complete learnable memory ViT, for efficient fine-tuning and potentially plays into continual learning

README.md

@@ -28,6 +28,7 @@
 - [Patch Merger](#patch-merger)
 - [Vision Transformer for Small Datasets](#vision-transformer-for-small-datasets)
 - [Parallel ViT](#parallel-vit)
+- [Learnable Memory ViT](#learnable-memory-vit)
 - [Dino](#dino)
 - [Accessing Attention](#accessing-attention)
 - [Research Ideas](#research-ideas)
@@ -903,6 +904,61 @@ img = torch.randn(4, 3, 256, 256)
 preds = v(img) # (4, 1000)
 ```
 
+## Learnable Memory ViT
+
+<img src="./images/learnable-memory-vit.png" width="350px"></img>
+
+This <a href="https://arxiv.org/abs/2203.15243">paper</a> shows that adding learnable memory tokens at each layer of a vision transformer can greatly enhance fine-tuning results (in addition to learnable task specific CLS token and adapter head).
+
+You can use this with a specially modified `ViT` as follows
+
+```python
+import torch
+from vit_pytorch.learnable_memory_vit import ViT, Adapter
+
+# normal base ViT
+
+v = ViT(
+    image_size = 256,
+    patch_size = 16,
+    num_classes = 1000,
+    dim = 1024,
+    depth = 6,
+    heads = 8,
+    mlp_dim = 2048,
+    dropout = 0.1,
+    emb_dropout = 0.1
+)
+
+img = torch.randn(4, 3, 256, 256)
+logits = v(img) # (4, 1000)
+
+# do your usual training with ViT
+# ...
+
+
+# then, to finetune, just pass the ViT into the Adapter class
+# you can do this for multiple Adapters, as shown below
+
+adapter1 = Adapter(
+    vit = v,
+    num_classes = 2,               # number of output classes for this specific task
+    num_memories_per_layer = 5     # number of learnable memories per layer, 10 was sufficient in paper
+)
+
+logits1 = adapter1(img) # (4, 2) - predict 2 classes off frozen ViT backbone with learnable memories and task specific head
+
+# yet another task to finetune on, this time with 4 classes
+
+adapter2 = Adapter(
+    vit = v,
+    num_classes = 4,
+    num_memories_per_layer = 10
+)
+
+logits2 = adapter2(img) # (4, 4) - predict 4 classes off frozen ViT backbone with learnable memories and task specific head
+
+```
 
 ## Dino
 
@@ -1442,6 +1498,14 @@ Coming from computer vision and new to transformers? Here are some resources tha
 }
 ```
 
+```bibtex
+@inproceedings{Sandler2022FinetuningIT,
+    title   = {Fine-tuning Image Transformers using Learnable Memory},
+    author  = {Mark Sandler and Andrey Zhmoginov and Max Vladymyrov and Andrew Jackson},
+    year    = {2022}
+}
+```
+
 ```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 = '0.30.0',
+  version = '0.31.0',
   license='MIT',
   description = 'Vision Transformer (ViT) - Pytorch',
   author = 'Phil Wang',

vit_pytorch/learnable_memory_vit.py

@@ -0,0 +1,216 @@
+import torch
+from torch import nn
+import torch.nn.functional as F
+
+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)
+
+# controlling freezing of layers
+
+def set_module_requires_grad_(module, requires_grad):
+    for param in module.parameters():
+        param.requires_grad = requires_grad
+
+def freeze_all_layers_(module):
+    set_module_requires_grad_(module, False)
+
+def unfreeze_all_layers_(module):
+    set_module_requires_grad_(module, True)
+
+# classes
+
+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)
+
+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.scale = dim_head ** -0.5
+        self.norm = nn.LayerNorm(dim)
+
+        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),
+            nn.Dropout(dropout)
+        )
+
+    def forward(self, x, attn_mask = None, memories = None):
+        x = self.norm(x)
+
+        x_kv = x # input for key / values projection
+
+        if exists(memories):
+            # add memories to key / values if it is passed in
+            memories = repeat(memories, 'n d -> b n d', b = x.shape[0]) if memories.ndim == 2 else memories
+            x_kv = torch.cat((x_kv, memories), dim = 1)
+
+        qkv = (self.to_q(x), *self.to_kv(x_kv).chunk(2, 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(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)
+            ]))
+
+    def forward(self, x, attn_mask = None, memories = None):
+        for ind, (attn, ff) in enumerate(self.layers):
+            layer_memories = memories[ind] if exists(memories) else None
+
+            x = attn(x, attn_mask = attn_mask, memories = layer_memories) + x
+            x = ff(x) + x
+        return 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.):
+        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
+        assert pool in {'cls', 'mean'}, 'pool type must be either cls (cls token) or mean (mean pooling)'
+
+        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, heads, dim_head, mlp_dim, dropout)
+
+        self.mlp_head = nn.Sequential(
+            nn.LayerNorm(dim),
+            nn.Linear(dim, num_classes)
+        )
+
+    def img_to_tokens(self, img):
+        x = self.to_patch_embedding(img)
+
+        cls_tokens = repeat(self.cls_token, '1 n d -> b n d', b = x.shape[0])
+        x = torch.cat((cls_tokens, x), dim = 1)
+
+        x += self.pos_embedding
+        x = self.dropout(x)
+        return x
+
+    def forward(self, img):
+        x = self.img_to_tokens(img)        
+
+        x = self.transformer(x)
+
+        cls_tokens = x[:, 0]
+        return self.mlp_head(cls_tokens)
+
+# adapter with learnable memories per layer, memory CLS token, and learnable adapter head
+
+class Adapter(nn.Module):
+    def __init__(
+        self,
+        *,
+        vit,
+        num_memories_per_layer = 10,
+        num_classes = 2,   
+    ):
+        super().__init__()
+        assert isinstance(vit, ViT)
+
+        # extract some model variables needed
+
+        dim = vit.cls_token.shape[-1]
+        layers = len(vit.transformer.layers)
+        num_patches = vit.pos_embedding.shape[-2]
+
+        self.vit = vit
+
+        # freeze ViT backbone - only memories will be finetuned
+
+        freeze_all_layers_(vit)
+
+        # learnable parameters
+
+        self.memory_cls_token = nn.Parameter(torch.randn(dim))
+        self.memories_per_layer = nn.Parameter(torch.randn(layers, num_memories_per_layer, dim))
+
+        self.mlp_head = nn.Sequential(
+            nn.LayerNorm(dim),
+            nn.Linear(dim, num_classes)
+        )
+
+        # specialized attention mask to preserve the output of the original ViT
+        # it allows the memory CLS token to attend to all other tokens (and the learnable memory layer tokens), but not vice versa        
+
+        attn_mask = torch.ones((num_patches, num_patches), dtype = torch.bool)
+        attn_mask = F.pad(attn_mask, (1, num_memories_per_layer), value = False)  # main tokens cannot attend to learnable memories per layer
+        attn_mask = F.pad(attn_mask, (0, 0, 1, 0), value = True)                  # memory CLS token can attend to everything
+        self.register_buffer('attn_mask', attn_mask)
+
+    def forward(self, img):
+        b = img.shape[0]
+
+        tokens = self.vit.img_to_tokens(img)
+
+        # add task specific memory tokens
+
+        memory_cls_tokens = repeat(self.memory_cls_token, 'd -> b 1 d', b = b)
+        tokens = torch.cat((memory_cls_tokens, tokens), dim = 1)        
+
+        # pass memories along with image tokens through transformer for attending
+
+        out = self.vit.transformer(tokens, memories = self.memories_per_layer, attn_mask = self.attn_mask)
+
+        # extract memory CLS tokens
+
+        memory_cls_tokens = out[:, 0]
+
+        # pass through task specific adapter head
+
+        return self.mlp_head(memory_cls_tokens)

vit_pytorch/vit.py

@@ -114,7 +114,7 @@ class ViT(nn.Module):
         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)]
         x = self.dropout(x)