X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=mygpt.py;h=809f79032a55204abd8add007b025ca54b1ad227;hb=60d829ba77c9769009d3d5a93a50d23c532d019a;hp=c58bea13414fccbffd7de432c142f06ca7ac83eb;hpb=702e672dcf9ebcfad11ae4034e64117f2c67ead5;p=culture.git

diff --git a/mygpt.py b/mygpt.py
index c58bea1..809f790 100755
--- a/mygpt.py
+++ b/mygpt.py
@@ -279,22 +279,24 @@ class MyGPT(nn.Module):
         self,
         input,
         ar_mask,
+        seq_logproba,
         temperature=1.0,
         deterministic_synthesis=False,
         forbidden_tokens=None,
         forced_biases=None,
     ):
-        sum_logits = 0
         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.log_softmax(dim=-1) / temperature
+            logits = (logits / temperature).log_softmax(dim=-1)
 
             if forbidden_tokens is not None:
                 logits = logits.masked_fill(forbidden_tokens, float("-inf"))
@@ -303,16 +305,16 @@ class MyGPT(nn.Module):
                 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()
-                sum_logits += logits.log_softmax(dim=-1)[
-                    torch.arange(t_next.size(0)), t_next
-                ].sum()
-            input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
 
-        return sum_logits
+            if seq_logproba is not None:
+                all_t = torch.arange(t_next.size(0))
+                seq_logproba += logits[all_t, t_next].sum(dim=-1)
+
+            input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
 
     def record_attention(self, v=True):
         for m in self.modules():