Authors: Thomas Leimkühler, Gurprit Singh, Karol Myszkowski, Hans-Peter Seidel, Tobias Ritschel.

The paper and supplemental material can be found on our project webpage.

@article{leimkuehler2019sigasia,
      title={Deep Point Correlation Design},
      author={Thomas Leimk\"uhler and Gurprit Singh and Karol Myszkowski and Hans-Peter Seidel and Tobias Ritschel},
      journal={ACM Transactions on Graphics (Proceedings of SIGGRAPH Asia 2019)},
      volume={38},
      number={6}
      year={2019},
      doi={10.1145/3355089.3356562}
    }

To check out the code on a local machine run

git clone https://github.com/sinbag/deepsampling.git

This code was developed using Tensorflow v1.9 with Python v3.6 and relies on the following packages:

matplotlib, imageio, openEXR

To compile the custom tensorflow filtering op, you can use cmake by adapting the paths INCLUDE_DIRECTORIES() and SET(tensorflowLib "") in

/ops/cmake/CMakeLists.txt

Execute the following commands on the terminal to compile and build the cmake file:

cd deepsampling/ops/
cmake/createBuild.sh
cmake --build ./build

After compilation, adapt CUSTOM_OP_PATH in

/ops/src/filter/dist_filter_gpu.py

to point to the libCustomOp.so file you created.

We are ready to run our first experiment. You can start with the example files provided in the src/experiments/folder. Go to the folder containing the main.py file and execute the following command to start training for Fig.8a from the paper.

python main.py -e experiment_fig8a

Note that the .py extension of the experiment file is not provided. To run any other experiment file, simply run

python main.py -e <your_experiment_filename>

where <your_experiment_filename> refers to the name of the file you can create following the instructions below.

To define a network and a design specification, you need to set up a Python experiment file and put it into

src/experiments/

You will find pre-defined experiment files in this directory, so you can just start from these. The pre-defined experiments correspond to figures in the paper and are meant for easy reproduction of results.

First, you need to define three global variables: LOGS_DIR is the output directory, where network outputs (spectra, realizations, ...) and trained weights are written to. TARGET_DIR points to the directory where target spectra and realizations are stored. FILE_EXT is the preferred visualization output format (usually ".pdf" or ".png").

Next, you need to implement the buildEnvironment() function.

Specify a TrainingSetup with the following parameters:

If you want to use the code in deployment mode, set trainIterations = 0 and load your network weights by specifying weightDir.

Now you need to specify any targets that you want to use. You can define any number of spectra and differential histograms by concatenating them into lists. Keep two lists: One for spectra, one for histograms.

To set up a spectrum target, create a FourierSetup with the following parameters:

Now you can load or estimate any spectrum using the methods provided in the FourierSetup class in src/setup.py, e.g., load a 2D spectrum from an EXR file using loadTarget2D().

To set up a histogram target, create a HistogramSetup with the following parameters:

Now you can estimate any differential histogram using the methods provided in the HistogramSetup class in src/setup.py, e.g., estimate a histogram from realizations using createTargetFromData().

Once you have created a training setup and a list of Fourier and/or differential histogram setups, wrap them into an environment and return it:

return setup.Environment(trainingSetup, fourierSetupList, histogramSetupList)

If you don't use a histogram or a Fourier setup, just put None.

Lastly, you need to specify the loss function by implementing lossSetup(env, outputNode). Through the environment env you have access to all information and targets you created earlier. The outputNode simply is the output of the last filtering operation of the network, i.e., the resulting point set(s).

You need to return the loss value and, optionally, (lists of) spectra and histograms you want to visualize. A simple example of a loss implementation which L1-compares the Monte Carlo estimate of the radially averaged spectrum of the network's output to a reference looks like this:

spectrumNode = fourier.radialSpectrumMC(outputNode, env.fourierSetupList[0])
lossNode = l1Loss(spectrumNode, env.fourierSetupList[0].target)
histogramNode =  None  
return lossNode, spectrumNode, histogramNode