Neural Volume Compression

Compressing 3D volumetric data with a blocked autoencoder approach. See report.pdf for an explanation.

Basic Usage

from codex import CodecDenseVAE
from utils import load_volume, show_volume

# default is the "more parameters" variant
codec = CodecDenseVAE()

# or load a custom model
codec = CodecDenseVAE(size=8, margin=0,
        model_pth="dense_vae_s8_m0_l32_h128_b64_e20.pt")

# encode and decode
vol = load_volume("data/tacc_turbulence_256x256x256_1x1x1_uint8.raw", size=256)
encoded = codec.encodes(vol)
decoded = codec.decodes(encoded)

# visualize
show_volume(decoded)

The model name s8_m0_l32_h128_b64_e20 refers to 8-wide blocks, 0 margin, 32 latent dims, 128 hidden dims, batch size of 64 and 20 epochs.

Project Structure

• utils.py Utilities for loading, saving, preprocessing, segmenting and visualizing volumes.
• models.py Pytorch neural models, such as DenseVAE.
• codex.py End-to-end codec classes, which wrap the NN models and include all pre/post processing.
• dataset.py Blocked dataset class used for training.
• training.py All utilities for training the NN models.
• eval.py Utilities for evaluating both raw models and end-to-end codecs.