Added the small weight embedding + id layer norm inits.

diff --git a/mygpt.py b/mygpt.py
index 3bce361..ebc9a83 100755 (executable)
--- a/mygpt.py
+++ b/mygpt.py
@@ -97,10 +97,6 @@ class MyGPT(nn.Module):
             AddPositionalEncoding(len_max),
         )
 
-        # Small embedding initialization
-        with torch.no_grad():
-            self.embedding[0].weight.normal_(0, 2e-2)
-
         trunk_blocks = [ ]
 
         for _ in range(nb_blocks):
@@ -128,6 +124,14 @@ class MyGPT(nn.Module):
 
         self.readout = nn.Linear(in_features = dim_model, out_features = vocabulary_size)
 
+        with torch.no_grad():
+            for m in self.modules():
+                if isinstance(m, nn.Embedding):
+                    m.weight.normal_(mean = 0, std = 2e-2)
+                elif isinstance(m, nn.LayerNorm):
+                    m.bias.zero_()
+                    m.weight.fill_(1.0)
+
     def forward(self, x):
         x = F.pad(x, (1, -1))
         x = self.embedding(x)