projects
/
picoclvr.git
/ blobdiff
commit
grep
author
committer
pickaxe
?
search:
re
summary
|
shortlog
|
log
|
commit
|
commitdiff
|
tree
raw
|
inline
| side by side
Update.
[picoclvr.git]
/
README.txt
diff --git
a/README.txt
b/README.txt
index
b5c4109
..
d4cb93d
100644
(file)
--- a/
README.txt
+++ b/
README.txt
@@
-1,9
+1,17
@@
-For the stack experiments:
+
+For the stack experiment:
./main.py --task=stack
./main.py --task=stack
-./main.py --task=stack --stack_fraction_values_for_train=0.75
+Takes ~1h10min on a 4090.
+
+For the arithmetic expressions experiments
+
+# 38M parameters / 250k samples
+
+./main.py --task=expr
-./main.py --task=stack --stack_fraction_values_for_train=0.75 --stack_nb_stacks=3
+# 352M parameters / 2.5M samples, reaches 99.80% after 12 epochs, the
+ learning rate schedule is obviously terrible
-Each takes ~1h10min on a 4090.
+./main.py --task=expr --nb_blocks=48 --dim_model=1024 --nb_train_samples=2500000 --result_dir=results_expr_48b_d1024_2.5M