be able to accept non-square patches, thanks to @FilipAndersson245

Add `use_glu` flag to `RvT`

README.md

@@ -38,6 +38,7 @@ preds = v(img) # (1, 1000)
 ```
 
 ## Parameters
+
 - `image_size`: int.  
 Image size. If you have rectangular images, make sure your image size is the maximum of the width and height
 - `patch_size`: int.  
@@ -93,7 +94,8 @@ distiller = DistillWrapper(
     student = v,
     teacher = teacher,
     temperature = 3,           # temperature of distillation
-    alpha = 0.5                # trade between main loss and distillation loss
+    alpha = 0.5,               # trade between main loss and distillation loss
+    hard = False               # whether to use soft or hard distillation
 )
 
 img = torch.randn(2, 3, 256, 256)
@@ -160,12 +162,13 @@ v = CaiT(
     patch_size = 32,
     num_classes = 1000,
     dim = 1024,
-    depth = 12,         # depth of transformer for patch to patch attention only
-    cls_depth = 2,      # depth of cross attention of CLS tokens to patch
+    depth = 12,             # depth of transformer for patch to patch attention only
+    cls_depth = 2,          # depth of cross attention of CLS tokens to patch
     heads = 16,
     mlp_dim = 2048,
     dropout = 0.1,
-    emb_dropout = 0.1
+    emb_dropout = 0.1,
+    layer_dropout = 0.05    # randomly dropout 5% of the layers
 )
 
 img = torch.randn(1, 3, 256, 256)
@@ -262,6 +265,32 @@ img = torch.randn(1, 3, 224, 224)
 preds = v(img) # (1, 1000)
 ```
 
+## LeViT
+
+<img src="./images/levit.png" width="300px"></img>
+
+<a href="https://arxiv.org/abs/2104.01136">This paper</a> proposes a number of changes, including (1) convolutional embedding instead of patch-wise projection (2) downsampling in stages (3) extra non-linearity in attention (4) 2d relative positional biases instead of initial absolute positional bias (5) batchnorm in place of layernorm.
+
+```python
+import torch
+from vit_pytorch.levit import LeViT
+
+levit = LeViT(
+    image_size = 224,
+    num_classes = 1000,
+    stages = 3,             # number of stages
+    dim = (256, 384, 512),  # dimensions at each stage
+    depth = 4,              # transformer of depth 4 at each stage
+    heads = (4, 6, 8),      # heads at each stage
+    mlp_mult = 2,
+    dropout = 0.1
+)
+
+img = torch.randn(1, 3, 224, 224)
+
+levit(img) # (1, 1000)
+```
+
 ## CvT
 
 <img src="./images/cvt.png" width="400px"></img>
@@ -306,6 +335,47 @@ img = torch.randn(1, 3, 224, 224)
 pred = v(img) # (1, 1000)
 ```
 
+## Twins SVT
+
+<img src="./images/twins_svt.png" width="400px"></img>
+
+This <a href="https://arxiv.org/abs/2104.13840">paper</a> proposes mixing local and global attention, along with position encoding generator (proposed in <a href="https://arxiv.org/abs/2102.10882">CPVT</a>) and global average pooling, to achieve the same results as <a href="https://arxiv.org/abs/2103.14030">Swin</a>, without the extra complexity of shifted windows, CLS tokens, nor positional embeddings.
+
+```python
+import torch
+from vit_pytorch.twins_svt import TwinsSVT
+
+model = TwinsSVT(
+    num_classes = 1000,       # number of output classes
+    s1_emb_dim = 64,          # stage 1 - patch embedding projected dimension
+    s1_patch_size = 4,        # stage 1 - patch size for patch embedding
+    s1_local_patch_size = 7,  # stage 1 - patch size for local attention
+    s1_global_k = 7,          # stage 1 - global attention key / value reduction factor, defaults to 7 as specified in paper
+    s1_depth = 1,             # stage 1 - number of transformer blocks (local attn -> ff -> global attn -> ff)
+    s2_emb_dim = 128,         # stage 2 (same as above)
+    s2_patch_size = 2,
+    s2_local_patch_size = 7,
+    s2_global_k = 7,
+    s2_depth = 1,
+    s3_emb_dim = 256,         # stage 3 (same as above)
+    s3_patch_size = 2,
+    s3_local_patch_size = 7,
+    s3_global_k = 7,
+    s3_depth = 5,
+    s4_emb_dim = 512,         # stage 4 (same as above)
+    s4_patch_size = 2,
+    s4_local_patch_size = 7,
+    s4_global_k = 7,
+    s4_depth = 4,
+    peg_kernel_size = 3,      # positional encoding generator kernel size
+    dropout = 0.              # dropout
+)
+
+img = torch.randn(1, 3, 224, 224)
+
+pred = model(img) # (1, 1000)
+```
+
 ## Masked Patch Prediction
 
 Thanks to <a href="https://github.com/zankner">Zach</a>, you can train using the original masked patch prediction task presented in the paper, with the following code.
@@ -514,6 +584,58 @@ img = torch.randn(1, 3, 224, 224)
 v(img) # (1, 1000)
 ```
 
+## FAQ
+
+- How do I pass in non-square images?
+
+You can already pass in non-square images - you just have to make sure your height and width is less than or equal to the `image_size`, and both divisible by the `patch_size`
+
+ex.
+
+```python
+import torch
+from vit_pytorch import ViT
+
+v = ViT(
+    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
+)
+
+img = torch.randn(1, 3, 256, 128) # <-- not a square
+
+preds = v(img) # (1, 1000)
+```
+
+- How do I pass in non-square patches?
+
+```python
+import torch
+from vit_pytorch import ViT
+
+v = ViT(
+    num_classes = 1000,
+    image_size = (256, 128),  # image size is a tuple of (height, width)
+    patch_size = (32, 16),    # patch size is a tuple of (height, width)
+    dim = 1024,
+    depth = 6,
+    heads = 16,
+    mlp_dim = 2048,
+    dropout = 0.1,
+    emb_dropout = 0.1
+)
+
+img = torch.randn(1, 3, 256, 128)
+
+preds = v(img)
+```
+
 ## Resources
 
 Coming from computer vision and new to transformers? Here are some resources that greatly accelerated my learning.
@@ -571,6 +693,17 @@ Coming from computer vision and new to transformers? Here are some resources tha
 }
 ```
 
+```bibtex
+@misc{touvron2021going,
+    title   = {Going deeper with Image Transformers}, 
+    author  = {Hugo Touvron and Matthieu Cord and Alexandre Sablayrolles and Gabriel Synnaeve and Hervé Jégou},
+    year    = {2021},
+    eprint  = {2103.17239},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.CV}
+}
+```
+
 ```bibtex
 @misc{chen2021crossvit,
     title   = {CrossViT: Cross-Attention Multi-Scale Vision Transformer for Image Classification},
@@ -604,6 +737,50 @@ Coming from computer vision and new to transformers? Here are some resources tha
 }
 ```
 
+```bibtex
+@misc{graham2021levit,
+    title   = {LeViT: a Vision Transformer in ConvNet's Clothing for Faster Inference},
+    author  = {Ben Graham and Alaaeldin El-Nouby and Hugo Touvron and Pierre Stock and Armand Joulin and Hervé Jégou and Matthijs Douze},
+    year    = {2021},
+    eprint  = {2104.01136},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.CV}
+}
+```
+
+```bibtex
+@misc{li2021localvit,
+    title   = {LocalViT: Bringing Locality to Vision Transformers},
+    author  = {Yawei Li and Kai Zhang and Jiezhang Cao and Radu Timofte and Luc Van Gool},
+    year    = {2021},
+    eprint  = {2104.05707},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.CV}
+}
+```
+
+```bibtex
+@misc{chu2021twins,
+    title   = {Twins: Revisiting Spatial Attention Design in Vision Transformers},
+    author  = {Xiangxiang Chu and Zhi Tian and Yuqing Wang and Bo Zhang and Haibing Ren and Xiaolin Wei and Huaxia Xia and Chunhua Shen},
+    year    = {2021},
+    eprint  = {2104.13840},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.CV}
+}
+```
+
+```bibtex
+@misc{su2021roformer,
+    title   = {RoFormer: Enhanced Transformer with Rotary Position Embedding}, 
+    author  = {Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu},
+    year    = {2021},
+    eprint  = {2104.09864},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.CL}
+}
+```
+
 ```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 = '0.14.1',
+  version = '0.17.3',
   license='MIT',
   description = 'Vision Transformer (ViT) - Pytorch',
   author = 'Phil Wang',

vit_pytorch/cait.py

@@ -1,3 +1,4 @@
+from random import randrange
 import torch
 from torch import nn, einsum
 import torch.nn.functional as F
@@ -10,11 +11,33 @@ from einops.layers.torch import Rearrange
 def exists(val):
     return val is not None
 
+def dropout_layers(layers, dropout):
+    if dropout == 0:
+        return layers
+
+    num_layers = len(layers)
+    to_drop = torch.zeros(num_layers).uniform_(0., 1.) < dropout
+
+    # make sure at least one layer makes it
+    if all(to_drop):
+        rand_index = randrange(num_layers)
+        to_drop[rand_index] = False
+
+    layers = [layer for (layer, drop) in zip(layers, to_drop) if not drop]
+    return layers
+
 # classes
 
 class LayerScale(nn.Module):
-    def __init__(self, dim, fn, init_eps = 0.1):
+    def __init__(self, dim, fn, depth):
         super().__init__()
+        if depth <= 18:  # epsilon detailed in section 2 of paper
+            init_eps = 0.1
+        elif depth > 18 and depth <= 24:
+            init_eps = 1e-5
+        else:
+            init_eps = 1e-6
+
         scale = torch.zeros(1, 1, dim).fill_(init_eps)
         self.scale = nn.Parameter(scale)
         self.fn = fn
@@ -81,16 +104,20 @@ class Attention(nn.Module):
         return self.to_out(out)
 
 class Transformer(nn.Module):
-    def __init__(self, dim, depth, heads, dim_head, mlp_dim, dropout = 0.):
+    def __init__(self, dim, depth, heads, dim_head, mlp_dim, dropout = 0., layer_dropout = 0.):
         super().__init__()
         self.layers = nn.ModuleList([])
-        for _ in range(depth):
+        self.layer_dropout = layer_dropout
+
+        for ind in range(depth):
             self.layers.append(nn.ModuleList([
-                LayerScale(dim, PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout))),
-                LayerScale(dim, PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout)))
+                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)
             ]))
     def forward(self, x, context = None):
-        for attn, ff in self.layers:
+        layers = dropout_layers(self.layers, dropout = self.layer_dropout)
+
+        for attn, ff in layers:
             x = attn(x, context = context) + x
             x = ff(x) + x
         return x
@@ -109,7 +136,8 @@ class CaiT(nn.Module):
         mlp_dim,
         dim_head = 64,
         dropout = 0.,
-        emb_dropout = 0.
+        emb_dropout = 0.,
+        layer_dropout = 0.
     ):
         super().__init__()
         assert image_size % patch_size == 0, 'Image dimensions must be divisible by the patch size.'
@@ -123,10 +151,11 @@ class CaiT(nn.Module):
 
         self.pos_embedding = nn.Parameter(torch.randn(1, num_patches, dim))
         self.cls_token = nn.Parameter(torch.randn(1, 1, dim))
+
         self.dropout = nn.Dropout(emb_dropout)
 
-        self.patch_transformer = Transformer(dim, depth, heads, dim_head, mlp_dim, dropout)
-        self.cls_transformer = Transformer(dim, cls_depth, heads, dim_head, mlp_dim, dropout)
+        self.patch_transformer = Transformer(dim, depth, heads, dim_head, mlp_dim, dropout, layer_dropout)
+        self.cls_transformer = Transformer(dim, cls_depth, heads, dim_head, mlp_dim, dropout, layer_dropout)
 
         self.mlp_head = nn.Sequential(
             nn.LayerNorm(dim),

vit_pytorch/cvt.py

@@ -22,15 +22,25 @@ def group_by_key_prefix_and_remove_prefix(prefix, d):
 
 # classes
 
+class LayerNorm(nn.Module): # layernorm, but done in the channel dimension #1
+    def __init__(self, dim, eps = 1e-5):
+        super().__init__()
+        self.eps = eps
+        self.g = nn.Parameter(torch.ones(1, dim, 1, 1))
+        self.b = nn.Parameter(torch.zeros(1, dim, 1, 1))
+
+    def forward(self, x):
+        std = torch.var(x, dim = 1, unbiased = False, keepdim = True).sqrt()
+        mean = torch.mean(x, dim = 1, keepdim = True)
+        return (x - mean) / (std + self.eps) * self.g + self.b
+
 class PreNorm(nn.Module):
     def __init__(self, dim, fn):
         super().__init__()
-        self.norm = nn.LayerNorm(dim)
+        self.norm = 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):
@@ -67,8 +77,8 @@ class Attention(nn.Module):
 
         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_q = DepthWiseConv2d(dim, inner_dim, proj_kernel, padding = padding, stride = 1, bias = False)
+        self.to_kv = DepthWiseConv2d(dim, inner_dim * 2, proj_kernel, padding = padding, stride = kv_proj_stride, bias = False)
 
         self.to_out = nn.Sequential(
             nn.Conv2d(inner_dim, dim, 1),
@@ -130,7 +140,7 @@ class CvT(nn.Module):
         s3_emb_stride = 2,
         s3_proj_kernel = 3,
         s3_kv_proj_stride = 2,
-        s3_heads = 4,
+        s3_heads = 6,
         s3_depth = 10,
         s3_mlp_mult = 4,
         dropout = 0.
@@ -146,6 +156,7 @@ class CvT(nn.Module):
 
             layers.append(nn.Sequential(
                 nn.Conv2d(dim, config['emb_dim'], kernel_size = config['emb_kernel'], padding = (config['emb_kernel'] // 2), stride = config['emb_stride']),
+                LayerNorm(config['emb_dim']),
                 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)
             ))
 

vit_pytorch/distill.py

@@ -104,7 +104,8 @@ class DistillWrapper(nn.Module):
         teacher,
         student,
         temperature = 1.,
-        alpha = 0.5
+        alpha = 0.5,
+        hard = False
     ):
         super().__init__()
         assert (isinstance(student, (DistillableViT, DistillableT2TViT, DistillableEfficientViT))) , 'student must be a vision transformer'
@@ -116,6 +117,7 @@ class DistillWrapper(nn.Module):
         num_classes = student.num_classes
         self.temperature = temperature
         self.alpha = alpha
+        self.hard = hard
 
         self.distillation_token = nn.Parameter(torch.randn(1, 1, dim))
 
@@ -137,11 +139,15 @@ class DistillWrapper(nn.Module):
 
         loss = F.cross_entropy(student_logits, labels)
 
-        distill_loss = F.kl_div(
-            F.log_softmax(distill_logits / T, dim = -1),
-            F.softmax(teacher_logits / T, dim = -1).detach(),
-        reduction = 'batchmean')
+        if not self.hard:
+            distill_loss = F.kl_div(
+                F.log_softmax(distill_logits / T, dim = -1),
+                F.softmax(teacher_logits / T, dim = -1).detach(),
+            reduction = 'batchmean')
+            distill_loss *= T ** 2
 
-        distill_loss *= T ** 2
+        else:
+            teacher_labels = teacher_logits.argmax(dim = -1)
+            distill_loss = F.cross_entropy(student_logits, teacher_labels)
 
-        return loss * alpha + distill_loss * (1 - alpha)
+        return loss * (1 - alpha) + distill_loss * alpha

vit_pytorch/levit.py

@@ -0,0 +1,193 @@
+from math import ceil
+
+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
+
+def cast_tuple(val, l = 3):
+    val = val if isinstance(val, tuple) else (val,)
+    return (*val, *((val[-1],) * max(l - len(val), 0)))
+
+def always(val):
+    return lambda *args, **kwargs: val
+
+# classes
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, mult, 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 Attention(nn.Module):
+    def __init__(self, dim, fmap_size, heads = 8, dim_key = 32, dim_value = 64, dropout = 0., dim_out = None, downsample = False):
+        super().__init__()
+        inner_dim_key = dim_key *  heads
+        inner_dim_value = dim_value *  heads
+        dim_out = default(dim_out, dim)
+
+        self.heads = heads
+        self.scale = dim_key ** -0.5
+
+        self.to_q = nn.Sequential(nn.Conv2d(dim, inner_dim_key, 1, stride = (2 if downsample else 1), bias = False), nn.BatchNorm2d(inner_dim_key))
+        self.to_k = nn.Sequential(nn.Conv2d(dim, inner_dim_key, 1, bias = False), nn.BatchNorm2d(inner_dim_key))
+        self.to_v = nn.Sequential(nn.Conv2d(dim, inner_dim_value, 1, bias = False), nn.BatchNorm2d(inner_dim_value))
+
+        self.attend = nn.Softmax(dim = -1)
+
+        out_batch_norm = nn.BatchNorm2d(dim_out)
+        nn.init.zeros_(out_batch_norm.weight)
+
+        self.to_out = nn.Sequential(
+            nn.GELU(),
+            nn.Conv2d(inner_dim_value, dim_out, 1),
+            out_batch_norm,
+            nn.Dropout(dropout)
+        )
+
+        # positional bias
+
+        self.pos_bias = nn.Embedding(fmap_size * fmap_size, heads)
+
+        q_range = torch.arange(0, fmap_size, step = (2 if downsample else 1))
+        k_range = torch.arange(fmap_size)
+
+        q_pos = torch.stack(torch.meshgrid(q_range, q_range), dim = -1)
+        k_pos = torch.stack(torch.meshgrid(k_range, k_range), dim = -1)
+
+        q_pos, k_pos = map(lambda t: rearrange(t, 'i j c -> (i j) c'), (q_pos, k_pos))
+        rel_pos = (q_pos[:, None, ...] - k_pos[None, :, ...]).abs()
+
+        x_rel, y_rel = rel_pos.unbind(dim = -1)
+        pos_indices = (x_rel * fmap_size) + y_rel
+
+        self.register_buffer('pos_indices', pos_indices)
+
+    def apply_pos_bias(self, fmap):
+        bias = self.pos_bias(self.pos_indices)
+        bias = rearrange(bias, 'i j h -> () h i j')
+        return fmap + bias
+
+    def forward(self, x):
+        b, n, *_, h = *x.shape, self.heads
+
+        q = self.to_q(x)
+        y = q.shape[2]
+
+        qkv = (q, self.to_k(x), self.to_v(x))
+        q, k, v = map(lambda t: rearrange(t, 'b (h d) ... -> b h (...) d', h = h), qkv)
+
+        dots = einsum('b h i d, b h j d -> b h i j', q, k) * self.scale
+
+        dots = self.apply_pos_bias(dots)
+
+        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 (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, fmap_size, depth, heads, dim_key, dim_value, mlp_mult = 2, dropout = 0., dim_out = None, downsample = False):
+        super().__init__()
+        dim_out = default(dim_out, dim)
+        self.layers = nn.ModuleList([])
+        self.attn_residual = (not downsample) and dim == dim_out
+
+        for _ in range(depth):
+            self.layers.append(nn.ModuleList([
+                Attention(dim, fmap_size = fmap_size, heads = heads, dim_key = dim_key, dim_value = dim_value, dropout = dropout, downsample = downsample, dim_out = dim_out),
+                FeedForward(dim_out, mlp_mult, dropout = dropout)
+            ]))
+    def forward(self, x):
+        for attn, ff in self.layers:
+            attn_res = (x if self.attn_residual else 0)
+            x = attn(x) + attn_res
+            x = ff(x) + x
+        return x
+
+class LeViT(nn.Module):
+    def __init__(
+        self,
+        *,
+        image_size,
+        num_classes,
+        dim,
+        depth,
+        heads,
+        mlp_mult,
+        stages = 3,
+        dim_key = 32,
+        dim_value = 64,
+        dropout = 0.,
+        num_distill_classes = None
+    ):
+        super().__init__()
+
+        dims = cast_tuple(dim, stages)
+        depths = cast_tuple(depth, stages)
+        layer_heads = cast_tuple(heads, stages)
+
+        assert all(map(lambda t: len(t) == stages, (dims, depths, layer_heads))), 'dimensions, depths, and heads must be a tuple that is less than the designated number of stages'
+
+        self.conv_embedding = nn.Sequential(
+            nn.Conv2d(3, 32, 3, stride = 2, padding = 1),
+            nn.Conv2d(32, 64, 3, stride = 2, padding = 1),
+            nn.Conv2d(64, 128, 3, stride = 2, padding = 1),
+            nn.Conv2d(128, dims[0], 3, stride = 2, padding = 1)
+        )
+
+        fmap_size = image_size // (2 ** 4)
+        layers = []
+
+        for ind, dim, depth, heads in zip(range(stages), dims, depths, layer_heads):
+            is_last = ind == (stages - 1)
+            layers.append(Transformer(dim, fmap_size, depth, heads, dim_key, dim_value, mlp_mult, dropout))
+
+            if not is_last:
+                next_dim = dims[ind + 1]
+                layers.append(Transformer(dim, fmap_size, 1, heads * 2, dim_key, dim_value, dim_out = next_dim, downsample = True))
+                fmap_size = ceil(fmap_size / 2)
+
+        self.backbone = nn.Sequential(*layers)
+
+        self.pool = nn.Sequential(
+            nn.AdaptiveAvgPool2d(1),
+            Rearrange('... () () -> ...')
+        )
+
+        self.distill_head = nn.Linear(dim, num_distill_classes) if exists(num_distill_classes) else always(None)
+        self.mlp_head = nn.Linear(dim, num_classes)
+
+    def forward(self, img):
+        x = self.conv_embedding(img)
+
+        x = self.backbone(x)        
+
+        x = self.pool(x)
+
+        out = self.mlp_head(x)
+        distill = self.distill_head(x)
+
+        if exists(distill):
+            return out, distill
+
+        return out

vit_pytorch/local_vit.py

@@ -0,0 +1,152 @@
+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
+
+# classes
+
+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 ExcludeCLS(nn.Module):
+    def __init__(self, fn):
+        super().__init__()
+        self.fn = fn
+
+    def forward(self, x, **kwargs):
+        cls_token, x = x[:, :1], x[:, 1:]
+        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):
+    def __init__(self, dim_in, dim_out, kernel_size, padding, stride = 1, 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)
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, hidden_dim, dropout = 0.):
+        super().__init__()
+        self.net = nn.Sequential(
+            nn.Conv2d(dim, hidden_dim, 1),
+            nn.Hardswish(),
+            DepthWiseConv2d(hidden_dim, hidden_dim, 3, padding = 1),
+            nn.Hardswish(),
+            nn.Dropout(dropout),
+            nn.Conv2d(hidden_dim, dim, 1),
+            nn.Dropout(dropout)
+        )
+    def forward(self, x):
+        h = w = int(sqrt(x.shape[-2]))
+        x = rearrange(x, 'b (h w) c -> b c h w', h = h, w = w)
+        x = self.net(x)
+        x = rearrange(x, 'b c h w -> b (h w) c')
+        return 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_qkv = nn.Linear(dim, inner_dim * 3, bias = False)
+
+        self.to_out = nn.Sequential(
+            nn.Linear(inner_dim, dim),
+            nn.Dropout(dropout)
+        )
+
+    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([
+                Residual(PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout))),
+                ExcludeCLS(Residual(PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout))))
+            ]))
+    def forward(self, x):
+        for attn, ff in self.layers:
+            x = attn(x)
+            x = ff(x)
+        return x
+
+# main class
+
+class LocalViT(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.):
+        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 = 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),
+            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 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)]
+        x = self.dropout(x)
+
+        x = self.transformer(x)
+
+        return self.mlp_head(x[:, 0])

vit_pytorch/pit.py

@@ -89,8 +89,8 @@ 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)
+            nn.Conv2d(dim_in, dim_out, kernel_size = kernel_size, padding = padding, groups = dim_in, stride = stride, bias = bias),
+            nn.Conv2d(dim_out, dim_out, kernel_size = 1, bias = bias)
         )
     def forward(self, x):
         return self.net(x)
@@ -162,8 +162,9 @@ class PiT(nn.Module):
                 layers.append(Pool(dim))
                 dim *= 2
 
-        self.layers = nn.Sequential(
-            *layers,
+        self.layers = nn.Sequential(*layers)
+
+        self.mlp_head = nn.Sequential(
             nn.LayerNorm(dim),
             nn.Linear(dim, num_classes)
         )
@@ -177,4 +178,6 @@ class PiT(nn.Module):
         x += self.pos_embedding
         x = self.dropout(x)
 
-        return self.layers(x)
+        x = self.layers(x)
+
+        return self.mlp_head(x[:, 0])

vit_pytorch/rvt.py

@@ -0,0 +1,209 @@
+from math import sqrt, pi, log
+
+import torch
+from torch import nn, einsum
+import torch.nn.functional as F
+
+from einops import rearrange, repeat
+from einops.layers.torch import Rearrange
+
+# rotary embeddings
+
+def rotate_every_two(x):
+    x = rearrange(x, '... (d j) -> ... d j', j = 2)
+    x1, x2 = x.unbind(dim = -1)
+    x = torch.stack((-x2, x1), dim = -1)
+    return rearrange(x, '... d j -> ... (d j)')
+
+class AxialRotaryEmbedding(nn.Module):
+    def __init__(self, dim, max_freq = 10):
+        super().__init__()
+        self.dim = dim
+        scales = torch.logspace(0., log(max_freq / 2) / log(2), self.dim // 4, base = 2)
+        self.register_buffer('scales', scales)
+
+    def forward(self, x):
+        device, dtype, n = x.device, x.dtype, int(sqrt(x.shape[-2]))
+
+        seq = torch.linspace(-1., 1., steps = n, device = device)
+        seq = seq.unsqueeze(-1)
+
+        scales = self.scales[(*((None,) * (len(seq.shape) - 1)), Ellipsis)]
+        scales = scales.to(x)
+
+        seq = seq * scales * pi
+
+        x_sinu = repeat(seq, 'i d -> i j d', j = n)
+        y_sinu = repeat(seq, 'j d -> i j d', i = n)
+
+        sin = torch.cat((x_sinu.sin(), y_sinu.sin()), dim = -1)
+        cos = torch.cat((x_sinu.cos(), y_sinu.cos()), dim = -1)
+
+        sin, cos = map(lambda t: rearrange(t, 'i j d -> (i j) d'), (sin, cos))
+        sin, cos = map(lambda t: repeat(t, 'n d -> () n (d j)', j = 2), (sin, cos))
+        return sin, cos
+
+class DepthWiseConv2d(nn.Module):
+    def __init__(self, dim_in, dim_out, kernel_size, padding, stride = 1, 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)
+
+# 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__()
+        self.conv = DepthWiseConv2d(dim_in, dim_out, kernel, padding = kernel // 2, bias = False)
+        self.cls_proj = nn.Linear(dim_in, dim_out) if dim_in != dim_out else nn.Identity()
+
+    def forward(self, x, fmap_dims):
+        cls_token, x = x[:, :1], x[:, 1:]
+        x = rearrange(x, 'b (h w) d -> b d h w', **fmap_dims)
+        x = self.conv(x)
+        x = rearrange(x, 'b d h w -> b (h w) d')
+        cls_token = self.cls_proj(cls_token)
+        return torch.cat((cls_token, x), dim = 1)
+
+class GEGLU(nn.Module):
+    def forward(self, x):
+        x, gates = x.chunk(2, dim = -1)
+        return F.gelu(gates) * x
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, hidden_dim, dropout = 0., use_glu = True):
+        super().__init__()
+        self.net = nn.Sequential(
+            nn.Linear(dim, hidden_dim * 2 if use_glu else hidden_dim),
+            GEGLU() if use_glu else 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., use_rotary = True, use_ds_conv = True, conv_query_kernel = 5):
+        super().__init__()
+        inner_dim = dim_head *  heads
+        self.use_rotary = use_rotary
+        self.heads = heads
+        self.scale = dim_head ** -0.5
+
+        self.attend = nn.Softmax(dim = -1)
+
+        self.use_ds_conv = use_ds_conv
+
+        self.to_q = SpatialConv(dim, inner_dim, conv_query_kernel, bias = False) if use_ds_conv else 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, pos_emb, fmap_dims):
+        b, n, _, h = *x.shape, self.heads
+
+        to_q_kwargs = {'fmap_dims': fmap_dims} if self.use_ds_conv else {}
+        q = self.to_q(x, **to_q_kwargs)
+
+        qkv = (q, *self.to_kv(x).chunk(2, dim = -1))
+
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h = h), qkv)
+
+        if self.use_rotary:
+            # apply 2d rotary embeddings to queries and keys, excluding CLS tokens
+
+            sin, cos = pos_emb
+            dim_rotary = sin.shape[-1]
+
+            (q_cls, q), (k_cls, k) = map(lambda t: (t[:, :1], t[:, 1:]), (q, k))
+
+            # handle the case where rotary dimension < head dimension
+
+            (q, q_pass), (k, k_pass) = map(lambda t: (t[..., :dim_rotary], t[..., dim_rotary:]), (q, k))
+            q, k = map(lambda t: (t * cos) + (rotate_every_two(t) * sin), (q, k))
+            q, k = map(lambda t: torch.cat(t, dim = -1), ((q, q_pass), (k, k_pass)))
+
+            # concat back the CLS tokens
+
+            q = torch.cat((q_cls, q), dim = 1)
+            k = torch.cat((k_cls, k), dim = 1)
+
+        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) n d -> b n (h d)', h = h)
+        return self.to_out(out)
+
+class Transformer(nn.Module):
+    def __init__(self, dim, depth, heads, dim_head, mlp_dim, dropout = 0., use_rotary = True, use_ds_conv = True, use_glu = True):
+        super().__init__()
+        self.layers = nn.ModuleList([])
+        self.pos_emb = AxialRotaryEmbedding(dim_head)
+        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))
+            ]))
+    def forward(self, x, fmap_dims):
+        pos_emb = self.pos_emb(x[:, 1:])
+
+        for attn, ff in self.layers:
+            x = attn(x, pos_emb = pos_emb, fmap_dims = fmap_dims) + x
+            x = ff(x) + x
+        return x
+
+# Rotary Vision Transformer
+
+class RvT(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., use_rotary = True, use_ds_conv = True, use_glu = True):
+        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 = channels * patch_size ** 2
+
+        self.patch_size = patch_size
+        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.cls_token = nn.Parameter(torch.randn(1, 1, dim))
+        self.transformer = Transformer(dim, depth, heads, dim_head, mlp_dim, dropout, use_rotary, use_ds_conv, use_glu)
+
+        self.mlp_head = nn.Sequential(
+            nn.LayerNorm(dim),
+            nn.Linear(dim, num_classes)
+        )
+
+    def forward(self, img):
+        b, _, h, w, p = *img.shape, self.patch_size
+
+        x = self.to_patch_embedding(img)
+        n = x.shape[1]
+
+        cls_tokens = repeat(self.cls_token, '() n d -> b n d', b = b)
+        x = torch.cat((cls_tokens, x), dim=1)
+
+        fmap_dims = {'h': h // p, 'w': w // p}
+        x = self.transformer(x, fmap_dims = fmap_dims)
+
+        return self.mlp_head(x[:, 0])

vit_pytorch/twins_svt.py

@@ -0,0 +1,229 @@
+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 Residual(nn.Module):
+    def __init__(self, fn):
+        super().__init__()
+        self.fn = fn
+
+    def forward(self, x, **kwargs):
+        return self.fn(x, **kwargs) + x
+
+class LayerNorm(nn.Module):
+    def __init__(self, dim, eps = 1e-5):
+        super().__init__()
+        self.eps = eps
+        self.g = nn.Parameter(torch.ones(1, dim, 1, 1))
+        self.b = nn.Parameter(torch.zeros(1, dim, 1, 1))
+
+    def forward(self, x):
+        std = torch.var(x, dim = 1, unbiased = False, keepdim = True).sqrt()
+        mean = torch.mean(x, dim = 1, keepdim = True)
+        return (x - mean) / (std + self.eps) * 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(
+            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 PatchEmbedding(nn.Module):
+    def __init__(self, *, dim, dim_out, patch_size):
+        super().__init__()
+        self.dim = dim
+        self.dim_out = dim_out
+        self.patch_size = patch_size
+        self.proj = nn.Conv2d(patch_size ** 2 * dim, dim_out, 1)
+
+    def forward(self, fmap):
+        p = self.patch_size
+        fmap = rearrange(fmap, 'b c (h p1) (w p2) -> b (c p1 p2) h w', p1 = p, p2 = p)
+        return self.proj(fmap)
+
+class PEG(nn.Module):
+    def __init__(self, dim, kernel_size = 3):
+        super().__init__()
+        self.proj = Residual(nn.Conv2d(dim, dim, kernel_size = kernel_size, padding = kernel_size // 2, groups = dim, stride = 1))
+
+    def forward(self, x):
+        return self.proj(x)
+
+class LocalAttention(nn.Module):
+    def __init__(self, dim, heads = 8, dim_head = 64, dropout = 0., patch_size = 7):
+        super().__init__()
+        inner_dim = dim_head *  heads
+        self.patch_size = patch_size
+        self.heads = heads
+        self.scale = dim_head ** -0.5
+
+        self.to_q = nn.Conv2d(dim, inner_dim, 1, bias = False)
+        self.to_kv = nn.Conv2d(dim, inner_dim * 2, 1, bias = False)
+
+        self.to_out = nn.Sequential(
+            nn.Conv2d(inner_dim, dim, 1),
+            nn.Dropout(dropout)
+        )
+
+    def forward(self, 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))
+
+        fmap = rearrange(fmap, 'b c (x p1) (y p2) -> (b x y) c p1 p2', p1 = p, p2 = p)
+
+        q, k, v = (self.to_q(fmap), *self.to_kv(fmap).chunk(2, dim = 1))
+        q, k, v = map(lambda t: rearrange(t, 'b (h d) p1 p2 -> (b h) (p1 p2) d', h = h), (q, k, v))
+
+        dots = einsum('b i d, b j d -> b i j', q, k) * self.scale
+
+        attn = dots.softmax(dim = - 1)
+
+        out = einsum('b i j, b j d -> b i d', attn, v)
+        out = rearrange(out, '(b x y h) (p1 p2) d -> b (h d) (x p1) (y p2)', h = h, x = x, y = y, p1 = p, p2 = p)
+        return self.to_out(out)
+
+class GlobalAttention(nn.Module):
+    def __init__(self, dim, heads = 8, dim_head = 64, dropout = 0., k = 7):
+        super().__init__()
+        inner_dim = dim_head *  heads
+        self.heads = heads
+        self.scale = dim_head ** -0.5
+
+        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)
+
+        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 = dots.softmax(dim = -1)
+
+        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, depth, heads = 8, dim_head = 64, mlp_mult = 4, local_patch_size = 7, global_k = 7, dropout = 0., has_local = True):
+        super().__init__()
+        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)))
+            ]))
+    def forward(self, x):
+        for local_attn, ff1, global_attn, ff2 in self.layers:
+            x = local_attn(x)
+            x = ff1(x)
+            x = global_attn(x)
+            x = ff2(x)
+        return x
+
+class TwinsSVT(nn.Module):
+    def __init__(
+        self,
+        *,
+        num_classes,
+        s1_emb_dim = 64,
+        s1_patch_size = 4,
+        s1_local_patch_size = 7,
+        s1_global_k = 7,
+        s1_depth = 1,
+        s2_emb_dim = 128,
+        s2_patch_size = 2,
+        s2_local_patch_size = 7,
+        s2_global_k = 7,
+        s2_depth = 1,
+        s3_emb_dim = 256,
+        s3_patch_size = 2,
+        s3_local_patch_size = 7,
+        s3_global_k = 7,
+        s3_depth = 5,
+        s4_emb_dim = 512,
+        s4_patch_size = 2,
+        s4_local_patch_size = 7,
+        s4_global_k = 7,
+        s4_depth = 4,
+        peg_kernel_size = 3,
+        dropout = 0.
+    ):
+        super().__init__()
+        kwargs = dict(locals())
+
+        dim = 3
+        layers = []
+
+        for prefix in ('s1', 's2', 's3', 's4'):
+            config, kwargs = group_by_key_prefix_and_remove_prefix(f'{prefix}_', kwargs)
+            is_last = prefix == 's4'
+
+            dim_next = config['emb_dim']
+
+            layers.append(nn.Sequential(
+                PatchEmbedding(dim = dim, dim_out = dim_next, patch_size = config['patch_size']),
+                Transformer(dim = dim_next, depth = 1, local_patch_size = config['local_patch_size'], global_k = config['global_k'], dropout = dropout, has_local = not is_last),
+                PEG(dim = dim_next, kernel_size = peg_kernel_size),
+                Transformer(dim = dim_next, depth = config['depth'],  local_patch_size = config['local_patch_size'], global_k = config['global_k'], dropout = dropout, has_local = not is_last)
+            ))
+
+            dim = dim_next
+
+        self.layers = nn.Sequential(
+            *layers,
+            nn.AdaptiveAvgPool2d(1),
+            Rearrange('... () () -> ...'),
+            nn.Linear(dim, num_classes)
+        )
+
+    def forward(self, x):
+        return self.layers(x)

vit_pytorch/vit.py

@@ -5,6 +5,13 @@ import torch.nn.functional as F
 from einops import rearrange, repeat
 from einops.layers.torch import Rearrange
 
+# helpers
+
+def pair(t):
+    return t if isinstance(t, tuple) else (t, t)
+
+# classes
+
 class PreNorm(nn.Module):
     def __init__(self, dim, fn):
         super().__init__()
@@ -74,13 +81,17 @@ class Transformer(nn.Module):
 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__()
-        assert image_size % patch_size == 0, 'Image dimensions must be divisible by the patch size.'
-        num_patches = (image_size // patch_size) ** 2
-        patch_dim = channels * patch_size ** 2
+        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_size, p2 = patch_size),
+            Rearrange('b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1 = patch_height, p2 = patch_width),
             nn.Linear(patch_dim, dim),
         )