X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=mygpt.py;h=b5ac916f1939d0fda01f70931dabb2cb7d0abe60;hb=3c6931f8ddc8160550e026d9e9610ef71260ce10;hp=24ba34591bb6833e9f0a550a4e104074486cec4a;hpb=43fbfaac1850098f5b1a9470c8e6ca3d5ab479fe;p=mygptrnn.git

diff --git a/mygpt.py b/mygpt.py
index 24ba345..b5ac916 100755
--- a/mygpt.py
+++ b/mygpt.py
@@ -655,17 +655,14 @@ class Caterpillar(nn.Module):
         self.rec_V[:, :, t0:t1] = next_V.flatten(2, 3)
         self.rec_K[:, :, t0:t1] = next_K.flatten(2, 3)
 
-        if self.training and self.proba_flashback:
-            # insert_flash_back(
-            # self.rec_V,
-            # V,
-            # self.rec_K,
-            # K,
-            # t0,
-            # t1,
-            # CL,
-            # proba=self.proba_flashback / CL,
-            # )
+        if self.training and self.proba_flashback > 0.0:
+            # insert_flash_back(self.rec_V,V,self.rec_K,K,t0,t1,CL,proba=self.proba_flashback / CL,)
+
+            # This piece of code makes the assumption that there is
+            # nothing informative before t0, otherwise we'd have to
+            # implement a cache for V and K too. This should not be
+            # too much of a problem since this is used only during
+            # train, where full sequence are available
 
             n = torch.arange(N, device=X.device)[:, None, None, None]
             t = torch.arange(t0, t1, device=X.device)[None, None, :, None]