Code for reproducing experiment results of our paper

Requirements:

1. Python 3.6.5
2. Pytorch 0.4.1
3. torchvision 0.2.1
4. tqdm 4.26.0
5. numpy, scipy

Model Evaluation is Done on Datasets:

Algorithm evaluation in our experiment section is done with dataset MNIST and FashionMNIST.

Classification Task Code is at Classification Repos

Usage Instruction:

• Train model with:

$ python train.py <hparams> 

If hparams.Mixture.naive = True in configuration file , then GenMM is going to be trained. If hparams.naive is setting False, LatMM is going to be trained.

• Training example:

python train.py hparams/LatMM/fashion-mnist5.json fashion-mnist directory_to_dataset

For generating samples and latent space interpolation, the directory path to a trained model (either GenMM or LatMM) should be specificed for Infer.pre_trained in configuration file . To generate samples or do latent space interpolation, run:

python infer_mix.py <hparams> <dataset_root> <mode>

For option , it can either be Generating or Interpolation.

Model implementation and evaluation:

For generator implementation in our GenMM and LatMM, the gnerators are implemented as flow-based models. The specific glow structure among flow models is used. The glow structure in glow-pytorch as pytorch implemention of paper "Glow: Generative Flow with Invertible 1×1 Convolutions" is adopted.

For implementation of our model/algorithm evaluation regarding different metrics, we use the generative model evaluation framework