I've just released a new version of the Python package used 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-modelerFull release notes are on GitHub.
What's new?
This one's got a ton of stuff that I'm excited to roll out.
Extensions
This is the one that I'm most excited for. I've created a way for others to easily build customizations for the local trainers. (If all goes well, then I may look at doing it for the online trainer as well, though the use of Jupyter notebooks within Colab presents some limitations.) Here are a few of the sorts of things that will be possible in the GUI trainer without needing to modify the source code of my repository:
Add new metadata fields to the trainer (Reamp input levels, Knob settings, etc).
Allow for using different architectures (e.g. LSTM instead of WaveNet) or custom hyperparameters within the standardized trainer (bigger-than-standard models?)
Oversampling the training files to get models that run at higher sample rates than 48k.
Using different training audio signals than "v3_0_0.wav".
Of course, the full-featured trainer can already do all of these, but users who prefer its flexibility can also improve their workflows by e.g. automatically registering new architectures automatically with an extension.
To write an extension, developers simply write a Python module and place it under ~/.neural-amp-modeler/extensions/. The GUI trainer will detect any Python modules there and attempt to run them.
This is somewhat similar to how "plug-ins" work in digital audio workstations--the DAW takes care of some core functionality, and users can customize their experience by installing their favorite plug-ins (like the NAM plug-in!) My hope is that this will make me less of a bottleneck for how people would like to experience the trainer and help others more easily share their innovations on top of NAM.
Because it's possible to add in functionality now like this, I'll be making some breaking changes in the future to remove support for some older functionality (e.g. older test files than v3_0_0.wav) since folks will be able to add these back as extensions if they'd like.
If you're interested to make your own extensions, you can either check out the Issue or see this repo for demonstrations.
[GUI Trainer] Upfront validation of all files before training
One thing that's been a pain point for folks doing batch training is when a file in the middle of a batch fails validation. This left users with two options: come back and find out that training stopped with a failure, or ignore the checks and not know if there is a model lurking in the middle of a batch of models that has something wrong with it.
So that folks can feel confident that training will run smoothly and produce good results, I've moved the data validation up to the front of the batch run so that you can immediately see whether there's anything you'll want to look into, and trust that batch training will go smoothly once you send it on its way.
Use the full-featured trainer without cloning the repo
The simplified GUI trainer has been available via the command "nam" if you install the repo using PyPI. However, if you wanted to use the full-featured trainer, you had to clone the repo from GitHub, which is more complicated. With this release, users can call
nam-fullfrom the command line to get the full functionality that was previously in the script at bin/train/main.py. (which is now removed; its code lives in the repo at nam/train/full.py.
More training metadata for the standardized trainers
I've added a section to the metadata that captures information from the standardized trainers. This includes things like the measured latency between the input and output signals, replicate ESR, whether any checks failed (and whether they were ignored by the user), etc. The hope is that this will make it easier to troubleshoot if a model comes out sounding off in some way or another.
[GUI trainer] Improvements to visuals on macOS when using dark mode
Just a few cosmetic touch-ups for example, the text isn't black on a near-black background :)
Enjoy!