switch to using linear attention for the sparse attention layers within unet, given success in GAN projects

dalle2_pytorch/dalle2_pytorch.py

@@ -1050,6 +1050,42 @@ class GridAttention(nn.Module):
         out = rearrange(out, '(b h w) (w1 w2) c -> b c (w1 h) (w2 w)', w1 = wsz, w2 = wsz, h = h // wsz, w = w // wsz)
         return out
 
+class LinearAttention(nn.Module):
+    def __init__(
+        self,
+        dim,
+        dim_head = 32,
+        heads = 8
+    ):
+        super().__init__()
+        self.scale = dim_head ** -0.5
+        self.heads = heads
+        inner_dim = dim_head * heads
+        self.norm = ChanLayerNorm(dim)
+
+        self.nonlin = nn.GELU()
+        self.to_qkv = nn.Conv2d(dim, inner_dim * 3, 1, bias = False)
+        self.to_out = nn.Conv2d(inner_dim, dim, 1, bias = False)
+
+    def forward(self, fmap):
+        h, x, y = self.heads, *fmap.shape[-2:]
+
+        fmap = self.norm(fmap)
+        q, k, v = self.to_qkv(fmap).chunk(3, dim = 1)
+        q, k, v = rearrange_many((q, k, v), 'b (h c) x y -> (b h) (x y) c', h = h)
+
+        q = q.softmax(dim = -1)
+        k = k.softmax(dim = -2)
+
+        q = q * self.scale
+
+        context = einsum('b n d, b n e -> b d e', k, v)
+        out = einsum('b n d, b d e -> b n e', q, context)
+        out = rearrange(out, '(b h) (x y) d -> b (h d) x y', h = h, x = x, y = y)
+
+        out = self.nonlin(out)
+        return self.to_out(out)
+
 class Unet(nn.Module):
     def __init__(
         self,
@@ -1067,7 +1103,6 @@ class Unet(nn.Module):
         attn_heads = 16,
         lowres_cond = False, # for cascading diffusion - https://cascaded-diffusion.github.io/
         sparse_attn = False,
-        sparse_attn_window = 8,  # window size for sparse attention
         attend_at_middle = True, # whether to have a layer of attention at the bottleneck (can turn off for higher resolution in cascading DDPM, before bringing in efficient attention)
         cond_on_text_encodings = False,
         max_text_len = 256,
@@ -1161,7 +1196,7 @@ class Unet(nn.Module):
 
             self.downs.append(nn.ModuleList([
                 ConvNextBlock(dim_in, dim_out, time_cond_dim = time_cond_dim, norm = ind != 0),
-                Residual(GridAttention(dim_out, window_size = sparse_attn_window, **attn_kwargs)) if sparse_attn else nn.Identity(),
+                Residual(LinearAttention(dim_out, **attn_kwargs)) if sparse_attn else nn.Identity(),
                 ConvNextBlock(dim_out, dim_out, cond_dim = layer_cond_dim, time_cond_dim = time_cond_dim),
                 Downsample(dim_out) if not is_last else nn.Identity()
             ]))
@@ -1178,7 +1213,7 @@ class Unet(nn.Module):
 
             self.ups.append(nn.ModuleList([
                 ConvNextBlock(dim_out * 2, dim_in, cond_dim = layer_cond_dim, time_cond_dim = time_cond_dim),
-                Residual(GridAttention(dim_in, window_size = sparse_attn_window, **attn_kwargs)) if sparse_attn else nn.Identity(),
+                Residual(LinearAttention(dim_in, **attn_kwargs)) if sparse_attn else nn.Identity(),
                 ConvNextBlock(dim_in, dim_in, cond_dim = layer_cond_dim, time_cond_dim = time_cond_dim),
                 Upsample(dim_in)
             ]))