-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathtrain.sh
46 lines (36 loc) · 3.3 KB
/
train.sh
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
# Experiment names in their respective table are included as comments:
## FB augmented training runs:
python train_with_gradient_descent.py name=baseline_sgd hyp=base_sgd # Baseline SGD
python train_with_gradient_descent.py name=fbaug_1 hyp=fb1 # Baseline FB
python train_with_gradient_descent.py name=fbaug_2 hyp=fb2 # FB train longer
python train_with_gradient_descent.py name=fbaug_clip hyp=fbclip # FB clipped
python train_with_gradient_descent.py name=fbaug_gradreg_lr08 hyp=gradreg # FB regularized
python train_with_gradient_descent.py name=fbaug_highreg_lr08 hyp=gradreg data.batch_size=32 # FB strong reg.
python train_with_gradient_descent.py name=fbaug_highreg_lr08_shuffle hyp=gradreg data.batch_size=32 hyp.shuffle=True # FB in practice
## FB fixed dataset:
# no augmentations:
python train_with_gradient_descent.py name=noaug_sgd data.augmentations_train= hyp=base_sgd # Baseline SGD
python train_with_gradient_descent.py name=fb_noaug_1 data.augmentations_train= hyp=fb1 # Baseline FB
python train_with_gradient_descent.py name=fb_noaug_2 data.augmentations_train= hyp=fb2 # FB train longer
python train_with_gradient_descent.py name=fb_noaug_clip data.augmentations_train= hyp=fbclip # FB clipped
python train_with_gradient_descent.py name=fb_noaug_gradreg_lr08 data.augmentations_train= hyp=gradreg # FB regularized
python train_with_gradient_descent.py name=fb_noaug_highreg_lr08 data.augmentations_train= hyp=gradreg data.batch_size=32 # FB strong reg.
# # 10x CIFAR:
python train_with_gradient_descent.py name=SGD_10_CIFAR hyp=base_sgd data/db=LMDB data.augmentations_train= data.db.rounds=10 hyp.train_semi_stochastic=True # Baseline SGD
#
# python train_with_gradient_descent.py name=fb_10_1 data/db=LMDB data.augmentations_train= data.db.rounds=10 hyp=fb1 # Baseline FB
# python train_with_gradient_descent.py name=fb_10_2 data/db=LMDB data.augmentations_train= data.db.rounds=10 hyp=fb2 # FB train longer
# python train_with_gradient_descent.py name=fb_10_clip data/db=LMDB data.augmentations_train= data.db.rounds=10 hyp=fbclip # FB clipped
# python train_with_gradient_descent.py name=fb_10_gradreg_lr08 data/db=LMDB data.augmentations_train= data.db.rounds=10 hyp=gradreg # FB regularized
# python train_with_gradient_descent.py name=fb_10_highreg_lr08 data/db=LMDB data.augmentations_train= data.db.rounds=10 hyp=gradreg data.batch_size=32 # FB strong reg.
#
# #40x CIFAR:
python train_with_gradient_descent.py name=SGD_40_CIFAR data/db=LMDB data.augmentations_train= data.db.rounds=40 hyp=base_sgd hyp.train_semi_stochastic=True
#
# python train_with_gradient_descent.py name=fb_40_1 data/db=LMDB data.augmentations_train= data.db.rounds=40 hyp=fb1
# python train_with_gradient_descent.py name=fb_40_2 data/db=LMDB data.augmentations_train= data.db.rounds=40 hyp=fb2
# python train_with_gradient_descent.py name=fb_40_clip data/db=LMDB data.augmentations_train= data.db.rounds=40 hyp=fbclip
# python train_with_gradient_descent.py name=fb_40_gradreg_lr08 data/db=LMDB data.augmentations_train= data.db.rounds=40 hyp=gradreg
# python train_with_gradient_descent.py name=fb_40_highreg_lr08 data/db=LMDB data.augmentations_train= data.db.rounds=40 hyp=gradreg data.batch_size=32
# Use checkpointing or multi-GPUs setups to finish the later settings in a reasonable time.
# Both are implemented and more info can be found in the config folder.