2 ======================================================================
3 For the stack experiment:
7 Takes ~1h10min on a 4090.
9 ======================================================================
10 For the arithmetic expressions experiments
12 # 38M parameters / 250k samples
16 # 352M parameters / 2.5M samples, reaches 99.80% after 12 epochs, the
17 learning rate schedule is obviously terrible
19 ./main.py --task=expr --nb_blocks=48 --dim_model=1024 --nb_train_samples=2500000 --result_dir=results_expr_48b_d1024_2.5M
20 ======================================================================
23 ./main.py --task=sandbox --nb_train_samples=10000 --nb_test_samples=1000 --nb_blocks=4 --nb_heads=1 --nb_epochs=20