Awesome

craig

ICML Paper: Data-efficient Training of Machine Learning Models

Training on MNIST:

Change the flags in the code (line 22-23 mnist.py)

Traing on random subsets: subset, random = True, True

Traing on craig subsets: subset, random = True, False

Training ResNet on CIFAR10:

Traing on random subsets: python train_resnet.py -s 0.1 -w -b 512

Traing on craig subsets: python train_resnet.py -s 0.1 -w -b 512 -g --smtk 0

Training Logistic Regression:

Traing on random subsets: python logistic.py --data covtype --method sgd -s 0.1 --greedy 0

Traing on craig subsets: python logistic.py --data covtype --method sgd -s 0.1 --greedy 1

You can use -b, -g to specify the learning rate, otherwise the learning rate will be tuned.

Please note that we used the greedy implementation from summary analythics, and the running times are reported accordingly. To use the provided python implementation, please use the flag smtk=0.