v0.3.0

Added reduction argument and fixed tensor shape handling

• Added reduction argument to control how the loss/metric is aggregated, supporting 'mean' (default), 'sum', and 'none' options
• Fixed tensor shape inconsistencies throughout the codebase to consistently use [batch, channel, stem, time] format
• Updated docstrings and comments to correctly document expected tensor shapes
• Optimized scaling_factor broadcasting
• Added comprehensive test for all reduction options

Full Changelog: v0.2.9...v0.3.0

v0.2.9

What's Changed

• Added bypass_filter argument that will bypass frequency weighting if True.
• Add RMS_EPS in divisor to avoid NaN by @iver56 in #3

New Contributors

• @iver56 made their first contribution in #3

Full Changelog: v0.2.7...v0.2.9

v0.2.7

Added alias file so imports can either be torch_log_wmse or torch_log_wmse_audio_quality (for legacy use).
Full Changelog: v0.2.3...v0.2.7

v0.2.3

Corrected file reference to pkl filter file.
Full Changelog: v0.2.2...v0.2.3

v0.2.2

Changed GitHub repo name & references

Updated GitHub repo name to appropriate torch-log-wmse. torch-log-wmse-audio-quality can still be installed, but everything will reference the new name (torch-log-wmse) moving forward.

Full Changelog: v0.2.1...v0.2.2

v0.2.1

Package Name Change

• Added pip package torch-log-wmse. Now the package can be installed as either torch-log-wmse or torch-log-wmse-audio-quality.
• Changed name of imports to torch_log_wmse independent of installation name (i.e. pip install torch-log-wmse or pip install torch-log-wmse-audio-quality). Imports now MUST be done as the following:

from torch_log_wmse import LogWMSE

Full Changelog: v0.1.8...v0.2.1

v0.1.8

Changed scaling factor so RMS doesn't need to = 0, rather just be lower than the error threshold to replace with min value. If it only could be 0, then very tiny numbers even closer to 0 would still go through.

Full Changelog: v0.1.7...v0.1.8

v0.1.7

If the unprocessed audio was silent, a value was immediately being returned unrelated to the model predictions. While this is how the original metric was implemented here, the RMS value should actually be replaced with epsilon so that the difference between a non-silent output and silent output can be measured.

Added test for silent input & target.

Adjusted circular shift to account for IR with odd number of samples.

Full Changelog: v0.1.6...v0.1.7

v0.1.6

Full Changelog: v0.1.5...v0.1.6

v0.1.5

Full Changelog: v0.1.4...v0.1.5