Update.

[culture.git] / mygpt.py

diff --git a/mygpt.py b/mygpt.py
index 45b7b59..ab4ccbc 100755 (executable)
--- 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)
 
 
 ######################################################################
@@ -116,7 +116,13 @@ class AddPositionalEncoding(nn.Module):
 
 class QKVAttention(nn.Module):
     def __init__(
-        self, dim_in, dim_qk, dim_v, nb_heads=1, causal=False, attention_dropout=0.0
+        self,
+        dim_in,
+        dim_qk,
+        dim_v,
+        nb_heads=1,
+        causal=False,
+        attention_dropout=0.0,
     ):
         super().__init__()
 
@@ -125,6 +131,7 @@ class QKVAttention(nn.Module):
 
         self.causal = causal
         self.attention_dropout = attention_dropout
+        self.record_attention = False
 
         self.w_q = randw(nb_heads, dim_qk, dim_in)
         self.w_k = randw(nb_heads, dim_qk, dim_in)
@@ -176,6 +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(
@@ -253,6 +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)
@@ -264,25 +276,58 @@ class MyGPT(nn.Module):
     # unchanged.
 
     def masked_inplace_autoregression(
-        self, input, ar_mask, forbidden_tokens=None, deterministic_synthesis=False
+        self,
+        input,
+        ar_mask,
+        summed_logits,
+        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()
+                if summed_logits is not None:
+                    summed_logits += logits[torch.arange(t_next.size(0)), 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():
+            if isinstance(m, QKVAttention):
+                m.record_attention = v
+
+    def retrieve_attention(self):
+        a = []
+        for m in self.modules():
+            if isinstance(m, QKVAttention):
+                a.append(m.a)
+        return a
+
 
 ######################################################################
 
@@ -298,13 +343,12 @@ if __name__ == "__main__":
         dim_keys=2,
         dim_hidden=2,
         nb_heads=2,
-        nb_blocks=1,
+        nb_blocks=2,
         dropout=0.1,
         causal=True,
     )
 
     model.eval()
-
     y1 = model(BracketedSequence(x)).x
     y2 = torch.randn_like(y1)
     for s in range(x.size(1)):