v2.0.0 Better Variable Compression Autoencoder

The newest version of the autoencoder can encode arbitrary length continuous data using a combination of LSTM and residual vector quantization. Signal features are sequentially encoded and subtracted from the original, allowing a variable amount of compression from 51.2x to 819.2x.

When encoding complex-valued signals of shape (batch_size, length), the pre-trained model will produce tokens of shape (batch_size, length // 256, 16) where 16 represents 16 codebooks of decreasing importance. The model can reconstruct the input with any number of codebooks.

For detailed usage see README.md. Pre-trained weights are available as an attachment to this release.

Aerospace Open-Source-Software release OSS24-0004-Glaucus.

Full Changelog: v1.2.0...v2.0.0

v1.2.0 Variational Autoencoder

Check out the README.md for an example of how to instantiate the pre-trained variational autoencoder (VAE) using the attached weights.

• RFLoss now has optional overlap parameter for spectrogram calculation
• Glaucus can now be instantated as an RF Unet by using blockgen with new modes unet-encoder and unet-decoder.
• Add GitHub workflow
• Warnings eliminated
• Eliminate noise layer from basic AE; could be re-enabled for denoising autoencoder

Unlimited Public Release per Decipher # OSS22-0005 approved 2022-08-30.

v1.1.4 Easily Adapt Weights

Added some code to help folks just learning how to use the autoencoder.

• add function to adapt quantized weights to non-quantized model
• move all configuration into pyproject.toml

Full Changelog: v1.1.3...v1.1.4

Unlimited Public Release per Decipher # OSS22-0005 approved 2022-08-30.

v1.1.3 Sig53 Transfer Learning Model

Model compatible with Sig53 dataset coincident with release containing model weights created using transfer learning from prior best model (glaucus-1024-761-c49063fd.pth).

Unlimited Public Release per Decipher # OSS22-0005 approved 2022-08-30.

v1.1.0 Initial Release with IEEE Papers

Glaucus release coinciding with IEEE Aerospace Conference 2023.

• code included
• models included

Unlimited Public Release per Decipher # OSS22-0005 approved 2022-08-30.