prepare non-causal attention, for use in the unet in the decoder

dalle2_pytorch/dalle2_pytorch.py

@@ -116,12 +116,14 @@ class Attention(nn.Module):
         dim,
         dim_head = 64,
         heads = 8,
-        dropout = 0.
+        dropout = 0.,
+        causal = False
     ):
         super().__init__()
         self.scale = dim_head ** -0.5
         inner_dim = dim_head * heads
 
+        self.causal = causal
         self.norm = RMSNorm(dim)
         self.dropout = nn.Dropout(dropout)
 
@@ -154,6 +156,7 @@ class Attention(nn.Module):
             mask = rearrange(mask, 'b j -> b 1 1 j')
             sim = sim.masked_fill(~mask, max_neg_value)
 
+        if self.causal:
             causal_mask = torch.ones((n, n), dtype = torch.bool, device = device).triu(1)
             sim = sim.masked_fill(causal_mask, max_neg_value)
 
@@ -165,7 +168,7 @@ class Attention(nn.Module):
         out = rearrange(out, 'b h n d -> b n (h d)')
         return self.to_out(out)
 
-class Transformer(nn.Module):
+class CausalTransformer(nn.Module):
     def __init__(
         self,
         *,
@@ -184,7 +187,7 @@ class Transformer(nn.Module):
         self.layers = nn.ModuleList([])
         for _ in range(depth):
             self.layers.append(nn.ModuleList([
-                Attention(dim = dim, dim_head = dim_head, heads = heads, dropout = attn_dropout),
+                Attention(dim = dim, causal = True, dim_head = dim_head, heads = heads, dropout = attn_dropout),
                 FeedForward(dim = dim, mult = ff_mult, dropout = ff_dropout)
             ]))
 
@@ -211,7 +214,7 @@ class DiffusionPriorNetwork(nn.Module):
         super().__init__()
         self.time_embeddings = nn.Embedding(num_timesteps, dim)  # also offer a continuous version of timestep embeddings, with a 2 layer MLP
         self.learned_query = nn.Parameter(torch.randn(dim))
-        self.causal_transformer = Transformer(**kwargs)
+        self.causal_transformer = CausalTransformer(**kwargs)
 
     def forward_with_cond_scale(
         self,