Releases: the-aerospace-corporation/glaucus
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 modesunet-encoder
andunet-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.