add adaptive token sampling paper

README.md

@@ -16,6 +16,7 @@
 - [LeViT](#levit)
 - [CvT](#cvt)
 - [Twins SVT](#twins-svt)
+- [CrossFormer](#crossformer)
 - [RegionViT](#regionvit)
 - [NesT](#nest)
 - [Masked Autoencoder](#masked-autoencoder)
@@ -493,7 +494,7 @@ img = torch.randn(1, 3, 224, 224)
 pred = model(img) # (1, 1000)
 ```
 
-## CrossFormer (wip)
+## CrossFormer
 
 <img src="./images/crossformer.png" width="400px"></img>
 
@@ -595,6 +596,8 @@ A new <a href="https://arxiv.org/abs/2111.06377">Kaiming He paper</a> proposes a
 
 <a href="https://www.youtube.com/watch?v=LKixq2S2Pz8">DeepReader quick paper review</a>
 
+<a href="https://www.youtube.com/watch?v=Dp6iICL2dVI">AI Coffeebreak with Letitia</a>
+
 You can use it with the following code
 
 ```python
@@ -676,6 +679,39 @@ for _ in range(100):
 torch.save(model.state_dict(), './pretrained-net.pt')
 ```
 
+## Adaptive Token Sampling
+
+<img src="./images/ats.png" width="400px"></img>
+
+This <a href="https://arxiv.org/abs/2111.15667">paper</a> proposes to use the CLS attention scores, re-weighed by the norms of the value heads, as means to discard unimportant tokens at different layers.
+
+```python
+import torch
+from vit_pytorch.ats_vit import ViT
+
+v = ViT(
+    image_size = 256,
+    patch_size = 16,
+    num_classes = 1000,
+    dim = 1024,
+    depth = 6,
+    max_tokens_per_depth = (256, 128, 64, 32, 16, 8), # a tuple that denotes the maximum number of tokens that any given layer should have. if the layer has greater than this amount, it will undergo adaptive token sampling
+    heads = 16,
+    mlp_dim = 2048,
+    dropout = 0.1,
+    emb_dropout = 0.1
+)
+
+img = torch.randn(4, 3, 256, 256)
+
+preds = v(img) # (1, 1000)
+
+# you can also get a list of the final sampled patch ids
+# a value of -1 denotes padding
+
+preds, token_ids = v(img, return_sampled_token_ids = True) # (1, 1000), (1, <=8)
+```
+
 ## Dino
 
 <img src="./images/dino.png" width="350px"></img>
@@ -1116,6 +1152,17 @@ Coming from computer vision and new to transformers? Here are some resources tha
 }
 ```
 
+```bibtex
+@misc{fayyaz2021ats,
+    title   = {ATS: Adaptive Token Sampling For Efficient Vision Transformers},
+    author  = {Mohsen Fayyaz and Soroush Abbasi Kouhpayegani and Farnoush Rezaei Jafari and Eric Sommerlade and Hamid Reza Vaezi Joze and Hamed Pirsiavash and Juergen Gall},
+    year    = {2021},
+    eprint  = {2111.15667},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.CV}
+}
+```
+
 ```bibtex
 @misc{vaswani2017attention,
     title   = {Attention Is All You Need},

setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'vit-pytorch',
   packages = find_packages(exclude=['examples']),
-  version = '0.24.0',
+  version = '0.24.3',
   license='MIT',
   description = 'Vision Transformer (ViT) - Pytorch',
   author = 'Phil Wang',

vit_pytorch/ats_vit.py

@@ -0,0 +1,262 @@
+import torch
+import torch.nn.functional as F
+from torch.nn.utils.rnn import pad_sequence
+from torch import nn, einsum
+
+from einops import rearrange, repeat
+from einops.layers.torch import Rearrange
+
+# helpers
+
+def exists(val):
+    return val is not None
+
+def pair(t):
+    return t if isinstance(t, tuple) else (t, t)
+
+# adaptive token sampling functions and classes
+
+def log(t, eps = 1e-6):
+    return torch.log(t + eps)
+
+def sample_gumbel(shape, device, dtype, eps = 1e-6):
+    u = torch.empty(shape, device = device, dtype = dtype).uniform_(0, 1)
+    return -log(-log(u, eps), eps)
+
+def batched_index_select(values, indices, dim = 1):
+    value_dims = values.shape[(dim + 1):]
+    values_shape, indices_shape = map(lambda t: list(t.shape), (values, indices))
+    indices = indices[(..., *((None,) * len(value_dims)))]
+    indices = indices.expand(*((-1,) * len(indices_shape)), *value_dims)
+    value_expand_len = len(indices_shape) - (dim + 1)
+    values = values[(*((slice(None),) * dim), *((None,) * value_expand_len), ...)]
+
+    value_expand_shape = [-1] * len(values.shape)
+    expand_slice = slice(dim, (dim + value_expand_len))
+    value_expand_shape[expand_slice] = indices.shape[expand_slice]
+    values = values.expand(*value_expand_shape)
+
+    dim += value_expand_len
+    return values.gather(dim, indices)
+
+class AdaptiveTokenSampling(nn.Module):
+    def __init__(self, output_num_tokens, eps = 1e-6):
+        super().__init__()
+        self.eps = eps
+        self.output_num_tokens = output_num_tokens
+
+    def forward(self, attn, value, mask):
+        heads, output_num_tokens, eps, device, dtype = attn.shape[1], self.output_num_tokens, self.eps, attn.device, attn.dtype
+
+        # first get the attention values for CLS token to all other tokens
+
+        cls_attn = attn[..., 0, 1:]
+
+        # calculate the norms of the values, for weighting the scores, as described in the paper
+
+        value_norms = value[..., 1:, :].norm(dim = -1)
+
+        # weigh the attention scores by the norm of the values, sum across all heads
+
+        cls_attn = einsum('b h n, b h n -> b n', cls_attn, value_norms)
+
+        # normalize to 1
+
+        normed_cls_attn = cls_attn / (cls_attn.sum(dim = -1, keepdim = True) + eps)
+
+        # instead of using inverse transform sampling, going to invert the softmax and use gumbel-max sampling instead
+
+        pseudo_logits = log(normed_cls_attn)
+
+        # mask out pseudo logits for gumbel-max sampling
+
+        mask_without_cls = mask[:, 1:]
+        mask_value = -torch.finfo(attn.dtype).max / 2
+        pseudo_logits = pseudo_logits.masked_fill(~mask_without_cls, mask_value)
+
+        # expand k times, k being the adaptive sampling number
+
+        pseudo_logits = repeat(pseudo_logits, 'b n -> b k n', k = output_num_tokens)
+        pseudo_logits = pseudo_logits + sample_gumbel(pseudo_logits.shape, device = device, dtype = dtype)
+
+        # gumble-max and add one to reserve 0 for padding / mask
+
+        sampled_token_ids = pseudo_logits.argmax(dim = -1) + 1
+
+        # calculate unique using torch.unique and then pad the sequence from the right
+
+        unique_sampled_token_ids_list = [torch.unique(t, sorted = True) for t in torch.unbind(sampled_token_ids)]
+        unique_sampled_token_ids = pad_sequence(unique_sampled_token_ids_list, batch_first = True)
+
+        # calculate the new mask, based on the padding
+
+        new_mask = unique_sampled_token_ids != 0
+
+        # CLS token never gets masked out (gets a value of True)
+
+        new_mask = F.pad(new_mask, (1, 0), value = True)
+
+        # prepend a 0 token id to keep the CLS attention scores
+
+        unique_sampled_token_ids = F.pad(unique_sampled_token_ids, (1, 0), value = 0)
+        expanded_unique_sampled_token_ids = repeat(unique_sampled_token_ids, 'b n -> b h n', h = heads)
+
+        # gather the new attention scores
+
+        new_attn = batched_index_select(attn, expanded_unique_sampled_token_ids, dim = 2)
+
+        # return the sampled attention scores, new mask (denoting padding), as well as the sampled token indices (for the residual)
+        return new_attn, new_mask, unique_sampled_token_ids
+
+# 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.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)
+
+class Attention(nn.Module):
+    def __init__(self, dim, heads = 8, dim_head = 64, dropout = 0., output_num_tokens = None):
+        super().__init__()
+        inner_dim = dim_head *  heads
+        self.heads = heads
+        self.scale = dim_head ** -0.5
+
+        self.attend = nn.Softmax(dim = -1)
+        self.to_qkv = nn.Linear(dim, inner_dim * 3, bias = False)
+
+        self.output_num_tokens = output_num_tokens
+        self.ats = AdaptiveTokenSampling(output_num_tokens) if exists(output_num_tokens) else None
+
+        self.to_out = nn.Sequential(
+            nn.Linear(inner_dim, dim),
+            nn.Dropout(dropout)
+        )
+
+    def forward(self, x, *, mask):
+        num_tokens = x.shape[1]
+
+        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
+
+        if exists(mask):
+            dots_mask = rearrange(mask, 'b i -> b 1 i 1') * rearrange(mask, 'b j -> b 1 1 j')
+            mask_value = -torch.finfo(dots.dtype).max
+            dots = dots.masked_fill(~dots_mask, mask_value)
+
+        attn = self.attend(dots)
+
+        sampled_token_ids = None
+
+        # if adaptive token sampling is enabled
+        # and number of tokens is greater than the number of output tokens
+        if exists(self.output_num_tokens) and (num_tokens - 1) > self.output_num_tokens:
+            attn, mask, sampled_token_ids = self.ats(attn, v, mask = mask)
+
+        out = torch.matmul(attn, v)
+        out = rearrange(out, 'b h n d -> b n (h d)')
+
+        return self.to_out(out), mask, sampled_token_ids
+
+class Transformer(nn.Module):
+    def __init__(self, dim, depth, max_tokens_per_depth, heads, dim_head, mlp_dim, dropout = 0.):
+        super().__init__()
+        assert len(max_tokens_per_depth) == depth, 'max_tokens_per_depth must be a tuple of length that is equal to the depth of the transformer'
+        assert sorted(max_tokens_per_depth, reverse = True) == list(max_tokens_per_depth), 'max_tokens_per_depth must be in decreasing order'
+        assert min(max_tokens_per_depth) > 0, 'max_tokens_per_depth must have at least 1 token at any layer'
+
+        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))
+            ]))
+
+    def forward(self, x):
+        b, n, device = *x.shape[:2], x.device
+
+        # use mask to keep track of the paddings when sampling tokens
+        # as the duplicates (when sampling) are just removed, as mentioned in the paper
+        mask = torch.ones((b, n), device = device, dtype = torch.bool)
+
+        token_ids = torch.arange(n, device = device)
+        token_ids = repeat(token_ids, 'n -> b n', b = b)
+
+        for attn, ff in self.layers:
+            attn_out, mask, sampled_token_ids = attn(x, mask = mask)
+
+            # when token sampling, one needs to then gather the residual tokens with the sampled token ids
+            if exists(sampled_token_ids):
+                x = batched_index_select(x, sampled_token_ids, dim = 1)
+                token_ids = batched_index_select(token_ids, sampled_token_ids, dim = 1)
+
+            x = x + attn_out
+
+            x = ff(x) + x
+
+        return x, token_ids
+
+class ViT(nn.Module):
+    def __init__(self, *, image_size, patch_size, num_classes, dim, depth, max_tokens_per_depth, heads, mlp_dim, channels = 3, dim_head = 64, dropout = 0., emb_dropout = 0.):
+        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.'
+
+        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),
+            nn.Linear(patch_dim, dim),
+        )
+
+        self.pos_embedding = nn.Parameter(torch.randn(1, num_patches + 1, dim))
+        self.cls_token = nn.Parameter(torch.randn(1, 1, dim))
+        self.dropout = nn.Dropout(emb_dropout)
+
+        self.transformer = Transformer(dim, depth, max_tokens_per_depth, heads, dim_head, mlp_dim, dropout)
+
+        self.mlp_head = nn.Sequential(
+            nn.LayerNorm(dim),
+            nn.Linear(dim, num_classes)
+        )
+
+    def forward(self, img, return_sampled_token_ids = False):
+        x = self.to_patch_embedding(img)
+        b, n, _ = x.shape
+
+        cls_tokens = repeat(self.cls_token, '() n d -> b n d', b = b)
+        x = torch.cat((cls_tokens, x), dim=1)
+        x += self.pos_embedding[:, :(n + 1)]
+        x = self.dropout(x)
+
+        x, token_ids = self.transformer(x)
+
+        logits = self.mlp_head(x[:, 0])
+
+        if return_sampled_token_ids:
+            # remove CLS token and decrement by 1 to make -1 the padding
+            token_ids = token_ids[:, 1:] - 1
+            return logits, token_ids
+
+        return logits

vit_pytorch/crossformer.py

@@ -6,18 +6,9 @@ import torch.nn.functional as F
 
 # helpers
 
-def exists(val):
-    return val is not None
-
-def default(val, d):
-    return val if exists(val) else d
-
 def cast_tuple(val, length = 1):
     return val if isinstance(val, tuple) else ((val,) * length)
 
-def divisible_by(val, d):
-    return (val % d) == 0
-
 # cross embed layer
 
 class CrossEmbedLayer(nn.Module):
@@ -103,12 +94,25 @@ class Attention(nn.Module):
         self.attn_type = attn_type
         self.window_size = window_size
 
-        self.dpb = DynamicPositionBias(dim // 4)
-
         self.norm = LayerNorm(dim)
         self.to_qkv = nn.Conv2d(dim, inner_dim * 3, 1, bias = False)
         self.to_out = nn.Conv2d(inner_dim, dim, 1)
 
+        # positions
+
+        self.dpb = DynamicPositionBias(dim // 4)
+
+        # calculate and store indices for retrieving bias
+
+        pos = torch.arange(window_size)
+        grid = torch.stack(torch.meshgrid(pos, pos))
+        grid = rearrange(grid, 'c i j -> (i j) c')
+        rel_pos = grid[:, None] - grid[None, :]
+        rel_pos += window_size - 1
+        rel_pos_indices = (rel_pos * torch.tensor([2 * window_size - 1, 1])).sum(dim = -1)
+
+        self.register_buffer('rel_pos_indices', rel_pos_indices, persistent = False)
+
     def forward(self, x):
         *_, height, width, heads, wsz, device = *x.shape, self.heads, self.window_size, x.device
 
@@ -136,12 +140,11 @@ class Attention(nn.Module):
 
         # add dynamic positional bias
 
-        i_pos = torch.arange(wsz, device = device)
-        j_pos = torch.arange(wsz, device = device)
-        grid = torch.stack(torch.meshgrid(i_pos, j_pos))
-        grid = rearrange(grid, 'c i j -> (i j) c')
-        rel_ij = grid[:, None] - grid[None, :]
-        rel_pos_bias = self.dpb(rel_ij.float())
+        pos = torch.arange(-wsz, wsz + 1, device = device)
+        rel_pos = torch.stack(torch.meshgrid(pos, pos))
+        rel_pos = rearrange(rel_pos, 'c i j -> (i j) c')
+        biases = self.dpb(rel_pos.float())
+        rel_pos_bias = biases[self.rel_pos_indices]
 
         sim = sim + rel_pos_bias
 

vit_pytorch/pit.py

@@ -129,14 +129,15 @@ class PiT(nn.Module):
         mlp_dim,
         dim_head = 64,
         dropout = 0.,
-        emb_dropout = 0.
+        emb_dropout = 0.,
+        channels = 3
     ):
         super().__init__()
         assert image_size % patch_size == 0, 'Image dimensions must be divisible by the patch size.'
         assert isinstance(depth, tuple), 'depth must be a tuple of integers, specifying the number of blocks before each downsizing'
         heads = cast_tuple(heads, len(depth))
 
-        patch_dim = 3 * patch_size ** 2
+        patch_dim = channels * patch_size ** 2
 
         self.to_patch_embedding = nn.Sequential(
             nn.Unfold(kernel_size = patch_size, stride = patch_size // 2),