projects
/
mygptrnn.git
/ blobdiff
commit
grep
author
committer
pickaxe
?
search:
re
summary
|
shortlog
|
log
|
commit
|
commitdiff
|
tree
raw
|
inline
| side by side
Update.
[mygptrnn.git]
/
mygpt.py
diff --git
a/mygpt.py
b/mygpt.py
index
eda8685
..
e7362b7
100755
(executable)
--- a/
mygpt.py
+++ b/
mygpt.py
@@
-37,7
+37,7
@@
import ffutils
# 1 for the successive tokens.
#
# Modules able to process brackets may implement a cache that is
# 1 for the successive tokens.
#
# Modules able to process brackets may implement a cache that is
-# resetted when
the input bracket starts at t=0
+# resetted when
init_cache is True
class BracketedSequence:
class BracketedSequence:
@@
-482,7
+482,7
@@
class Caterpillar(nn.Module):
self.attention_dropout = attention_dropout
warnings.warn("flash back", RuntimeWarning)
self.attention_dropout = attention_dropout
warnings.warn("flash back", RuntimeWarning)
- self.proba_flashback =
0.1
+ self.proba_flashback =
1e-2
self.w_G = randw(nb_heads, caterpillar_height, dim_model)
self.b_G = nn.Parameter(
self.w_G = randw(nb_heads, caterpillar_height, dim_model)
self.b_G = nn.Parameter(
@@
-585,8
+585,6
@@
class Caterpillar(nn.Module):
self.rec_K[:, :, t0:t1] = next_K.flatten(2, 3)
if self.training and self.proba_flashback > 0.0:
self.rec_K[:, :, t0:t1] = next_K.flatten(2, 3)
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
# 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
@@
-605,20
+603,18
@@
class Caterpillar(nn.Module):
src_time = t - u - t0
src_head = torch.randint(H, (N, CH, t1 - t0, 1), device=X.device)
src_time = t - u - t0
src_head = torch.randint(H, (N, CH, t1 - t0, 1), device=X.device)
- mask
_V
= (
+ mask = (
torch.rand(N, CH, t1 - t0, DV, device=X.device) <= self.proba_flashback
).long()
torch.rand(N, CH, t1 - t0, DV, device=X.device) <= self.proba_flashback
).long()
+
self.rec_V[:, :, t0:t1] = (
self.rec_V[:, :, t0:t1] = (
- mask
_V
* V[n, src_head, src_time, dv]
- + (1 - mask
_V
) * self.rec_V[:, :, t0:t1]
+ mask * V[n, src_head, src_time, dv]
+ + (1 - mask) * self.rec_V[:, :, t0:t1]
)
)
- mask_K = (
- torch.rand(N, CH, t1 - t0, DK, device=X.device) <= self.proba_flashback
- ).long()
self.rec_K[:, :, t0:t1] = (
self.rec_K[:, :, t0:t1] = (
- mask
_K
* K[n, src_head, src_time, dk]
- + (1 - mask
_K
) * self.rec_K[:, :, t0:t1]
+ mask * K[n, src_head, src_time, dk]
+ + (1 - mask) * self.rec_K[:, :, t0:t1]
)
######################################################################
)
######################################################################
@@
-775,7
+771,12
@@
class MyGPT(nn.Module):
):
super().__init__()
):
super().__init__()
- assert attention_layer in {"mha", "dumbrec", "kvrec", "caterpillar"}
+ assert attention_layer in {
+ "mha",
+ "dumbrec",
+ "kvrec",
+ "caterpillar",
+ }, f"Unknown attention operator {attention_layer}."
if attention_layer == "caterpillar":
assert nb_lines % caterpillar_height == 0
if attention_layer == "caterpillar":
assert nb_lines % caterpillar_height == 0