Merge pull request #910 from szmigacz/smigacz/mha_xavier_init_gain_fix

Fixed weight init for fused weight matrices in fused MHA by adding correct gain factor

apex/contrib/multihead_attn/encdec_multihead_attn.py

+import math
+
 import torch
 from torch import nn
 from torch.nn import Parameter
@@ -76,7 +78,11 @@ class EncdecMultiheadAttn(nn.Module):
 
     def reset_parameters(self):
         nn.init.xavier_uniform_(self.in_proj_weight_q)
-        nn.init.xavier_uniform_(self.in_proj_weight_kv)
+        # in_proj_weight_kv has shape [2 * hidden, hidden] but it should be
+        # initialized like a [hidden, hidden] matrix.
+        # sqrt(6 / (hidden + hidden)) / sqrt(6 / (2 * hidden + hidden)) = sqrt(1.5)
+        # therefore xavier_uniform gain should be set to sqrt(1.5).
+        nn.init.xavier_uniform_(self.in_proj_weight_kv, gain=math.sqrt(1.5))
         nn.init.xavier_uniform_(self.out_proj_weight)
         if self.bias:
             nn.init.constant_(self.in_proj_bias_q, 0.)

apex/contrib/multihead_attn/self_multihead_attn.py

+import math
+
 import torch
 from torch import nn
 from torch.nn import Parameter
@@ -98,7 +100,11 @@ class SelfMultiheadAttn(nn.Module):
             nn.init.xavier_uniform_(self.k_weight)
             nn.init.xavier_uniform_(self.v_weight)
         else:
-            nn.init.xavier_uniform_(self.in_proj_weight)
+            # in_proj_weight has shape [3 * hidden, hidden] but it should be
+            # initialized like a [hidden, hidden] matrix.
+            # sqrt(6 / (hidden + hidden)) / sqrt(6 / (3 * hidden + hidden)) = sqrt(2)
+            # therefore xavier_uniform gain should be set to sqrt(2).
+            nn.init.xavier_uniform_(self.in_proj_weight, gain=math.sqrt(2))
         nn.init.xavier_uniform_(self.out_proj_weight)
         if self.bias:
             if self.separate_qkv_params: