add yet another transformer stability measure

Author: lucidrains

dalle2_pytorch/dalle2_pytorch.py

@@ -760,6 +760,7 @@ def __init__(
         dim_head = 64,
         heads = 8,
         ff_mult = 4,
+        norm_in = False,
         norm_out = True,
         attn_dropout = 0.,
         ff_dropout = 0.,
@@ -768,6 +769,8 @@ def __init__(
         rotary_emb = True
     ):
         super().__init__()
+        self.init_norm = LayerNorm(dim) if norm_in else nn.Identity() # from latest BLOOM model and Yandex's YaLM
+
         self.rel_pos_bias = RelPosBias(heads = heads)
 
         rotary_emb = RotaryEmbedding(dim = min(32, dim_head)) if rotary_emb else None
@@ -785,6 +788,8 @@ def __init__(
     def forward(self, x):
         n, device = x.shape[1], x.device
 
+        x = self.init_norm(x)
+
         attn_bias = self.rel_pos_bias(n, n + 1, device = device)
 
         for attn, ff in self.layers:

dalle2_pytorch/train_configs.py

@@ -137,6 +137,7 @@ class DiffusionPriorNetworkConfig(BaseModel):
     dim_head: int = 64
     heads: int = 8
     ff_mult: int = 4
+    norm_in: bool = False
     norm_out: bool = True
     attn_dropout: float = 0.
     ff_dropout: float = 0.

dalle2_pytorch/version.py

@@ -1 +1 @@
-__version__ = '0.23.1'
+__version__ = '0.23.2'