X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;ds=inline;f=mygpt.py;h=f9547972e6a937c02ef299aabf639570b85b111f;hb=3ae0c8f3767e4285ab548e4548576a6ddf6003bb;hp=d6879dc08a29f05cac1998bc1ab16e46db07821c;hpb=52c6bd98650c846459f10e8303dd2e6c7ba2a68f;p=mygpt.git

diff --git a/mygpt.py b/mygpt.py
index d6879dc..f954797 100755
--- 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)
@@ -24,14 +24,12 @@ class Residual(nn.Module):
 
 ##############################
 
-class PositionalEncoding(nn.Module):
+class AddPositionalEncoding(nn.Module):
     def __init__(self, len_max):
         super().__init__()
         self.len_max = len_max
 
-    # From Vaswani et al 2018
-    # PE_{t,2i}   = sin(t/(L^{2i/D}))
-    # PE_{t,2i+1} = cos(t/(L^{2i/D}))
+    # [Vaswani et al 2018] PE_{t,2i} = sin(t/(L^{2i/D})), PE_{t,2i+1} = cos(t/(L^{2i/D}))
     def forward(self, x):
         t = torch.arange(x.size(1), dtype = x.dtype, device = x.device)[:, None]
         j = torch.arange(x.size(2), dtype = x.dtype, device = x.device)[None, :]
@@ -68,9 +66,9 @@ class QKVAttention(nn.Module):
         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'))
+            forbidden_attention = 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(forbidden_attention, float('-inf'))
 
         a = a.softmax(dim = 3)
         a = F.dropout(a, self.attention_dropout, self.training)
@@ -96,14 +94,14 @@ class MyGPT(nn.Module):
         self.embedding = nn.Sequential(
             nn.Embedding(vocabulary_size, dim_model),
             nn.Dropout(dropout),
-            PositionalEncoding(len_max),
+            AddPositionalEncoding(len_max),
         )
 
         trunk_blocks = [ ]
 
         for _ in range(nb_blocks):
             trunk_blocks += [
-                Residual(
+                WithResidual(
                     nn.LayerNorm((dim_model,)),
                     QKVAttention(
                         dim_in = dim_model,
@@ -113,7 +111,7 @@ class MyGPT(nn.Module):
                         causal = True, attention_dropout = dropout
                     ),
                 ),
-                Residual(
+                WithResidual(
                     nn.LayerNorm((dim_model,)),
                     nn.Linear(in_features = dim_model, out_features = dim_hidden),
                     nn.ReLU(),
@@ -126,12 +124,19 @@ 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, 0))
+        x = F.pad(x, (1, -1))
         x = self.embedding(x)
         x = self.trunk(x)
         x = self.readout(x)
-        x = F.pad(x, (0, 0, 0, -1))
         return x
 
 ######################################################################