Open almostimplemented opened 1 year ago
Hi, I have met a similar problem and read your comments, but still didn't solve the problem,can you help me? In my code I do the following:
from piano_transcription_inference import PianoTranscription, sample_rate, load_audio
# Load audio
(audio, _) = load_audio('一生等你钢琴.mp3', sr=sample_rate, mono=True)
# Transcriptor
transcriptor = PianoTranscription(device='cpu') # 'cuda' | 'cpu'
# Transcribe and write out to MIDI file
transcribed_dict = transcriptor.transcribe(audio, 'cut_liszt.mid')
And I meet:
I wonder what's wrong with the checkpoint. TAT
Hi @qiuqiangkong ,
Thanks for this tool.
I want to raise a bug and offer a solution.
In my code I do the following:
But then the code downloads the published model anyways and overwrites it:
It's because this logic here is quite brittle:
At the very least, the code should only execute this if the argument
checkpoint_path
wasNone
to begin with.It's good that I have copies of the checkpoint, but it would be a shame for a user to point this script at their
checkpoint
only to have your code overwrite the weights!