Trainer version 0.10.0 is released

I've released a new version of the Python package for training models!

As usual, the newest version is automatically used in the online trainer; folks making models locally on their own computers can install the update by typing

pip install --upgrade neural-amp-modeler

Full release notes are available on GitHub.

What's new?

In addition to a nice bunch of bug fixes, I've got some things I'm really excited to share in this release:

Reamp level calibration metadata

Over the past few years (and likely due to the proliferation of highly accurate amp modeling tech like NAM), there's been growing interest among guitarists in accurately calibrating their digital gear's input gain so that their digital models see the correct signal level.. (A search on YouTube reveals countless videos on the subject of amp sim input levels).

With v0.10.0, I've added metadata fields that can be easily filled using the standardized trainers for the analog input and output calibration levels for the gear that you're modeling. These can be retroactively filled into existing models (as the metadata spec is intentionally open to extension--you can put anything you want in there!), and I will be making changes to NeuralAmpModelerCore to integrate these into the real-time model classes, and NeuralAmpModelerPlugin to demonstrate how they might be used to automatically calibrate levels in plug-ins that builders are developing.

How this all works might be confusing, so I've created documentation that explains for model makers how to correctly determine the send and return level calibration for their recording setups. As always with metadata, it's optional, so you can skip it if you'd like.

Inexact reamp file lengths don't cause errors

In previous versions, I had the trainer check that the input and output files were the same length. Some users struggled with this, so I'm relaxing the requirement so that a reamp that's up to a second longer (cf. the 1-second tail that REAPER adds in its default rendering settings) will be allowed. The file must not be shorter, though--otherwise there is missing validation data. Also, anything more than a second longer will cause an error--something might be wrong.

No more "fit cab" checkbox

The checkbox encourages the model to prioritize matching the high-frequency part of its predictions with the target. This was important to ensure that models of signal chains including cabinets sound good, but it's almost always a welcome tweak to models. Therefore, in order to simplify, I've set it to always be on in the standardized (Colab & GUI) trainers. (If you don't agree with this choice, I developed the ability for the trainer to accept extensions to modify its behavior for exactly this reason.)

Improved automatic latency calibration

The standardized trainers have a convenience function that attempts to determine (and compensate for) the latency in your recording setup. Sometimes, it gets it wrong, resulting in models that have a large latency "baked in". I collected recordings from users that had this issue, and have improved the algorithm so that it gets all of those cases right.

Rename "v3_0_0.wav" to "input.wav"

Why not make things simpler? I've renamed the standardized input file. No other changes there :)

Enjoy!