cyd3r / pytorch-prototypeDL

Pytorch implementation for "Deep Learning for Case-Based Reasoning through Prototypes"

pytorch-prototypeDL

Pytorch implementation for "Deep Learning for Case-Based Reasoning through Prototypes: A Neural Network that Explains Its Predictions"

Requirements

• Python 3.6.1
• PyTorch 1.1.0 or higher
• Torchvision 0.2.1

Usage

$ python train.py --help
$ usage: train.py [-h] [--mode MODE] [--data_dir DATA_DIR] [--nworkers NWORKERS]
                [--bsize BSIZE] [--shuffle SHUFFLE] [--sigma SIGMA]
                [--alpha ALPHA] [--arch ARCH] [--n_prototypes N_PROTOTYPES]
                [--decoder_arch DECODER_ARCH] [--optim OPTIM] [--lr LR]
                [--wd WD] [--momentum MOMENTUM] [--epochs EPOCHS]
                [--max_norm MAX_NORM] [--start_epoch START_EPOCH]
                [--lambda_class LAMBDA_CLASS] [--lambda_ae LAMBDA_AE]
                [--lambda_1 LAMBDA_1] [--lambda_2 LAMBDA_2]
                [--save_path SAVE_PATH] [--log_dir LOG_DIR]
                [--log_iter LOG_ITER] [--resume RESUME] [--seed SEED]

1. Training MNIST

$ python train.py --mode train_mnist --data_dir data/mnist \
	--arch small --n_prototypes 15 --decoder_arch deconv

MNIST Decoder Samples

MNIST Prototypes

2. Training Fashion-MNIST

$ python train.py --mode train_fmnist --data_dir data/fmnist \
	--arch small --n_prototypes 15 --decoder_arch upconv

Fashion-MNIST Decoder Samples

Fashion-MNIST Prototypes

3. Training CIFAR10

$ python train.py --mode train_cifar --data_dir data/cifar \
	--arch large --n_prototypes 50 --decoder_arch upconv

CIFAR10 Decoder Samples

CIFAR10 Prototypes

Results

Dataset	Network	Decoder Arch	# of Prototypes	Test Accuracy	Notes
MNIST	small	deconv	15	0.9902	---
	small	upconv	15	0.9920	---
Fashion MNIST	small	deconv	15	0.8465	---
	small	upconv	15	0.9039	--lr 6e-4 --lambda_class 4
CIFAR10	small	deconv	50	0.6735	--lr 6e-4
	small	upconv	50	0.6845	--lr 6e-4 --lambda_class 5
	large	upconv	50	0.7338	--lr 6e-4 --lambda_class 10 --lambda_ae 0.1 used L1 instead of MSE for AE error