OCD update
authorFrancois Fleuret <francois@fleuret.org>
Mon, 25 Jul 2022 19:04:30 +0000 (21:04 +0200)
committerFrancois Fleuret <francois@fleuret.org>
Mon, 25 Jul 2022 19:04:30 +0000 (21:04 +0200)
mygpt.py

index 57cbbc6..37fe6af 100755 (executable)
--- a/mygpt.py
+++ b/mygpt.py
@@ -41,34 +41,43 @@ class PositionalEncoding(nn.Module):
 ##############################
 
 class QKVAttention(nn.Module):
-    def __init__(self, dim_in, dim_qk, dim_v,
-                 nb_heads = 1, causal = False, attention_dropout = 0.0):
+    def __init__(
+            self,
+            dim_in, dim_qk, dim_v,
+            nb_heads = 1, causal = False, attention_dropout = 0.0
+    ):
         super().__init__()
 
         def randw(*d):
-            return nn.Parameter(torch.empty(*d).normal_(0, 1 / math.sqrt(d[-1])))
+            return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
+
+        self.causal = causal
+        self.attention_dropout = attention_dropout
 
         self.w_q = randw(nb_heads, dim_qk, dim_in)
         self.w_k = randw(nb_heads, dim_qk, dim_in)
         self.w_v = randw(nb_heads, dim_v, dim_in)
         self.w_o = randw(dim_in, dim_v * nb_heads)
-        self.causal = causal
-        self.attention_dropout = attention_dropout
 
     def forward(self, x_q, x_kv = None):
         if x_kv is None: x_kv = x_q
+
         q = torch.einsum('ntc,hdc->nhtd', x_q, self.w_q)
         k = torch.einsum('ntc,hdc->nhtd', x_kv, self.w_k)
         v = torch.einsum('ntc,hdc->nhtd', x_kv, self.w_v)
+
         a = torch.einsum('nhtd,nhsd->nhts', q, k) / math.sqrt(q.size(3))
+
         if self.causal:
             mask = torch.arange(a.size(2), device = q.device)[None, None, :, None] \
                    < torch.arange(a.size(3), device = q.device)[None, None, None, :]
             a = a.masked_fill(mask, float('-inf'))
+
         a = a.softmax(dim = 3)
         a = F.dropout(a, self.attention_dropout, self.training)
-        y = torch.einsum('nhts,nhsd->nthd', a, v)
-        y = y.flatten(2) @ self.w_o
+        y = torch.einsum('nhts,nhsd->nthd', a, v).flatten(2)
+
+        y = y @ self.w_o
 
         return y