From: François Fleuret <francois@fleuret.org>
Date: Sun, 7 Jan 2024 15:13:50 +0000 (+0100)
Subject: Update.
X-Git-Url: https://www.fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=commitdiff_plain;h=aa09a883611d6e323b4e8e678b867097fc13afaf;hp=43fbfaac1850098f5b1a9470c8e6ca3d5ab479fe;p=mygptrnn.git

Update.
---

diff --git a/mygpt.py b/mygpt.py
index 24ba345..6e13ff8 100755
--- a/mygpt.py
+++ b/mygpt.py
@@ -656,6 +656,12 @@ class Caterpillar(nn.Module):
         self.rec_K[:, :, t0:t1] = next_K.flatten(2, 3)
 
         if self.training and self.proba_flashback:
+            # 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
+
             # insert_flash_back(
             # self.rec_V,
             # V,