X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=README.txt;h=489a79299fa53b4b56d54c37b17092d676e9cf7f;hb=209f48148a9695069de628ba679345e8f95efa30;hp=7ec9bb04916968397ecac3615e5419ac0a58e259;hpb=7d596f5ca37e7e9ac53047acdc7d00928821929f;p=picoclvr.git

diff --git a/README.txt b/README.txt
index 7ec9bb0..489a792 100644
--- a/README.txt
+++ b/README.txt
@@ -1,7 +1,28 @@
-For the stack experiments:
+18.10.2023
+
+./main.py --task=qmlp --model=352M --nb_train_samples=250000 --result_dir=results_qmlp_352M --batch_size=2
+
+~11h per epoch on 3090 Ti
+
+======================================================================
+For the stack experiment:
 
 ./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
+
+# 352M parameters / 2.5M samples, reaches 99.80% after 12 epochs, the
+  learning rate schedule is obviously terrible
+
+./main.py --task=expr --nb_blocks=48 --dim_model=1024 --nb_train_samples=2500000 --result_dir=results_expr_48b_d1024_2.5M
+======================================================================
+25.07.2023
 
-./main.py --task=stack --stack_fraction_values_for_train=0.75 --stack_nb_stacks=3
+./main.py --task=sandbox --nb_train_samples=10000 --nb_test_samples=1000 --nb_blocks=4 --nb_heads=1 --nb_epochs=20