作者大大您好，我在电脑上已经安装了ffmpeg，cmd里输入ffmpeg -version是可以查看到信息的，但运行example.py仍然报错：

Exception has occurred: NoBackendError
  File "C:\Users\XD_26\Downloads\piano_transcription_inference-master\piano_transcription_inference\utilities.py", line 504, in load_audio
    with audioread.audio_open(os.path.realpath(path), backends=backends) as input_file:
  File "C:\Users\XD_26\Downloads\piano_transcription_inference-master\example.py", line 22, in inference
    (audio, _) = load_audio(audio_path, sr=sample_rate, mono=True)
  File "C:\Users\XD_26\Downloads\piano_transcription_inference-master\example.py", line 38, in <module>
    inference('resources/cut_liszt.mp3', 'cut_liszt.mid')

我不太清楚我哪一步出了问题，希望能得到解答。非常感谢！:)

Hello

An error occurred while using the library and left an issue.

The error I'm currently getting is:

audio.exceptions.NoBackendError

As far as I know, this error occurs when the ffmpeg library is not present when loading the mp3 file format.

But the file I'm loading now is a file in wav format.

I think it's a version issue In version 3.10.0 Down to version 3.7.9 but still having this issue.

The environment I am currently using is as follows.

IDE: Pycharm Version : 3.7.9 , 3.10.0

OS: Windows 10

Thank you for publishing the model. In pypi's package, when the audio model can't turn notes, the program will report an error. Instead of outputting an empty midi. The wrong place is in line 397 of piano_transcription_inference/utilities.py. And there are some audio models that can't be identified.

This is added for a quick demonstration of how the code works, which also provides an easy platform for ppl to try it out without installing anything.

Thank you for sharing your amazing work.

In my case, inference time is the first priority. The paper includes two ablations in which the frame head does not condition on outputs of velocity and offset heads (Table I), and I think the first ablation (Regress cond.) might have a faster inference time. If it's so, could you share those checkpoints as well?

Also, what other suggestions do you have for improving inference times? Note that, the model I used is Regress_onset_offset_frame_velocity_CRNN because I do not need to sustain pedal predictions.

Many thanks in advance!

While I was trying to package piano_transcription_inference for Nix, the following error encountered:

  File "/usr/local/lib/python3.9/site-packages/piano_transcription_inference/utilities.py", line 556, in load_audio
    y = librosa.core.audio.resample(y, sr_native, sr, res_type=res_type)
TypeError: resample() takes 1 positional argument but 3 positional arguments (and 1 keyword-only argument) were given

Downgrade librosa to 0.8.1 and it worked fine.

作者你好，我在使用piano_transcription做训练的时候出现了梯度消失的情况。我更改了模型内部的一些参数（比如conv channel、fc等）和配置文件（utils/config.py）中的frames_per_second（从100改为20），在训练过程中出现了梯度消失的情况tensor(nan, grad_fn=)，我要如何解决这个问题呢？

Hi, Thanks for sharing this work.

I am using pretrained model from this repo. and I found that its performance score is much higher than the original paper's score. (Note F1 0.89, Onset F1 0.979)

Is this pretrained model trained by applying the augmentation technique included in the code you provided?

I may have made a mistake in the score calculation, so I would appreciate it if you let me know if audio augmentation is applied on this pretrained model.