X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;ds=inline;f=mygpt.py;h=a17849181edafb878c6c55cc45ad0422b01c1dce;hb=8a548630c88957264306db4354e880414b0fa8ef;hp=ac1c55e84d91fecb06b453533f1800aead640ed7;hpb=00b2d5ed01fb523fbc4e699f0419329efbee0ea8;p=culture.git

diff --git a/mygpt.py b/mygpt.py
index ac1c55e..a178491 100755
--- a/mygpt.py
+++ b/mygpt.py
@@ -46,7 +46,7 @@ class BracketedSequence:
         return self.x[:, self.first : self.first + self.nb]
 
     def complete(self):
-        return self.first == 0 and self.nb == x.size(1)
+        return self.first == 0 and self.nb == self.x.size(1)
 
 
 ######################################################################
@@ -169,9 +169,6 @@ class QKVAttention(nn.Module):
             "nhtd,nhsd->nhts", q, self.cache_k[:, :, : bs_q.first + bs_q.nb]
         ) / math.sqrt(self.w_q.size(1))
 
-        if self.record_attention:
-            self.a = a
-
         if self.causal:
             if bs_q.first == 0:
                 self.cache_attzero = (
@@ -186,6 +183,10 @@ class QKVAttention(nn.Module):
             )
 
         a = a.softmax(dim=3)
+
+        if self.record_attention:
+            self.a = a
+
         a = F.dropout(a, self.attention_dropout, self.training)
 
         y = torch.einsum(
@@ -263,6 +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)
@@ -274,8 +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(
@@ -283,16 +292,23 @@ class MyGPT(nn.Module):
             )  # 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 = output[:, s] / temperature
             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):