diff --git a/models.py b/models.py
index b262ffd..ccb1ce9 100644
--- a/models.py
+++ b/models.py
@@ -4,6 +4,7 @@ import torch.nn.functional as F
 import numpy as np
 from functools import partial
 from einops.layers.torch import Rearrange, Reduce
+from einops import rearrange, repeat
 from utils import *
 from layers import *
 from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
@@ -868,6 +869,203 @@ def window_reverse(windows, window_size, H, W):
     x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, -1, H, W)
     return x
 
+def cast_tuple(val, length = 1):
+    return val if isinstance(val, tuple) else ((val,) * length)
+
+# helper classes
+
+class ChanLayerNorm(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):
+        var = torch.var(x, dim = 1, unbiased = False, keepdim = True)
+        mean = torch.mean(x, dim = 1, keepdim = True)
+        return (x - mean) / (var + self.eps).sqrt() * self.g + self.b
+
+class OverlappingPatchEmbed(nn.Module):
+    def __init__(self, dim_in, dim_out, stride = 2):
+        super().__init__()
+        kernel_size = stride * 2 - 1
+        padding = kernel_size // 2
+        self.conv = nn.Conv2d(dim_in, dim_out, kernel_size, stride = stride, padding = padding)
+
+    def forward(self, x):
+        return self.conv(x)
+
+class PEG(nn.Module):
+    def __init__(self, dim, kernel_size = 3):
+        super().__init__()
+        self.proj = nn.Conv2d(dim, dim, kernel_size = kernel_size, padding = kernel_size // 2, groups = dim, stride = 1)
+
+    def forward(self, x):
+        return self.proj(x) + x
+
+# feedforward
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, mult = 4, dropout = 0.):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        self.net = nn.Sequential(
+            ChanLayerNorm(dim),
+            nn.Conv2d(dim, inner_dim, 1),
+            nn.GELU(),
+            nn.Dropout(dropout),
+            nn.Conv2d(inner_dim, dim, 1),
+            nn.Dropout(dropout)
+        )
+    def forward(self, x):
+        return self.net(x)
+
+# attention
+
+class DSSA(nn.Module):
+    def __init__(
+        self,
+        dim,
+        heads = 8,
+        dim_head = 32,
+        dropout = 0.,
+        window_size = 7
+    ):
+        super().__init__()
+        self.heads = heads
+        self.scale = dim_head ** -0.5
+        self.window_size = window_size
+        inner_dim = dim_head * heads
+
+        self.norm = ChanLayerNorm(dim)
+
+        self.attend = nn.Sequential(
+            nn.Softmax(dim = -1),
+            nn.Dropout(dropout)
+        )
+
+        self.to_qkv = nn.Conv1d(dim, inner_dim * 3, 1, bias = False)
+
+        # window tokens
+
+        self.window_tokens = nn.Parameter(torch.randn(dim))
+
+        # prenorm and non-linearity for window tokens
+        # then projection to queries and keys for window tokens
+
+        self.window_tokens_to_qk = nn.Sequential(
+            nn.LayerNorm(dim_head),
+            nn.GELU(),
+            Rearrange('b h n c -> b (h c) n'),
+            nn.Conv1d(inner_dim, inner_dim * 2, 1),
+            Rearrange('b (h c) n -> b h n c', h = heads),
+        )
+
+        # window attention
+
+        self.window_attend = nn.Sequential(
+            nn.Softmax(dim = -1),
+            nn.Dropout(dropout)
+        )
+
+        self.to_out = nn.Sequential(
+            nn.Conv2d(inner_dim, dim, 1),
+            nn.Dropout(dropout)
+        )
+
+    def forward(self, x):
+        """
+        einstein notation
+
+        b - batch
+        c - channels
+        w1 - window size (height)
+        w2 - also window size (width)
+        i - sequence dimension (source)
+        j - sequence dimension (target dimension to be reduced)
+        h - heads
+        x - height of feature map divided by window size
+        y - width of feature map divided by window size
+        """
+
+        batch, height, width, heads, wsz = x.shape[0], *x.shape[-2:], self.heads, self.window_size
+        assert (height % wsz) == 0 and (width % wsz) == 0, f'height {height} and width {width} must be divisible by window size {wsz}'
+        num_windows = (height // wsz) * (width // wsz)
+
+        x = self.norm(x)
+
+        # fold in windows for "depthwise" attention - not sure why it is named depthwise when it is just "windowed" attention
+
+        x = rearrange(x, 'b c (h w1) (w w2) -> (b h w) c (w1 w2)', w1 = wsz, w2 = wsz)
+
+        # add windowing tokens
+
+        w = repeat(self.window_tokens, 'c -> b c 1', b = x.shape[0])
+        x = torch.cat((w, x), dim = -1)
+
+        # project for queries, keys, value
+
+        q, k, v = self.to_qkv(x).chunk(3, dim = 1)
+
+        # split out heads
+
+        q, k, v = map(lambda t: rearrange(t, 'b (h d) ... -> b h (...) d', h = heads), (q, k, v))
+
+        # scale
+
+        q = q * self.scale
+
+        # similarity
+
+        dots = einsum('b h i d, b h j d -> b h i j', q, k)
+
+        # attention
+
+        attn = self.attend(dots)
+
+        # aggregate values
+
+        out = torch.matmul(attn, v)
+
+        # split out windowed tokens
+
+        window_tokens, windowed_fmaps = out[:, :, 0], out[:, :, 1:]
+
+        # early return if there is only 1 window
+
+        if num_windows == 1:
+            fmap = rearrange(windowed_fmaps, '(b x y) h (w1 w2) d -> b (h d) (x w1) (y w2)', x = height // wsz, y = width // wsz, w1 = wsz, w2 = wsz)
+            return self.to_out(fmap)
+
+        # carry out the pointwise attention, the main novelty in the paper
+
+        window_tokens = rearrange(window_tokens, '(b x y) h d -> b h (x y) d', x = height // wsz, y = width // wsz)
+        windowed_fmaps = rearrange(windowed_fmaps, '(b x y) h n d -> b h (x y) n d', x = height // wsz, y = width // wsz)
+
+        # windowed queries and keys (preceded by prenorm activation)
+
+        w_q, w_k = self.window_tokens_to_qk(window_tokens).chunk(2, dim = -1)
+
+        # scale
+
+        w_q = w_q * self.scale
+
+        # similarities
+
+        w_dots = einsum('b h i d, b h j d -> b h i j', w_q, w_k)
+
+        w_attn = self.window_attend(w_dots)
+
+        # aggregate the feature maps from the "depthwise" attention step (the most interesting part of the paper, one i haven't seen before)
+
+        aggregated_windowed_fmap = einsum('b h i j, b h j w d -> b h i w d', w_attn, windowed_fmaps)
+
+        # fold back the windows and then combine heads for aggregation
+
+        fmap = rearrange(aggregated_windowed_fmap, 'b h (x y) (w1 w2) d -> b (h d) (x w1) (y w2)', x = height // wsz, y = width // wsz, w1 = wsz, w2 = wsz)
+        return self.to_out(fmap)
+
 class PoolFormerBlock(nn.Module):
     """
     Implementation of one PoolFormer block.
@@ -893,8 +1091,10 @@ class PoolFormerBlock(nn.Module):
         #self.token_mixer = Pooling(pool_size=pool_size)
         # self.token_mixer = FNetBlock()
         self.window_size = 4
+        self.attn_heads = 4
         self.attn_mask = None
-        self.token_mixer = WindowAttention(dim=dim, window_size=to_2tuple(self.window_size), num_heads=4)
+        # self.token_mixer = WindowAttention(dim=dim, window_size=to_2tuple(self.window_size), num_heads=4)
+        self.token_mixer = DSSA(dim, heads=self.attn_heads, window_size=self.window_size, dropout=0.5)
         self.norm2 = norm_layer(dim)
         mlp_hidden_dim = int(dim * mlp_ratio)
         self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, 
@@ -912,11 +1112,12 @@ class PoolFormerBlock(nn.Module):
 
     def forward(self, x):
         B, C, H, W = x.shape
-        x_windows = window_partition(x, self.window_size)
-        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)
-        attn_windows = self.token_mixer(x_windows, mask=self.attn_mask)
-        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
-        x_attn = window_reverse(attn_windows, self.window_size, H, W)
+        # x_windows = window_partition(x, self.window_size)
+        # x_windows = x_windows.view(-1, self.window_size * self.window_size, C)
+        # attn_windows = self.token_mixer(x_windows, mask=self.attn_mask)
+        # attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
+        # x_attn = window_reverse(attn_windows, self.window_size, H, W)
+        x_attn = self.token_mixer(x)
         if self.use_layer_scale:
             x = x + self.drop_path(
                 self.layer_scale_1.unsqueeze(-1).unsqueeze(-1)