Initial release
Removed device assignment that was causing errors on distributed training setups
Error in tensor shapes. Was processing as the incorrect [batch, channels, stem, samples] instead of the correct [batch, stem, channels, samples] in some places.
Moved frequency weighting image to be included in the package.
Moving image again and correcting reference.
Added repo logo.
The convolution operation was previously introducing an unintended time shift due to incorrect padding and trimming. This was causing models to inadvertently learn these time shifts when the operation was used as a loss function. This issue has now been corrected. The convolution operation is now time-invariant, meaning it will not introduce any unwanted time shifts.
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.
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.
Publishing as torch-log-wmse (for brevity) as well as torch-log-wmse-audio-quality.
Updated all references 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
Updated badge references in the README.
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.
Corrected file reference to pkl filter file.
added alias file so imports can either be torch_log_wmse or torch_log_wmse_audio_quality.
Added bypass_filter argument that will bypass frequency weighting if True.
Fixed bug that returns NaN when one of the entries in the batch is a digital silence triplet - Thanks to Iver Jordal for the issue & PR!
Updated README to reflect 0.2.8 bypass_filter update.