chhwang / cmcl

This code is for the paper "Confident Multiple Choice Learning".

Confident Multiple Choice Learning

This code is for the paper "Confident Multiple Choice Learning".

Preliminaries

It is tested under Ubuntu Linux 16.04.1 and Python 2.7 environment, and requries following Python packages to be installed:

• TensorFlow: version 1.0.0 or above. Only GPU version is available.
• Torchfile: version 0.0.2 or above.

Simple torchfile installation:

pip install torchfile

Dataset

We provide the following datasets in torch format:

• CIFAR-10 whitened: pre-processed data (1.37GB)
• SVHN (excluding the extra dataset): pre-processed data (2.27GB)

Example scripts

• run_CMCL.sh: train the models using "Confident multiple choice learning".
• run_MCL.sh: train the models using "Multiple choice learning".
• run_IE.sh: train the models using "Independent ensemble".

All training options:

python src/ensemble.py \
--dataset=cifar \
--model_type=resnet \
--batch_size=128 \
--num_model=5 \
--loss_type=cmcl_v0 \
--k=4 \
--beta=0.75 \
--feature_sharing=True \
--test=False

• dataset : supports cifar and svhn.
• model_type : supports vggnet, googlenet, and resnet.
• batch_size : we use batch size 128.
• num_model : number of models to ensemble.
• loss_type : supports independent, mcl, cmcl_v0, and cmcl_v1.
• k : overlap parameter.
• beta : penalty parameter.
• feature_sharing : use feature sharing if True.
• test : if True, test the result of previous training, otherwise run a new training.