X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=mygpt.py;h=77c29ce909549fca9487e9e50564ce7e01f67932;hb=0b8185b90014369f0d39892e128ad04a7d9ae872;hp=8cd015297b3f29b037eea0c48d4449c2204b8180;hpb=f680fa1486b0a70c37f0951cedd7b5c56b5808bb;p=picoclvr.git

diff --git a/mygpt.py b/mygpt.py
index 8cd0152..77c29ce 100755
--- a/mygpt.py
+++ b/mygpt.py
@@ -45,6 +45,9 @@ class BracketedSequence:
     def slice(self):
         return self.x[:, self.first : self.first + self.nb]
 
+    def complete(self):
+        return self.first == 0 and self.nb == self.x.size(1)
+
 
 ######################################################################
 
@@ -113,16 +116,22 @@ 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__()
 
         def randw(*d):
             return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
 
-        assert causal, "TODO: Switch off the cache when non-causal!!!"
         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)
@@ -132,6 +141,10 @@ class QKVAttention(nn.Module):
     def forward(self, bs_q):
         x_q = bs_q.x
 
+        assert (
+            self.causal or bs_q.complete()
+        ), "Partial evaluation is only possible for causal models"
+
         if bs_q.first == 0:
             self.cache_k = x_q.new_zeros(
                 x_q.size(0), self.w_k.size(0), x_q.size(1), self.w_k.size(1)
@@ -170,6 +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(
@@ -258,7 +275,12 @@ class MyGPT(nn.Module):
     # unchanged.
 
     def masked_inplace_autoregression(
-        self, input, ar_mask, forbidden_tokens=None, deterministic_synthesis=False
+        self,
+        input,
+        ar_mask,
+        deterministic_synthesis=False,
+        forbidden_tokens=None,
+        forced_biases=None,
     ):
         to_generate = (ar_mask.sum(0) > 0).nonzero()
         if to_generate.min() > 0:
@@ -270,6 +292,8 @@ 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:
@@ -277,6 +301,18 @@ class MyGPT(nn.Module):
                 t_next = dist.sample()
             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
+
 
 ######################################################################
 
@@ -292,13 +328,12 @@ 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)):