From aa09a883611d6e323b4e8e678b867097fc13afaf Mon Sep 17 00:00:00 2001 From: =?utf8?q?Fran=C3=A7ois=20Fleuret?= Date: Sun, 7 Jan 2024 16:13:50 +0100 Subject: [PATCH] Update. --- mygpt.py | 6 ++++++ 1 file changed, 6 insertions(+) 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, -- 2.39.5