = channels * patch_size ** 2
 
         self.to_patch_embedding = nn.Sequential(
             Rearrange('b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1 = patch_size, p2 = patch_size),
@@ -223,11 +224,12 @@ class CrossViT(nn.Module):
         cross_attn_dim_head = 64,
         depth = 3,
         dropout = 0.1,
-        emb_dropout = 0.1
+        emb_dropout = 0.1,
+        channels = 3
     ):
         super().__init__()
-        self.sm_image_embedder = ImageEmbedder(dim = sm_dim, image_size = image_size, patch_size = sm_patch_size, dropout = emb_dropout)
-        self.lg_image_embedder = ImageEmbedder(dim = lg_dim, image_size = image_size, patch_size = lg_patch_size, dropout = emb_dropout)
+        self.sm_image_embedder = ImageEmbedder(dim = sm_dim, channels= channels, image_size = image_size, patch_size = sm_patch_size, dropout = emb_dropout)
+        self.lg_image_embedder = ImageEmbedder(dim = lg_dim, channels = channels, image_size = image_size, patch_size = lg_patch_size, dropout = emb_dropout)
 
         self.multi_scale_encoder = MultiScaleEncoder(
             depth = depth,

vit_pytorch/cvt.py

@@ -140,12 +140,13 @@ class CvT(nn.Module):
         s3_heads = 6,
         s3_depth = 10,
         s3_mlp_mult = 4,
-        dropout = 0.
+        dropout = 0.,
+        channels = 3
     ):
         super().__init__()
         kwargs = dict(locals())
 
-        dim = 3
+        dim = channels
         layers = []
 
         for prefix in ('s1', 's2', 's3'):

vit_pytorch/distill.py

@@ -1,6 +1,8 @@
 import torch
-import torch.nn.functional as F
 from torch import nn
+from torch.nn import Module
+import torch.nn.functional as F
+
 from vit_pytorch.vit import ViT
 from vit_pytorch.t2t import T2TViT
 from vit_pytorch.efficient import ViT as EfficientViT
@@ -12,6 +14,9 @@ from einops import rearrange, repeat
 def exists(val):
     return val is not None
 
+def default(val, d):
+    return val if exists(val) else d
+
 # classes
 
 class DistillMixin:
@@ -20,12 +25,12 @@ class DistillMixin:
         x = self.to_patch_embedding(img)
         b, n, _ = x.shape
 
-        cls_tokens = repeat(self.cls_token, '() n d -> b n d', b = b)
+        cls_tokens = repeat(self.cls_token, '1 n d -> b n d', b = b)
         x = torch.cat((cls_tokens, x), dim = 1)
         x += self.pos_embedding[:, :(n + 1)]
 
         if distilling:
-            distill_tokens = repeat(distill_token, '() n d -> b n d', b = b)
+            distill_tokens = repeat(distill_token, '1 n d -> b n d', b = b)
             x = torch.cat((x, distill_tokens), dim = 1)
 
         x = self._attend(x)
@@ -97,7 +102,7 @@ class DistillableEfficientViT(DistillMixin, EfficientViT):
 
 # knowledge distillation wrapper
 
-class DistillWrapper(nn.Module):
+class DistillWrapper(Module):
     def __init__(
         self,
         *,
@@ -105,7 +110,8 @@ class DistillWrapper(nn.Module):
         student,
         temperature = 1.,
         alpha = 0.5,
-        hard = False
+        hard = False,
+        mlp_layernorm = False
     ):
         super().__init__()
         assert (isinstance(student, (DistillableViT, DistillableT2TViT, DistillableEfficientViT))) , 'student must be a vision transformer'
@@ -122,14 +128,14 @@ class DistillWrapper(nn.Module):
         self.distillation_token = nn.Parameter(torch.randn(1, 1, dim))
 
         self.distill_mlp = nn.Sequential(
-            nn.LayerNorm(dim),
+            nn.LayerNorm(dim) if mlp_layernorm else nn.Identity(),
             nn.Linear(dim, num_classes)
         )
 
     def forward(self, img, labels, temperature = None, alpha = None, **kwargs):
-        b, *_ = img.shape
-        alpha = alpha if exists(alpha) else self.alpha
-        T = temperature if exists(temperature) else self.temperature
+
+        alpha = default(alpha, self.alpha)
+        T = default(temperature, self.temperature)
 
         with torch.no_grad():
             teacher_logits = self.teacher(img)

vit_pytorch/look_vit.py

@@ -0,0 +1,267 @@
+import torch
+from torch import nn
+import torch.nn.functional as F
+from torch.nn import Module, ModuleList
+
+from einops import einsum, rearrange, repeat, reduce
+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 divisible_by(num, den):
+    return (num % den) == 0
+
+# simple vit sinusoidal pos emb
+
+def posemb_sincos_2d(t, temperature = 10000):
+    h, w, d, device = *t.shape[1:], t.device
+    y, x = torch.meshgrid(torch.arange(h, device = device), torch.arange(w, device = device), indexing = 'ij')
+    assert (d % 4) == 0, "feature dimension must be multiple of 4 for sincos emb"
+    omega = torch.arange(d // 4, device = device) / (d // 4 - 1)
+    omega = temperature ** -omega
+
+    y = y.flatten()[:, None] * omega[None, :]
+    x = x.flatten()[:, None] * omega[None, :]
+    pos = torch.cat((x.sin(), x.cos(), y.sin(), y.cos()), dim = 1)
+
+    return pos.float()
+
+# bias-less layernorm with unit offset trick (discovered by Ohad Rubin)
+
+class LayerNorm(Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.ln = nn.LayerNorm(dim, elementwise_affine = False)
+        self.gamma = nn.Parameter(torch.zeros(dim))
+
+    def forward(self, x):
+        normed = self.ln(x)
+        return normed * (self.gamma + 1)
+
+# mlp
+
+def MLP(dim, factor = 4, dropout = 0.):
+    hidden_dim = int(dim * factor)
+    return nn.Sequential(
+        LayerNorm(dim),
+        nn.Linear(dim, hidden_dim),
+        nn.GELU(),
+        nn.Dropout(dropout),
+        nn.Linear(hidden_dim, dim),
+        nn.Dropout(dropout)
+    )
+
+# attention
+
+class Attention(Module):
+    def __init__(
+        self,
+        dim,
+        heads = 8,
+        dim_head = 64,
+        dropout = 0.,
+        reuse_attention = False
+    ):
+        super().__init__()
+        inner_dim = dim_head *  heads
+
+        self.scale = dim_head ** -0.5
+        self.heads = heads
+        self.reuse_attention = reuse_attention
+
+        self.split_heads = Rearrange('b n (h d) -> b h n d', h = heads)
+
+        self.norm = LayerNorm(dim) if not reuse_attention else nn.Identity()
+        self.attend = nn.Softmax(dim = -1)
+        self.dropout = nn.Dropout(dropout)
+
+        self.to_q = nn.Linear(dim, inner_dim, bias = False) if not reuse_attention else None
+        self.to_k = nn.Linear(dim, inner_dim, bias = False) if not reuse_attention else None
+        self.to_v = nn.Linear(dim, inner_dim, bias = False)
+
+        self.to_out = nn.Sequential(
+            Rearrange('b h n d -> b n (h d)'),
+            nn.Linear(inner_dim, dim, bias = False),
+            nn.Dropout(dropout)
+        )
+
+    def forward(
+        self,
+        x,
+        context = None,
+        return_attn = False,
+        attn = None
+    ):
+        x = self.norm(x)
+        context = default(context, x)
+
+        v = self.to_v(context)
+        v = self.split_heads(v)
+
+        if not self.reuse_attention:
+            qk = (self.to_q(x), self.to_k(context))
+            q, k = tuple(self.split_heads(t) for t in qk)
+
+            q = q * self.scale
+            sim = einsum(q, k, 'b h i d, b h j d -> b h i j')
+
+            attn = self.attend(sim)
+            attn = self.dropout(attn)
+        else:
+            assert exists(attn), 'attention matrix must be passed in for reusing previous attention'
+
+        out = einsum(attn, v, 'b h i j, b h j d -> b h i d')
+        out = self.to_out(out)
+
+        if not return_attn:
+            return out
+
+        return out, attn
+
+# LookViT
+
+class LookViT(Module):
+    def __init__(
+        self,
+        *,
+        dim,
+        image_size,
+        num_classes,
+        depth = 3,
+        patch_size = 16,
+        heads = 8,
+        mlp_factor = 4,
+        dim_head = 64,
+        highres_patch_size = 12,
+        highres_mlp_factor = 4,
+        cross_attn_heads = 8,
+        cross_attn_dim_head = 64,
+        patch_conv_kernel_size = 7,
+        dropout = 0.1,
+        channels = 3
+    ):
+        super().__init__()
+        assert divisible_by(image_size, highres_patch_size)
+        assert divisible_by(image_size, patch_size)
+        assert patch_size > highres_patch_size, 'patch size of the main vision transformer should be smaller than the highres patch sizes (that does the `lookup`)'
+        assert not divisible_by(patch_conv_kernel_size, 2)
+
+        self.dim = dim
+        self.image_size = image_size
+        self.patch_size = patch_size
+
+        kernel_size = patch_conv_kernel_size
+        patch_dim = (highres_patch_size * highres_patch_size) * channels
+
+        self.to_patches = nn.Sequential(
+            Rearrange('b c (h p1) (w p2) -> b (p1 p2 c) h w', p1 = highres_patch_size, p2 = highres_patch_size),
+            nn.Conv2d(patch_dim, dim, kernel_size, padding = kernel_size // 2),
+            Rearrange('b c h w -> b h w c'),
+            LayerNorm(dim),
+        )
+
+        # absolute positions
+
+        num_patches = (image_size // highres_patch_size) ** 2
+        self.pos_embedding = nn.Parameter(torch.randn(num_patches, dim))
+
+        # lookvit blocks
+
+        layers = ModuleList([])
+
+        for _ in range(depth):
+            layers.append(ModuleList([
+                Attention(dim = dim, dim_head = dim_head, heads = heads, dropout = dropout),
+                MLP(dim = dim, factor = mlp_factor, dropout = dropout),
+                Attention(dim = dim, dim_head = cross_attn_dim_head, heads = cross_attn_heads, dropout = dropout),
+                Attention(dim = dim, dim_head = cross_attn_dim_head, heads = cross_attn_heads, dropout = dropout, reuse_attention = True),
+                LayerNorm(dim),
+                MLP(dim = dim, factor = highres_mlp_factor, dropout = dropout)
+            ]))
+
+        self.layers = layers
+
+        self.norm = LayerNorm(dim)
+        self.highres_norm = LayerNorm(dim)
+
+        self.to_logits = nn.Linear(dim, num_classes, bias = False)
+
+    def forward(self, img):
+        assert img.shape[-2:] == (self.image_size, self.image_size)
+
+        # to patch tokens and positions
+
+        highres_tokens = self.to_patches(img)
+        size = highres_tokens.shape[-2]
+
+        pos_emb = posemb_sincos_2d(highres_tokens)
+        highres_tokens = highres_tokens + rearrange(pos_emb, '(h w) d -> h w d', h = size)
+
+        tokens = F.interpolate(
+            rearrange(highres_tokens, 'b h w d -> b d h w'),
+            img.shape[-1] // self.patch_size,
+            mode = 'bilinear'
+        )
+
+        tokens = rearrange(tokens, 'b c h w -> b (h w) c')
+        highres_tokens = rearrange(highres_tokens, 'b h w c -> b (h w) c')
+
+        # attention and feedforwards
+
+        for attn, mlp, lookup_cross_attn, highres_attn, highres_norm, highres_mlp in self.layers:
+
+            # main tokens cross attends (lookup) on the high res tokens
+
+            lookup_out, lookup_attn = lookup_cross_attn(tokens, highres_tokens, return_attn = True)  # return attention as they reuse the attention matrix
+            tokens = lookup_out + tokens
+
+            tokens = attn(tokens) + tokens
+            tokens = mlp(tokens) + tokens
+
+            # attention-reuse
+
+            lookup_attn = rearrange(lookup_attn, 'b h i j -> b h j i') # transpose for reverse cross attention
+
+            highres_tokens = highres_attn(highres_tokens, tokens, attn = lookup_attn) + highres_tokens
+            highres_tokens = highres_norm(highres_tokens)
+
+            highres_tokens = highres_mlp(highres_tokens) + highres_tokens
+
+        # to logits
+
+        tokens = self.norm(tokens)
+        highres_tokens = self.highres_norm(highres_tokens)
+
+        tokens = reduce(tokens, 'b n d -> b d', 'mean')
+        highres_tokens = reduce(highres_tokens, 'b n d -> b d', 'mean')
+
+        return self.to_logits(tokens + highres_tokens)
+
+# main
+
+if __name__ == '__main__':
+    v = LookViT(
+        image_size = 256,
+        num_classes = 1000,
+        dim = 512,
+        depth = 2,
+        heads = 8,
+        dim_head = 64,
+        patch_size = 32,
+        highres_patch_size = 8,
+        highres_mlp_factor = 2,
+        cross_attn_heads = 8,
+        cross_attn_dim_head = 64,
+        dropout = 0.1
+    ).cuda()
+
+    img = torch.randn(2, 3, 256, 256).cuda()
+    pred = v(img)
+
+    assert pred.shape == (2, 1000)

vit_pytorch/na_vit.py

@@ -198,7 +198,7 @@ class NaViT(nn.Module):
             self.calc_token_dropout = token_dropout_prob
 
         elif isinstance(token_dropout_prob, (float, int)):
-            assert 0. < token_dropout_prob < 1.
+            assert 0. <= token_dropout_prob < 1.
             token_dropout_prob = float(token_dropout_prob)
             self.calc_token_dropout = lambda height, width: token_dropout_prob
 
@@ -249,7 +249,7 @@ class NaViT(nn.Module):
         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)
+        p, c, device, has_token_dropout = self.patch_size, self.channels, self.device, exists(self.calc_token_dropout) and self.training
 
         arange = partial(torch.arange, device = device)
         pad_sequence = partial(orig_pad_sequence, batch_first = True)
@@ -260,7 +260,7 @@ class NaViT(nn.Module):
             batched_images = group_images_by_max_seq_len(
                 batched_images,
                 patch_size = self.patch_size,
-                calc_token_dropout = self.calc_token_dropout,
+                calc_token_dropout = self.calc_token_dropout if self.training else None,
                 max_seq_len = group_max_seq_len
             )
 
@@ -314,8 +314,8 @@ class NaViT(nn.Module):
         # 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')
+        seq_arange = arange(lengths.amax().item())
+        key_pad_mask = rearrange(seq_arange, 'n -> 1 n') < rearrange(lengths, 'b -> b 1')
 
         # derive attention mask, and combine with key padding mask from above
 

vit_pytorch/rvt.py

@@ -3,12 +3,14 @@ from math import sqrt, pi, log
 import torch
 from torch import nn, einsum
 import torch.nn.functional as F
+from torch.cuda.amp import autocast
 
 from einops import rearrange, repeat
 from einops.layers.torch import Rearrange
 
 # rotary embeddings
 
+@autocast(enabled = False)
 def rotate_every_two(x):
     x = rearrange(x, '... (d j) -> ... d j', j = 2)
     x1, x2 = x.unbind(dim = -1)
@@ -22,6 +24,7 @@ class AxialRotaryEmbedding(nn.Module):
         scales = torch.linspace(1., max_freq / 2, self.dim // 4)
         self.register_buffer('scales', scales)
 
+    @autocast(enabled = False)
     def forward(self, x):
         device, dtype, n = x.device, x.dtype, int(sqrt(x.shape[-2]))
 

vit_pytorch/simple_flash_attn_vit_3d.py

@@ -0,0 +1,171 @@
+from packaging import version
+from collections import namedtuple
+
+import torch
+from torch import nn
+import torch.nn.functional as F
+from torch.nn import Module, ModuleList
+
+from einops import rearrange
+from einops.layers.torch import Rearrange
+
+# constants
+
+Config = namedtuple('FlashAttentionConfig', ['enable_flash', 'enable_math', 'enable_mem_efficient'])
+
+# helpers
+
+def pair(t):
+    return t if isinstance(t, tuple) else (t, t)
+
+def posemb_sincos_3d(patches, temperature = 10000, dtype = torch.float32):
+    _, f, h, w, dim, device, dtype = *patches.shape, patches.device, patches.dtype
+
+    z, y, x = torch.meshgrid(
+        torch.arange(f, device = device),
+        torch.arange(h, device = device),
+        torch.arange(w, device = device),
+    indexing = 'ij')
+
+    fourier_dim = dim // 6
+
+    omega = torch.arange(fourier_dim, device = device) / (fourier_dim - 1)
+    omega = 1. / (temperature ** omega)
+
+    z = z.flatten()[:, None] * omega[None, :]
+    y = y.flatten()[:, None] * omega[None, :]
+    x = x.flatten()[:, None] * omega[None, :] 
+
+    pe = torch.cat((x.sin(), x.cos(), y.sin(), y.cos(), z.sin(), z.cos()), dim = 1)
+
+    pe = F.pad(pe, (0, dim - (fourier_dim * 6))) # pad if feature dimension not cleanly divisible by 6
+    return pe.type(dtype)
+
+# main class
+
+class Attend(Module):
+    def __init__(self, use_flash = False, config: Config = Config(True, True, True)):
+        super().__init__()
+        self.config = config
+        self.use_flash = use_flash
+        assert not (use_flash and version.parse(torch.__version__) < version.parse('2.0.0')), 'in order to use flash attention, you must be using pytorch 2.0 or above'
+
+    def flash_attn(self, q, k, v):
+        # flash attention - https://arxiv.org/abs/2205.14135
+        
+        with torch.backends.cuda.sdp_kernel(**self.config._asdict()):
+            out = F.scaled_dot_product_attention(q, k, v)
+
+        return out
+
+    def forward(self, q, k, v):
+        n, device, scale = q.shape[-2], q.device, q.shape[-1] ** -0.5
+
+        if self.use_flash:
+            return self.flash_attn(q, k, v)
+
+        # similarity
+
+        sim = einsum("b h i d, b j d -> b h i j", q, k) * scale
+
+        # attention
+
+        attn = sim.softmax(dim=-1)
+
+        # aggregate values
+
+        out = einsum("b h i j, b j d -> b h i d", attn, v)
+
+        return out
+
+# classes
+
+class FeedForward(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(Module):
+    def __init__(self, dim, heads = 8, dim_head = 64, use_flash = True):
+        super().__init__()
+        inner_dim = dim_head *  heads
+        self.heads = heads
+        self.scale = dim_head ** -0.5
+        self.norm = nn.LayerNorm(dim)
+
+        self.attend = Attend(use_flash = use_flash)
+
+        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)
+
+        out = self.attend(q, k, v)
+
+        out = rearrange(out, 'b h n d -> b n (h d)')
+        return self.to_out(out)
+
+class Transformer(Module):
+    def __init__(self, dim, depth, heads, dim_head, mlp_dim, use_flash):
+        super().__init__()
+        self.layers = ModuleList([])
+        for _ in range(depth):
+            self.layers.append(ModuleList([
+                Attention(dim, heads = heads, dim_head = dim_head, use_flash = use_flash),
+                FeedForward(dim, mlp_dim)
+            ]))
+
+    def forward(self, x):
+        for attn, ff in self.layers:
+            x = attn(x) + x
+            x = ff(x) + x
+
+        return x
+
+class SimpleViT(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, use_flash_attn = True):
+        super().__init__()
+        image_height, image_width = pair(image_size)
+        patch_height, patch_width = pair(image_patch_size)
+
+        assert image_height % patch_height == 0 and image_width % patch_width == 0, 'Image dimensions must be divisible by the patch size.'
+        assert frames % frame_patch_size == 0, 'Frames must be divisible by the frame patch size'
+
+        num_patches = (image_height // patch_height) * (image_width // patch_width) * (frames // frame_patch_size)
+        patch_dim = channels * patch_height * patch_width * frame_patch_size
+
+        self.to_patch_embedding = nn.Sequential(
+            Rearrange('b c (f pf) (h p1) (w p2) -> b f h w (p1 p2 pf c)', p1 = patch_height, p2 = patch_width, pf = frame_patch_size),
+            nn.LayerNorm(patch_dim),
+            nn.Linear(patch_dim, dim),
+            nn.LayerNorm(dim),
+        )
+
+        self.transformer = Transformer(dim, depth, heads, dim_head, mlp_dim, use_flash_attn)
+
+        self.to_latent = nn.Identity()
+        self.linear_head = nn.Linear(dim, num_classes)
+
+    def forward(self, video):
+        *_, h, w, dtype = *video.shape, video.dtype
+
+        x = self.to_patch_embedding(video)
+        pe = posemb_sincos_3d(x)
+        x = rearrange(x, 'b ... d -> b (...) d') + pe
+
+        x = self.transformer(x)
+        x = x.mean(dim = 1)
+
+        x = self.to_latent(x)
+        return self.linear_head(x)

vit_pytorch/simple_vit_with_fft.py

@@ -0,0 +1,162 @@
+import torch
+from torch.fft import fft2
+from torch import nn
+
+from einops import rearrange, reduce, 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, freq_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)
+        freq_patch_height, freq_patch_width = pair(freq_patch_size)
+
+        assert image_height % patch_height == 0 and image_width % patch_width == 0, 'Image dimensions must be divisible by the patch size.'
+        assert image_height % freq_patch_height == 0 and image_width % freq_patch_width == 0, 'Image dimensions must be divisible by the freq patch size.'
+
+        patch_dim = channels * patch_height * patch_width
+        freq_patch_dim = channels * 2 * freq_patch_height * freq_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.to_freq_embedding = nn.Sequential(
+            Rearrange("b c (h p1) (w p2) ri -> b (h w) (p1 p2 ri c)", p1 = freq_patch_height, p2 = freq_patch_width),
+            nn.LayerNorm(freq_patch_dim),
+            nn.Linear(freq_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.freq_pos_embedding = posemb_sincos_2d(
+            h = image_height // freq_patch_height,
+            w = image_width // freq_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):
+        device, dtype = img.device, img.dtype
+
+        x = self.to_patch_embedding(img)
+        freqs = torch.view_as_real(fft2(img))
+
+        f = self.to_freq_embedding(freqs)
+
+        x += self.pos_embedding.to(device, dtype = dtype)
+        f += self.freq_pos_embedding.to(device, dtype = dtype)
+
+        x, ps = pack((f, x), 'b * d')
+
+        x = self.transformer(x)
+
+        _, x = unpack(x, ps, 'b * d')
+        x = reduce(x, 'b n d -> b d', 'mean')
+
+        x = self.to_latent(x)
+        return self.linear_head(x)
+
+if __name__ == '__main__':
+    vit = SimpleViT(
+        num_classes = 1000,
+        image_size = 256,
+        patch_size = 8,
+        freq_patch_size = 8,
+        dim = 1024,
+        depth = 1,
+        heads = 8,
+        mlp_dim = 2048,
+    )
+
+    images = torch.randn(8, 3, 256, 256)
+
+    logits = vit(images)

vit_pytorch/t2t.py

@@ -61,10 +61,7 @@ class T2TViT(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

@@ -10,7 +10,7 @@ class FeedForward(nn.Module):
     def __init__(self, dim, hidden_dim, dropout = 0.):
         super().__init__()
         self.net = nn.Sequential(
-            nn.Layernorm(dim),
+            nn.LayerNorm(dim),
             nn.Linear(dim, hidden_dim),
             nn.GELU(),
             nn.Dropout(dropout),