X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=mygpt.py;h=3bce361424ec4a8b37150c300dc0df05b01de7ef;hb=c3621f9a75cd4d79410d90a29dc9fdec401eaa2d;hp=7c4e06df1259aa4bd372987589ababf5f5afd6b1;hpb=cd1cc80f711ca1f7188cc9854f18231e02470eba;p=mygpt.git

diff --git a/mygpt.py b/mygpt.py
index 7c4e06d..3bce361 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)
@@ -24,29 +24,25 @@ 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, :]
         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 +84,8 @@ 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 +94,19 @@ class MyGPT(nn.Module):
         self.embedding = nn.Sequential(
             nn.Embedding(vocabulary_size, dim_model),
             nn.Dropout(dropout),
-            PositionalEncoding(len_max = 1e5),
+            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):
             trunk_blocks += [
-                Residual(
-                    nn.LayerNorm(dim_model),
+                WithResidual(
+                    nn.LayerNorm((dim_model,)),
                     QKVAttention(
                         dim_in = dim_model,
                         dim_qk = dim_keys,
@@ -114,8 +115,8 @@ 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),
@@ -128,11 +129,11 @@ class MyGPT(nn.Module):
         self.readout = nn.Linear(in_features = dim_model, out_features = vocabulary_size)
 
     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)
-        return x[:, :-1]
+        return x
 
 ######################################################################