Update.

[culture.git] / mygpt.py

diff --git a/mygpt.py b/mygpt.py
index 0cf70e0..7117e76 100755 (executable)
--- 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)
@@ -275,8 +276,15 @@ class MyGPT(nn.Module):
     # unchanged.
 
     def masked_inplace_autoregression(
-        self, input, ar_mask, forbidden_tokens=None, deterministic_synthesis=False
+        self,
+        input,
+        ar_mask,
+        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(
@@ -287,13 +295,20 @@ class MyGPT(nn.Module):
             logits = output[:, s]
             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)
             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
+
     def record_attention(self, v=True):
         for m in self.modules():
             if isinstance(m, QKVAttention):