X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=README.txt;h=489a79299fa53b4b56d54c37b17092d676e9cf7f;hb=209f48148a9695069de628ba679345e8f95efa30;hp=d4cb93d405a81f0d3d3b6118ab06860f2953cb68;hpb=2175a87ad4304a97c63ac9cca6224d0c0b74c64e;p=picoclvr.git

diff --git a/README.txt b/README.txt
index d4cb93d..489a792 100644
--- a/README.txt
+++ b/README.txt
@@ -1,10 +1,17 @@
+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
 
 Takes ~1h10min on a 4090.
 
+======================================================================
 For the arithmetic expressions experiments
 
 # 38M parameters / 250k samples
@@ -15,3 +22,7 @@ For the arithmetic expressions experiments
   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=sandbox --nb_train_samples=10000 --nb_test_samples=1000 --nb_blocks=4 --nb_heads=1 --nb_epochs=20