Code to accompany the paper Iterative Amortized Inference by Marino et al., ICML 2018.
First, clone the repository by opening a terminal and running:
$ git clone https://github.com/joelouismarino/iterative_inference.git
The code uses PyTorch version 0.3.0.post4 and visdom version 0.1.7. To avoid conflicts with more recent versions of these packages, you may wish to create a conda environment:
$ conda create --name it_inf python=2.7
To enter the environment, run:
$ source activate it_inf
Within the environment, install PyTorch by visiting the list of versions here, and grabbing version 0.3.0.post4 for your version of CUDA (8.0, 9.0, 9.1, etc.). Note that the code requires CUDA. Be sure to also install torchvision.
To install vidsom, run
(it_inf) $ pip install visdom==0.1.7
You will also need to install dill, a serialization package, and scipy:
(it_inf) $ pip install dill scipy
To exit the environment, run:
(it_inf) $ source deactivate
To use visdom for plotting, open an terminal and run
python -m visdom.server
Note that if you created a conda environment to use visdom version 0.1.7, you will need to enter that environment before activating visdom (see above).
The code can be run from the terminal using command line arguments. The arguments are
dataset,model_type,inference_type,data_path, andlog_path.
For instance, to run a single-level model on MNIST with iterative inference, run:
python main.py --dataset 'mnist' --model_type 'single_level' --inference_type 'iterative' --data_path '/path/to/data/' --log_path '/path/to/logs'
Be sure to replace the paths for data_path and log_path with valid paths to where the data and logs should be saved, respectively.
You can watch the training progress by opening a browser window and navigating to http://localhost:8097, and selecting the visdom environment corresponding to the experiment.