add working PiT

README.md

@@ -432,6 +432,28 @@ Coming from computer vision and new to transformers? Here are some resources tha
 }
 ```
 
+```bibtex
+@misc{chen2021crossvit,
+    title   = {CrossViT: Cross-Attention Multi-Scale Vision Transformer for Image Classification},
+    author  = {Chun-Fu Chen and Quanfu Fan and Rameswar Panda},
+    year    = {2021},
+    eprint  = {2103.14899},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.CV}
+}
+```
+
+```bibtex
+@misc{wu2021cvt,
+    title   = {CvT: Introducing Convolutions to Vision Transformers},
+    author  = {Haiping Wu and Bin Xiao and Noel Codella and Mengchen Liu and Xiyang Dai and Lu Yuan and Lei Zhang},
+    year    = {2021},
+    eprint  = {2103.15808},
+    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 = '0.10.1',
+  version = '0.12.0',
   license='MIT',
   description = 'Vision Transformer (ViT) - Pytorch',
   author = 'Phil Wang',

vit_pytorch/cross_vit.py

@@ -0,0 +1,270 @@
+import torch
+from torch import nn, einsum
+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 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.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)
+
+# attention
+
+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.attend = nn.Softmax(dim = -1)
+        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, context = None, kv_include_self = False):
+        b, n, _, h = *x.shape, self.heads
+        context = default(context, x)
+
+        if kv_include_self:
+            context = torch.cat((x, context), dim = 1) # cross attention requires CLS token includes itself as key / value
+
+        qkv = (self.to_q(x), *self.to_kv(context).chunk(2, dim = -1))
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = h), qkv)
+
+        dots = einsum('b h i d, b h j d -> b h i j', q, k) * self.scale
+
+        attn = self.attend(dots)
+
+        out = einsum('b h i j, b h j d -> b h i d', attn, v)
+        out = rearrange(out, 'b h n d -> b n (h d)')
+        return self.to_out(out)
+
+# transformer encoder, for small and large patches
+
+class Transformer(nn.Module):
+    def __init__(self, dim, depth, heads, dim_head, mlp_dim, dropout = 0.):
+        super().__init__()
+        self.layers = nn.ModuleList([])
+        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))
+            ]))
+
+    def forward(self, x):
+        for attn, ff in self.layers:
+            x = attn(x) + x
+            x = ff(x) + x
+        return self.norm(x)
+
+# projecting CLS tokens, in the case that small and large patch tokens have different dimensions
+
+class ProjectInOut(nn.Module):
+    def __init__(self, dim_in, dim_out, fn):
+        super().__init__()
+        self.fn = fn
+
+        need_projection = dim_in != dim_out
+        self.project_in = nn.Linear(dim_in, dim_out) if need_projection else nn.Identity()
+        self.project_out = nn.Linear(dim_out, dim_in) if need_projection else nn.Identity()
+
+    def forward(self, x, *args, **kwargs):
+        x = self.project_in(x)
+        x = self.fn(x, *args, **kwargs)
+        x = self.project_out(x)
+        return x
+
+# cross attention transformer
+
+class CrossTransformer(nn.Module):
+    def __init__(self, sm_dim, lg_dim, depth, heads, dim_head, dropout):
+        super().__init__()
+        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)))
+            ]))
+
+    def forward(self, sm_tokens, lg_tokens):
+        (sm_cls, sm_patch_tokens), (lg_cls, lg_patch_tokens) = map(lambda t: (t[:, :1], t[:, 1:]), (sm_tokens, lg_tokens))
+
+        for sm_attend_lg, lg_attend_sm in self.layers:
+            sm_cls = sm_attend_lg(sm_cls, context = lg_patch_tokens, kv_include_self = True) + sm_cls
+            lg_cls = lg_attend_sm(lg_cls, context = sm_patch_tokens, kv_include_self = True) + lg_cls
+
+        sm_tokens = torch.cat((sm_cls, sm_patch_tokens), dim = 1)
+        lg_tokens = torch.cat((lg_cls, lg_patch_tokens), dim = 1)
+        return sm_tokens, lg_tokens
+
+# multi-scale encoder
+
+class MultiScaleEncoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        depth,
+        sm_dim,
+        lg_dim,
+        sm_enc_params,
+        lg_enc_params,
+        cross_attn_heads,
+        cross_attn_depth,
+        cross_attn_dim_head = 64,
+        dropout = 0.
+    ):
+        super().__init__()
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(nn.ModuleList([
+                Transformer(dim = sm_dim, dropout = dropout, **sm_enc_params),
+                Transformer(dim = lg_dim, dropout = dropout, **lg_enc_params),
+                CrossTransformer(sm_dim = sm_dim, lg_dim = lg_dim, depth = cross_attn_depth, heads = cross_attn_heads, dim_head = cross_attn_dim_head, dropout = dropout)
+            ]))
+
+    def forward(self, sm_tokens, lg_tokens):
+        for sm_enc, lg_enc, cross_attend in self.layers:
+            sm_tokens, lg_tokens = sm_enc(sm_tokens), lg_enc(lg_tokens)
+            sm_tokens, lg_tokens = cross_attend(sm_tokens, lg_tokens)
+
+        return sm_tokens, lg_tokens
+
+# patch-based image to token embedder
+
+class ImageEmbedder(nn.Module):
+    def __init__(
+        self,
+        *,
+        dim,
+        image_size,
+        patch_size,
+        dropout = 0.
+    ):
+        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
+
+        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),
+            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(dropout)
+
+    def forward(self, img):
+        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)]
+
+        return self.dropout(x)
+
+# cross ViT class
+
+class CrossViT(nn.Module):
+    def __init__(
+        self,
+        *,
+        image_size,
+        num_classes,
+        sm_dim,
+        lg_dim,
+        sm_patch_size = 12,
+        sm_enc_depth = 1,
+        sm_enc_heads = 8,
+        sm_enc_mlp_dim = 2048,
+        sm_enc_dim_head = 64,
+        lg_patch_size = 16,
+        lg_enc_depth = 4,
+        lg_enc_heads = 8,
+        lg_enc_mlp_dim = 2048,
+        lg_enc_dim_head = 64,
+        cross_attn_depth = 2,
+        cross_attn_heads = 8,
+        cross_attn_dim_head = 64,
+        depth = 3,
+        dropout = 0.1,
+        emb_dropout = 0.1
+    ):
+        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.multi_scale_encoder = MultiScaleEncoder(
+            depth = depth,
+            sm_dim = sm_dim,
+            lg_dim = lg_dim,
+            cross_attn_heads = cross_attn_heads,
+            cross_attn_dim_head = cross_attn_dim_head,
+            cross_attn_depth = cross_attn_depth,
+            sm_enc_params = dict(
+                depth = sm_enc_depth,
+                heads = sm_enc_heads,
+                mlp_dim = sm_enc_mlp_dim,
+                dim_head = sm_enc_dim_head
+            ),
+            lg_enc_params = dict(
+                depth = lg_enc_depth,
+                heads = lg_enc_heads,
+                mlp_dim = lg_enc_mlp_dim,
+                dim_head = lg_enc_dim_head
+            ),
+            dropout = dropout
+        )
+
+        self.sm_mlp_head = nn.Sequential(nn.LayerNorm(sm_dim), nn.Linear(sm_dim, num_classes))
+        self.lg_mlp_head = nn.Sequential(nn.LayerNorm(lg_dim), nn.Linear(lg_dim, num_classes))
+
+    def forward(self, img):
+        sm_tokens = self.sm_image_embedder(img)
+        lg_tokens = self.lg_image_embedder(img)
+
+        sm_tokens, lg_tokens = self.multi_scale_encoder(sm_tokens, lg_tokens)
+
+        sm_cls, lg_cls = map(lambda t: t[:, 0], (sm_tokens, lg_tokens))
+
+        sm_logits = self.sm_mlp_head(sm_cls)
+        lg_logits = self.lg_mlp_head(lg_cls)
+
+        return sm_logits + lg_logits

vit_pytorch/cvt.py

@@ -0,0 +1,162 @@
+import torch
+from torch import nn, einsum
+import torch.nn.functional as F
+
+from einops import rearrange, repeat
+from einops.layers.torch import Rearrange
+
+# helper methods
+
+def group_dict_by_key(cond, d):
+    return_val = [dict(), dict()]
+    for key in d.keys():
+        match = bool(cond(key))
+        ind = int(not match)
+        return_val[ind][key] = d[key]
+    return (*return_val,)
+
+def group_by_key_prefix_and_remove_prefix(prefix, d):
+    kwargs_with_prefix, kwargs = group_dict_by_key(lambda x: x.startswith(prefix), d)
+    kwargs_without_prefix = dict(map(lambda x: (x[0][len(prefix):], x[1]), tuple(kwargs_with_prefix.items())))
+    return kwargs_without_prefix, kwargs
+
+# 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):
+        x = rearrange(x, 'b c h w -> b h w c')
+        x = self.norm(x)
+        x = rearrange(x, 'b h w c -> b c h w')
+        return self.fn(x, **kwargs)
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, mult = 4, dropout = 0.):
+        super().__init__()
+        self.net = nn.Sequential(
+            nn.Conv2d(dim, dim * mult, 1),
+            nn.GELU(),
+            nn.Dropout(dropout),
+            nn.Conv2d(dim * mult, dim, 1),
+            nn.Dropout(dropout)
+        )
+    def forward(self, x):
+        return self.net(x)
+
+class DepthWiseConv2d(nn.Module):
+    def __init__(self, dim_in, dim_out, kernel_size, padding, stride, bias = True):
+        super().__init__()
+        self.net = nn.Sequential(
+            nn.Conv2d(dim_in, dim_in, kernel_size = kernel_size, padding = padding, groups = dim_in, stride = stride, bias = bias),
+            nn.BatchNorm2d(dim_in),
+            nn.Conv2d(dim_in, dim_out, kernel_size = 1, bias = bias)
+        )
+    def forward(self, x):
+        return self.net(x)
+
+class Attention(nn.Module):
+    def __init__(self, dim, proj_kernel, kv_proj_stride, heads = 8, dim_head = 64, dropout = 0.):
+        super().__init__()
+        inner_dim = dim_head *  heads
+        padding = proj_kernel // 2
+        self.heads = heads
+        self.scale = dim_head ** -0.5
+
+        self.attend = nn.Softmax(dim = -1)
+
+        self.to_q = DepthWiseConv2d(dim, inner_dim, 3, padding = padding, stride = 1, bias = False)
+        self.to_kv = DepthWiseConv2d(dim, inner_dim * 2, 3, padding = padding, stride = kv_proj_stride, bias = False)
+
+        self.to_out = nn.Sequential(
+            nn.Conv2d(inner_dim, dim, 1),
+            nn.Dropout(dropout)
+        )
+
+    def forward(self, 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))
+        q, k, v = map(lambda t: rearrange(t, 'b (h d) x y -> (b h) (x y) d', h = h), (q, k, v))
+
+        dots = einsum('b i d, b j d -> b i j', q, k) * self.scale
+
+        attn = self.attend(dots)
+
+        out = einsum('b i j, b j d -> b i d', attn, v)
+        out = rearrange(out, '(b h) (x y) d -> b (h d) x y', h = h, y = y)
+        return self.to_out(out)
+
+class Transformer(nn.Module):
+    def __init__(self, dim, proj_kernel, kv_proj_stride, depth, heads, dim_head = 64, mlp_mult = 4, dropout = 0.):
+        super().__init__()
+        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))
+            ]))
+    def forward(self, x):
+        for attn, ff in self.layers:
+            x = attn(x) + x
+            x = ff(x) + x
+        return x
+
+class CvT(nn.Module):
+    def __init__(
+        self,
+        *,
+        num_classes,
+        s1_emb_dim = 64,
+        s1_emb_kernel = 7,
+        s1_emb_stride = 4,
+        s1_proj_kernel = 3,
+        s1_kv_proj_stride = 2,
+        s1_heads = 1,
+        s1_depth = 1,
+        s1_mlp_mult = 4,
+        s2_emb_dim = 192,
+        s2_emb_kernel = 3,
+        s2_emb_stride = 2,
+        s2_proj_kernel = 3,
+        s2_kv_proj_stride = 2,
+        s2_heads = 3,
+        s2_depth = 2,
+        s2_mlp_mult = 4,
+        s3_emb_dim = 384,
+        s3_emb_kernel = 3,
+        s3_emb_stride = 2,
+        s3_proj_kernel = 3,
+        s3_kv_proj_stride = 2,
+        s3_heads = 4,
+        s3_depth = 10,
+        s3_mlp_mult = 4,
+        dropout = 0.
+    ):
+        super().__init__()
+        kwargs = dict(locals())
+
+        dim = 3
+        layers = []
+
+        for prefix in ('s1', 's2', 's3'):
+            config, kwargs = group_by_key_prefix_and_remove_prefix(f'{prefix}_', kwargs)
+
+            layers.append(nn.Sequential(
+                nn.Conv2d(dim, config['emb_dim'], kernel_size = config['emb_kernel'], padding = (config['emb_kernel'] // 2), stride = config['emb_stride']),
+                Transformer(dim = config['emb_dim'], proj_kernel = config['proj_kernel'], kv_proj_stride = config['kv_proj_stride'], depth = config['depth'], heads = config['heads'], mlp_mult = config['mlp_mult'], dropout = dropout)
+            ))
+
+            dim = config['emb_dim']
+
+        self.layers = nn.Sequential(
+            *layers,
+            nn.AdaptiveAvgPool2d(1),
+            Rearrange('... () () -> ...'),
+            nn.Linear(dim, num_classes)
+        )
+
+    def forward(self, x):
+        return self.layers(x)

vit_pytorch/pit.py

@@ -0,0 +1,180 @@
+from math import sqrt
+
+import torch
+from torch import nn, einsum
+import torch.nn.functional as F
+
+from einops import rearrange, repeat
+from einops.layers.torch import Rearrange
+
+# helpers
+
+def cast_tuple(val, num):
+    return val if isinstance(val, tuple) else (val,) * num
+
+def conv_output_size(image_size, kernel_size, stride, padding = 0):
+    return int(((image_size - kernel_size + (2 * padding)) / 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.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
+        project_out = not (heads == 1 and dim_head == dim)
+
+        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.to_out = nn.Sequential(
+            nn.Linear(inner_dim, dim),
+            nn.Dropout(dropout)
+        ) if project_out else nn.Identity()
+
+    def forward(self, x):
+        b, n, _, h = *x.shape, self.heads
+        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)
+
+        dots = einsum('b h i d, b h j d -> b h i j', q, k) * self.scale
+
+        attn = self.attend(dots)
+
+        out = einsum('b h i j, b h j d -> b h i d', 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([
+                PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout)),
+                PreNorm(dim, 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
+
+# depthwise convolution, for pooling
+
+class DepthWiseConv2d(nn.Module):
+    def __init__(self, dim_in, dim_out, kernel_size, padding, stride, bias = True):
+        super().__init__()
+        self.net = nn.Sequential(
+            nn.Conv2d(dim_in, dim_in, kernel_size = kernel_size, padding = padding, groups = dim_in, stride = stride, bias = bias),
+            nn.Conv2d(dim_in, dim_out, kernel_size = 1, bias = bias)
+        )
+    def forward(self, x):
+        return self.net(x)
+
+# pooling layer
+
+class Pool(nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.downsample = DepthWiseConv2d(dim, dim * 2, kernel_size = 3, stride = 2, padding = 1)
+        self.cls_ff = nn.Linear(dim, dim * 2)
+
+    def forward(self, x):
+        cls_token, tokens = x[:, :1], x[:, 1:]
+
+        cls_token = self.cls_ff(cls_token)
+
+        tokens = rearrange(tokens, 'b (h w) c -> b c h w', h = int(sqrt(tokens.shape[1])))
+        tokens = self.downsample(tokens)
+        tokens = rearrange(tokens, 'b c h w -> b (h w) c')
+
+        return torch.cat((cls_token, tokens), dim = 1)
+
+# main class
+
+class PiT(nn.Module):
+    def __init__(
+        self,
+        *,
+        image_size,
+        patch_size,
+        num_classes,
+        dim,
+        depth,
+        heads,
+        mlp_dim,
+        dim_head = 64,
+        dropout = 0.,
+        emb_dropout = 0.
+    ):
+        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
+
+        self.to_patch_embedding = nn.Sequential(
+            nn.Unfold(kernel_size = patch_size, stride = patch_size // 2),
+            Rearrange('b c n -> b n c'),
+            nn.Linear(patch_dim, dim)
+        )
+
+        output_size = conv_output_size(image_size, patch_size, patch_size // 2)
+        num_patches = output_size ** 2
+
+        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)
+
+        layers = []
+
+        for ind, (layer_depth, layer_heads) in enumerate(zip(depth, heads)):
+            not_last = ind < (len(depth) - 1)
+            
+            layers.append(Transformer(dim, layer_depth, layer_heads, dim_head, mlp_dim, dropout))
+
+            if not_last:
+                layers.append(Pool(dim))
+                dim *= 2
+
+        self.layers = nn.Sequential(
+            *layers,
+            nn.LayerNorm(dim),
+            nn.Linear(dim, num_classes)
+        )
+
+    def forward(self, img):
+        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
+        x = self.dropout(x)
+
+        return self.layers(x)

vit_pytorch/recorder.py

@@ -4,9 +4,6 @@ from torch import nn
 
 from vit_pytorch.vit import Attention
 
-def exists(val):
-    return val is not None
-
 def find_modules(nn_module, type):
     return [module for module in nn_module.modules() if isinstance(module, type)]
 
@@ -25,7 +22,7 @@ class Recorder(nn.Module):
         self.recordings.append(output.clone().detach())
 
     def _register_hook(self):
-        modules = find_modules(self, Attention)
+        modules = find_modules(self.vit.transformer, Attention)
         for module in modules:
             handle = module.attend.register_forward_hook(self._hook)
             self.hooks.append(handle)