This folder contains example codes that employ a PyTorch-based EmbraceNet model on the Fashion MNIST dataset. As in the main paper, the code divides the original images into left and right halves having a size of 14 x 28 pixels and considers them as bimodal datasets.
Place the Fashion MNIST data to the data/ folder and run the code based on the following script snippets.
Train the original EmbraceNet model:
python train.py --data_training --cuda_device=0 --train_path=/tmp/embracenet/without_dropoutTrain the EmbraceNet model with modality dropout (Section 5.1 of the main paper):
python train.py --data_training --cuda_device=0 --model_dropout --train_path=/tmp/embracenet/with_dropoutValidate the EmbraceNet model:
python validate.py --cuda_device=-1 --restore_path=/tmp/embracenet/with_dropout/model_50000.pthValidate the EmbraceNet model only on left halves (i.e., drop right halves):
python validate.py --cuda_device=-1 --model_drop_right --restore_path=/tmp/embracenet/with_dropout/model_50000.pthValidate the EmbraceNet model only on right halves (i.e., drop left halves):
python validate.py --cuda_device=-1 --model_drop_left --restore_path=/tmp/embracenet/with_dropout/model_50000.pthValidate the EmbraceNet model with output self-ensemble (as in this paper):
python validate.py --cuda_device=-1 --restore_path=/tmp/embracenet/with_dropout/model_50000.pth --ensemble_repeats=5