X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;ds=sidebyside;f=mygpt.py;h=70478493f100584588e7389b465303054c62af1f;hb=2186d96fccfc525884f1b3fb722c40642891ab0a;hp=77c29ce909549fca9487e9e50564ce7e01f67932;hpb=621231cc5bb94f983c556a1b450b66067bec4165;p=culture.git

diff --git a/mygpt.py b/mygpt.py
index 77c29ce..7047849 100755
--- a/mygpt.py
+++ b/mygpt.py
@@ -264,6 +264,7 @@ class MyGPT(nn.Module):
                     m.weight.fill_(1.0)
 
     def forward(self, bs):
+        # print(f"GENERATE {bs.first} {bs.first+bs.nb}")
         bs = BracketedSequence(F.pad(bs.x, (1, -1)), bs.first, bs.nb)
         bs = self.embedding(bs)
         bs = self.trunk(bs)
@@ -278,27 +279,40 @@ class MyGPT(nn.Module):
         self,
         input,
         ar_mask,
+        seq_logproba,
+        temperature=1.0,
         deterministic_synthesis=False,
         forbidden_tokens=None,
         forced_biases=None,
     ):
         to_generate = (ar_mask.sum(0) > 0).nonzero()
+
         if to_generate.min() > 0:
             self(
                 BracketedSequence(input, 0, to_generate.min())
             )  # Needed to initialize the model's cache
         for s in range(to_generate.min(), to_generate.max() + 1):
             output = self(BracketedSequence(input, s, 1)).x
+
             logits = output[:, s]
+
+            logits = (logits / temperature).log_softmax(dim=-1)
+
             if forbidden_tokens is not None:
                 logits = logits.masked_fill(forbidden_tokens, float("-inf"))
+
             if forced_biases is not None:
                 logits = logits + forced_biases[None, :]
+
             if deterministic_synthesis:
-                t_next = logits.argmax(1)
+                t_next = logits.argmax(-1)
             else:
                 dist = torch.distributions.categorical.Categorical(logits=logits)
                 t_next = dist.sample()
+
+            all_n = torch.arange(t_next.size(0))
+            seq_logproba += logits[all_n, t_next].sum(dim=-1)
+
             input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
 
     def record_attention(self, v=True):