Awesome

AmbientGAN: Generative models from lossy measurements

This repository provides code to reproduce results from the paper AmbientGAN: Generative models from lossy measurements.

The training setup is as in the following diagram:

Here are a few example results:

Measured	Baseline	AmbientGAN (ours)
<img src="https://github.com/AshishBora/ambient-gan/blob/master/images/celebA_drop_patch/measured.png" width="300">	<img src="https://github.com/AshishBora/ambient-gan/blob/master/images/celebA_drop_patch/inpaint_ns.png" width="300">	<img src="https://github.com/AshishBora/ambient-gan/blob/master/images/celebA_drop_patch/ambient.png" width="300">
<img src="https://github.com/AshishBora/ambient-gan/blob/master/images/celebA_blur_addnoise/measured.png" width="300">	<img src="https://github.com/AshishBora/ambient-gan/blob/master/images/celebA_blur_addnoise/wiener_deconv.png" width="300">	<img src="https://github.com/AshishBora/ambient-gan/blob/master/images/celebA_blur_addnoise/ambient.png" width="300">
<img src="https://github.com/AshishBora/ambient-gan/blob/master/images/cifar_drop_independent/measured.png" width="300">	<img src="https://github.com/AshishBora/ambient-gan/blob/master/images/cifar_drop_independent/unmeasure_blur.png" width="300">	<img src="https://github.com/AshishBora/ambient-gan/blob/master/images/cifar_drop_independent/ambient.png" width="300">

Few more samples from AmbientGAN models trained with 1-D projections:

Pad-Rotate-Project	Pad-Rotate-Project-theta
<img src="https://github.com/AshishBora/ambient-gan/blob/master/images/mnist_pad_rotate_project/ambient.png" width="200">	<img src="https://github.com/AshishBora/ambient-gan/blob/master/images/mnist_pad_rotate_project_with_theta/ambient.png" width="200">

The rest of the README describes how to reproduce the results.

Requirements

Python 2.7
Tensorflow >= 1.4.0
matplotlib
scipy
numpy
cvxpy
scikit-learn
tqdm
opencv-python
pandas

For pip installation, use $ pip install -r requirements.txt

Get the data

MNIST data is automatically downloaded
Get the celebA dataset here and put the jpeg files in ./data/celebA/
Get the CIFAR-10 python data from here and put it in ./data/cifar10/cifar-10-batches-py/*

Get inference models

We need inference models for computing the inception score.

For MNIST, you can train your own by
```
cd ./src/mnist/inf
python train.py
```
[TODO]: Provide a pretrained model.
Inception model for use with CIFAR-10 is automatically downloaded.

Create experiment scripts

Run ./create_scripts/create_scripts.sh

This will create scripts for all the experiments in the paper.

[Optional] If you want to run only a subset of experiments you can define the grid in ./create_scripts/DATASET_NAME/grid_*.sh or if you wish to tweak a lot of parameters, you can change ./create_scripts/DATASET_NAME/base_script.sh. Then run ./create_scripts/create_scripts.sh as above to create the corresponding scripts (remember to remove any previous files from ./scripts/)

Run experiments

We provide scripts to train on multiple GPUs in parallel. For example, if you wish to use 4 GPUs, you can run: ./run_scripts/run_sequentially_parallel.sh "0 1 2 3"

This will start 4 GNU screens. Each program within the screen will attempt to acquire and run experiments from ./scripts/, one at a time. Each experiment run will save samples, checkpoints, etc. to ./results/.

See results as you train

Samples

You can see samples for each experiment in ./results/samples/EXPT_DIR/

EXPT_DIR is defined based on the hyperparameters of the experiment. See ./src/commons/dir_def.py to see how this is done.

Quantitative plots

Run

python src/aggregator_mnist.py
python src/aggregator_cifar.py

This will create pickle files in ./results/ with the relevant data in a Pandas dataframe.

Now use the ipython notebooks ./plotting_mnist.ipynb and ./plotting_cifar.ipynb to get the relevant plots. The generated plots are also saved to ./results/plots/ (make sure this directory exists)