Pretrained PyTorch CRNN chorus detection model (best-validation checkpoint).
Test-set performance: F1 0.871, Precision 0.867, Recall 0.875, Accuracy 0.893, Loss 0.253.
The checkpoint contains the model state dict and its configuration. It is downloaded automatically by the web app on first run, or can be loaded directly:
from pytorch_core.model import load_CRNN_model
model = load_CRNN_model()
The original TensorFlow implementation and its .h5 model remain available on the tensorflow branch.