Update.

[mygpt.git] / mygpt.py

diff --git a/mygpt.py b/mygpt.py
index 7c4e06d..9da2e68 100755 (executable)
--- a/mygpt.py
+++ b/mygpt.py
@@ -14,7 +14,7 @@ from torch.nn import functional as F
 
 ##############################
 
-class Residual(nn.Module):
+class WithResidual(nn.Module):
     def __init__(self, *f):
         super().__init__()
         self.f = f[0] if len(f) == 1 else nn.Sequential(*f)
@@ -37,16 +37,14 @@ class PositionalEncoding(nn.Module):
         j = torch.arange(x.size(2), dtype = x.dtype, device = x.device)[None, :]
         k = j%2
         pe = torch.sin(t / (self.len_max ** ((j - k) / x.size(2))) + math.pi/2 * k)
-        return x + pe # Let broadcasting to its job
+        return x + pe
 
 ##############################
 
 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):
@@ -88,7 +86,8 @@ class MyGPT(nn.Module):
     def __init__(self,
                  vocabulary_size,
                  dim_model, dim_keys, dim_hidden,
-                 nb_heads, nb_blocks, dropout = 0.):
+                 nb_heads, nb_blocks,
+                 dropout = 0.0, len_max = 1e5):
 
         super().__init__()
 
@@ -97,15 +96,15 @@ class MyGPT(nn.Module):
         self.embedding = nn.Sequential(
             nn.Embedding(vocabulary_size, dim_model),
             nn.Dropout(dropout),
-            PositionalEncoding(len_max = 1e5),
+            PositionalEncoding(len_max),
         )
 
         trunk_blocks = [ ]
 
         for _ in range(nb_blocks):
             trunk_blocks += [
-                Residual(
-                    nn.LayerNorm(dim_model),
+                WithResidual(
+                    nn.LayerNorm((dim_model,)),
                     QKVAttention(
                         dim_in = dim_model,
                         dim_qk = dim_keys,
@@ -114,8 +113,8 @@ class MyGPT(nn.Module):
                         causal = True, attention_dropout = dropout
                     ),
                 ),
-                Residual(
-                    nn.LayerNorm(dim_model),
+                WithResidual(
+                    nn.LayerNorm((dim_model,)),
                     nn.Linear(in_features = dim_model, out_features = dim_hidden),
                     nn.ReLU(),
                     nn.Linear(in_features = dim_hidden, out_features = dim_model),
@@ -132,7 +131,8 @@ class MyGPT(nn.Module):
         x = self.embedding(x)
         x = self.trunk(x)
         x = self.readout(x)
-        return x[:, :-1]
+        x = F.pad(x, (0, 0, 0, -1))
+        return x
 
 ######################################################################