use cross attention for conditioning unet based on image embedding tokens (which opens up the door on conditioning on text encodings as well

dalle2_pytorch/dalle2_pytorch.py

@@ -118,14 +118,13 @@ class ChanRMSNorm(RMSNorm):
         inv_norm = torch.rsqrt(squared_sum + self.eps)
         return x * inv_norm * rearrange(self.gamma, 'c -> 1 c 1 1') * self.scale
 
-class PreNormResidual(nn.Module):
-    def __init__(self, dim, fn):
+class Residual(nn.Module):
+    def __init__(self, fn):
         super().__init__()
         self.fn = fn
-        self.norm = RMSNorm(dim)
 
     def forward(self, x, **kwargs):
-        return self.fn(self.norm(x), **kwargs) + x
+        return self.fn(x, **kwargs) + x
 
 # mlp
 
@@ -240,6 +239,7 @@ class Attention(nn.Module):
 
         sim = sim - sim.amax(dim = -1, keepdim = True)
         attn = sim.softmax(dim = -1)
+        attn = self.dropout(attn)
 
         out = einsum('b h i j, b j d -> b h i d', attn, v)
 
@@ -571,6 +571,17 @@ def Upsample(dim):
 def Downsample(dim):
     return nn.Conv2d(dim, dim, 4, 2, 1)
 
+class Blur(nn.Module):
+    def __init__(self):
+        super().__init__()
+        filt = torch.Tensor([1, 2, 1])
+        self.register_buffer('filt', filt)
+
+    def forward(self, x):
+        filt = self.filt
+        filt = rearrange(filt, '... j -> ... 1 j') * rearrange(flit, '... i -> ... i 1')
+        return filter2d(x, filt, normalized = True)
+
 class SinusoidalPosEmb(nn.Module):
     def __init__(self, dim):
         super().__init__()
@@ -598,10 +609,17 @@ class ConvNextBlock(nn.Module):
         super().__init__()
         need_projection = dim != dim_out
 
-        self.mlp = nn.Sequential(
-            nn.GELU(),
-            nn.Linear(cond_dim, dim)
-        ) if exists(cond_dim) else None
+        self.cross_attn = None
+
+        if exists(cond_dim):
+            self.cross_attn = EinopsToAndFrom(
+                'b c h w',
+                'b (h w) c',
+                CrossAttention(
+                    dim = dim,
+                    context_dim = cond_dim
+                )
+            )
 
         self.ds_conv = nn.Conv2d(dim, dim, 7, padding = 3, groups = dim)
 
@@ -618,21 +636,82 @@ class ConvNextBlock(nn.Module):
     def forward(self, x, cond = None):
         h = self.ds_conv(x)
 
-        if exists(self.mlp):
+        if exists(self.cross_attn):
             assert exists(cond)
-            condition = self.mlp(cond)
-            h = h + rearrange(condition, 'b c -> b c 1 1')
+            h = self.cross_attn(h, context = cond) + h
 
         h = self.net(h)
+
         return h + self.res_conv(x)
 
+class CrossAttention(nn.Module):
+    def __init__(
+        self,
+        dim,
+        *,
+        context_dim = None,
+        dim_head = 64,
+        heads = 8,
+        dropout = 0.,
+    ):
+        super().__init__()
+        self.scale = dim_head ** -0.5
+        self.heads = heads
+        inner_dim = dim_head * heads
+
+        context_dim = default(context_dim, dim)
+
+        self.norm = RMSNorm(dim)
+        self.norm_context = RMSNorm(context_dim)
+        self.dropout = nn.Dropout(dropout)
+
+        self.null_kv = nn.Parameter(torch.randn(2, dim_head))
+        self.to_q = nn.Linear(dim, inner_dim, bias = False)
+        self.to_kv = nn.Linear(context_dim, inner_dim * 2, bias = False)
+        self.to_out = nn.Linear(inner_dim, dim, bias = False)
+
+    def forward(self, x, context, mask = None):
+        b, n, device = *x.shape[:2], x.device
+
+        x = self.norm(x)
+        context = self.norm_context(context)
+
+        q, k, v = (self.to_q(x), *self.to_kv(context).chunk(2, dim = -1))
+
+        q, k, v = rearrange_many((q, k, v), 'b n (h d) -> b h n d', h = self.heads)
+
+        # add null key / value for classifier free guidance in prior net
+
+        nk, nv = repeat_many(self.null_kv.unbind(dim = -2), 'd -> b h 1 d', h = self.heads,  b = b)
+
+        k = torch.cat((nk, k), dim = -2)
+        v = torch.cat((nv, v), dim = -2)
+
+        q = q * self.scale
+
+        sim = einsum('b h i d, b h j d -> b h i j', q, k)
+        max_neg_value = -torch.finfo(sim.dtype).max
+
+        if exists(mask):
+            mask = F.pad(mask, (1, 0), value = True)
+            mask = rearrange(mask, 'b j -> b 1 1 j')
+            sim = sim.masked_fill(~mask, max_neg_value)
+
+        sim = sim - sim.amax(dim = -1, keepdim = True)
+        attn = sim.softmax(dim = -1)
+
+        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 Unet(nn.Module):
     def __init__(
         self,
         dim,
         *,
         image_embed_dim,
-        time_dim = None,
+        cond_dim = None,
+        num_image_tokens = 4,
         out_dim = None,
         dim_mults=(1, 2, 4, 8),
         channels = 3,
@@ -643,18 +722,28 @@ class Unet(nn.Module):
         dims = [channels, *map(lambda m: dim * m, dim_mults)]
         in_out = list(zip(dims[:-1], dims[1:]))
 
-        time_dim = default(time_dim, dim)
+        # time and image embeddings
+
+        cond_dim = default(cond_dim, dim)
 
         self.time_mlp = nn.Sequential(
             SinusoidalPosEmb(dim),
             nn.Linear(dim, dim * 4),
             nn.GELU(),
-            nn.Linear(dim * 4, dim)
+            nn.Linear(dim * 4, cond_dim),
+            Rearrange('b d -> b 1 d')
         )
 
-        self.null_image_embed = nn.Parameter(torch.randn(image_embed_dim))
+        self.image_to_cond = nn.Sequential(
+            nn.Linear(image_embed_dim, cond_dim * num_image_tokens),
+            Rearrange('b (n d) -> b n d', n = num_image_tokens)
+        ) if image_embed_dim != cond_dim else nn.Identity()
 
-        cond_dim = time_dim + image_embed_dim
+        # for classifier free guidance
+
+        self.null_image_embed = nn.Parameter(torch.randn(1, num_image_tokens, cond_dim))
+
+        # layers
 
         self.downs = nn.ModuleList([])
         self.ups = nn.ModuleList([])
@@ -664,7 +753,7 @@ class Unet(nn.Module):
             is_last = ind >= (num_resolutions - 1)
 
             self.downs.append(nn.ModuleList([
-                ConvNextBlock(dim_in, dim_out, cond_dim = cond_dim, norm = ind != 0),
+                ConvNextBlock(dim_in, dim_out, norm = ind != 0),
                 ConvNextBlock(dim_out, dim_out, cond_dim = cond_dim),
                 Downsample(dim_out) if not is_last else nn.Identity()
             ]))
@@ -672,7 +761,7 @@ class Unet(nn.Module):
         mid_dim = dims[-1]
 
         self.mid_block1 = ConvNextBlock(mid_dim, mid_dim, cond_dim = cond_dim)
-        self.mid_attn = EinopsToAndFrom('b c h w', 'b (h w) c', PreNormResidual(mid_dim, Attention(mid_dim)))
+        self.mid_attn = EinopsToAndFrom('b c h w', 'b (h w) c', Residual(Attention(mid_dim)))
         self.mid_block2 = ConvNextBlock(mid_dim, mid_dim, cond_dim = cond_dim)
 
         for ind, (dim_in, dim_out) in enumerate(reversed(in_out[1:])):
@@ -714,52 +803,43 @@ class Unet(nn.Module):
         cond_drop_prob = 0.
     ):
         batch_size, device = x.shape[0], x.device
-        t = self.time_mlp(time)
+        time_tokens = self.time_mlp(time)
 
         cond_prob_mask = prob_mask_like((batch_size,), cond_drop_prob, device = device)
 
         # mask out image embedding depending on condition dropout
         # for classifier free guidance
 
-        image_embed = torch.where(
-            rearrange(cond_prob_mask, 'b -> b 1'),
-            image_embed,
-            rearrange(self.null_image_embed, 'd -> 1 d')
+        image_tokens = self.image_to_cond(image_embed)
+
+        image_tokens = torch.where(
+            rearrange(cond_prob_mask, 'b -> b 1 1'),
+            image_tokens,
+            self.null_image_embed
         )
 
-        t = torch.cat((t, image_embed), dim = -1)
+        c = torch.cat((time_tokens, image_tokens), dim = -2) # c for condition
 
         hiddens = []
 
         for convnext, convnext2, downsample in self.downs:
-            x = convnext(x, t)
-            x = convnext2(x, t)
+            x = convnext(x, c)
+            x = convnext2(x, c)
             hiddens.append(x)
             x = downsample(x)
 
-        x = self.mid_block1(x, t)
+        x = self.mid_block1(x, c)
         x = self.mid_attn(x)
-        x = self.mid_block2(x, t)
+        x = self.mid_block2(x, c)
 
         for convnext, convnext2, upsample in self.ups:
             x = torch.cat((x, hiddens.pop()), dim=1)
-            x = convnext(x, t)
-            x = convnext2(x, t)
+            x = convnext(x, c)
+            x = convnext2(x, c)
             x = upsample(x)
 
         return self.final_conv(x)
 
-class Blur(nn.Module):
-    def __init__(self):
-        super().__init__()
-        filt = torch.Tensor([1, 2, 1])
-        self.register_buffer('filt', filt)
-
-    def forward(self, x):
-        filt = self.filt
-        filt = rearrange(filt, '... j -> ... 1 j') * rearrange(flit, '... i -> ... i 1')
-        return filter2d(x, filt, normalized = True)
-
 class Decoder(nn.Module):
     def __init__(
         self,