add ability to turn on normformer settings, given @borisdayma reported good results and some personal anecdata

README.md

@@ -897,4 +897,14 @@ Once built, images will be saved to the same directory the command is invoked
 }
 ```
 
+```bibtex
+@article{Shleifer2021NormFormerIT,
+    title   = {NormFormer: Improved Transformer Pretraining with Extra Normalization},
+    author  = {Sam Shleifer and Jason Weston and Myle Ott},
+    journal = {ArXiv},
+    year    = {2021},
+    volume  = {abs/2110.09456}
+}
+```
+
 *Creating noise from data is easy; creating data from noise is generative modeling.* - Yang Song's <a href="https://arxiv.org/abs/2011.13456">paper</a>

dalle2_pytorch/dalle2_pytorch.py

@@ -499,7 +499,12 @@ class SwiGLU(nn.Module):
         x, gate = x.chunk(2, dim = -1)
         return x * F.silu(gate)
 
-def FeedForward(dim, mult = 4, dropout = 0., post_activation_norm = False):
+def FeedForward(
+    dim,
+    mult = 4,
+    dropout = 0.,
+    post_activation_norm = False
+):
     """ post-activation norm https://arxiv.org/abs/2110.09456 """
 
     inner_dim = int(mult * dim)
@@ -522,7 +527,8 @@ class Attention(nn.Module):
         dim_head = 64,
         heads = 8,
         dropout = 0.,
-        causal = False
+        causal = False,
+        post_norm = False
     ):
         super().__init__()
         self.scale = dim_head ** -0.5
@@ -537,7 +543,11 @@ class Attention(nn.Module):
         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(dim, dim_head * 2, bias = False)
-        self.to_out = nn.Linear(inner_dim, dim, bias = False)
+
+        self.to_out = nn.Sequential(
+            nn.Linear(inner_dim, dim, bias = False),
+            LayerNorm(dim) if post_norm else nn.Identity()
+        )
 
     def forward(self, x, mask = None, attn_bias = None):
         b, n, device = *x.shape[:2], x.device
@@ -602,7 +612,8 @@ class CausalTransformer(nn.Module):
         norm_out = True,
         attn_dropout = 0.,
         ff_dropout = 0.,
-        final_proj = True
+        final_proj = True,
+        normformer = False
     ):
         super().__init__()
         self.rel_pos_bias = RelPosBias(heads = heads)
@@ -610,8 +621,8 @@ class CausalTransformer(nn.Module):
         self.layers = nn.ModuleList([])
         for _ in range(depth):
             self.layers.append(nn.ModuleList([
-                Attention(dim = dim, causal = True, dim_head = dim_head, heads = heads, dropout = attn_dropout),
+                Attention(dim = dim, causal = True, dim_head = dim_head, heads = heads, dropout = attn_dropout, post_norm = normformer),
-                FeedForward(dim = dim, mult = ff_mult, dropout = ff_dropout)
+                FeedForward(dim = dim, mult = ff_mult, dropout = ff_dropout, post_activation_norm = normformer)
             ]))
 
         self.norm = LayerNorm(dim) if norm_out else nn.Identity()  # unclear in paper whether they projected after the classic layer norm for the final denoised image embedding, or just had the transformer output it directly: plan on offering both options

setup.py

@@ -10,7 +10,7 @@ setup(
       'dream = dalle2_pytorch.cli:dream'
     ],
   },
-  version = '0.0.92',
+  version = '0.0.93',
   license='MIT',
   description = 'DALL-E 2',
   author = 'Phil Wang',