Yuzhe Lu, Kairong Jiang, Joshua A. Levine, Matthew Berger
This project contains code, based on the above paper, that addresses the compression of volumetric scalar fields using coordinate-based multilayer perceptrons.
We have included a slim requirements.txt file for relevant libraries.
To train a model, please see the train.py script. The only required option is the volume itself, which we expect to be a 3D numpy array. We have included a small volume for testing purposes, located at volumes/test_vol.npy:
python train.py --volume volumes/test_vol.npy
By default, CUDA is disabled, but can be enabled via the flag --cuda. There are a number of other options for training, the two most relevant are: (1) --compression_ratio, the ratio of volume resolution to network size, thus defining the number of network parameters, and (2) --grad_lambda, for specifying a value for gradient regularization.
Once training has complete, two files are written: the network weights, and the network configuration. By default, these are written to files thenet.pth and thenet.json, but can be changed via specifying options for --network and --config.
Furthermore, if you would like to obtain the volume represented by the network once training has completed, then enable the flag --enable-vol-debug. This will write out, both, the input volume, and network-predicted volume, as vti files that you can load in external visualization tools, e.g. ParaView.
To quantize the weights of the network, please see the net_compress.py script. To run:
python net_compress.py --net thenet.pth --config thenet.json --compressed compressed_out
The --compressed option specifies the binary file to write to for the compressed representation.
To decompress the network, please see the net_decompress.py script. To run:
python net_decompress.py --volume volumes/test_vol.npy --compressed compressed_out
The above assumes that we have access to the original volume for comparison purposes, though this is simple enough to modify if the volume is not present.