Solving "TypeError: __init__() got an unexpected keyword argument 'max_iter'

[J4BEZ]

Thank you for providing us with an interesting project.

While exploring Riffusion through streamlit app,
I discovered an issue and a solution for it

So I came here to share the solution.

TL;DR

pin torchaudio version to 2.0.1
same solution as #158 and #166

I'm sorry to "gu-ma" that I recognize the pull-request of gu-ma after I completed to wrote this article.
If this PR could give you a little help, I would like to request to check PR of gu-ma, not just mine alone for the development of the open-source ecosystem.
Hope you have a merry Christmas🎄

---

What is the Issue🤔

TypeError: __init__() got an unexpected keyword argument 'max_iter'

When does it occur⏰

While running 'Text to Audio' of Riffusion Playground.
The process of converting text into a spectrogram image works very well but there is an issue of the process of extracting audio through converted image

What is the cause of the issue❔

Due to update of torchaudio in pytorch
It is believed to be a structural change in torchaudio.transforms.InverseMelScale.

riffusion-inference\riffusion\spectrogram_converter.py", line 87, in __init__
has keyword arguments named 'max_iter', 'torlerance_loss', 'tolerance_change', 'sgdargs'

        # https://pytorch.org/audio/stable/generated/torchaudio.transforms.InverseMelScale.html
        self.inverse_mel_scaler = torchaudio.transforms.InverseMelScale(
            n_stft=params.n_fft // 2 + 1,
            n_mels=params.num_frequencies,
            sample_rate=params.sample_rate,
            f_min=params.min_frequency,
            f_max=params.max_frequency,
            max_iter=params.max_mel_iters,
            tolerance_loss=1e-5,
            tolerance_change=1e-8,
            sgdargs=None,
            norm=params.mel_scale_norm,
            mel_scale=params.mel_scale_type,
        ).to(self.device)

📑 A good article to refer to
< InverseMelScale - Torchaudio 2.0.1 documentation >
< InverseMelScale - Torchaudio 2.1.1 documentation >

How to solve the issue☑

We can solve this issue by installing torchaudio which v.2.0.x

1. Please remove the existing torchaudio, torchvision and torch.

2. Please revise the requirements.txt as below to install pytorch 2.0.1 Before:

...
torch
torchaudio
torchvision
...

After:

...
torch==2.0.1
torchaudio==2.0.2
torchvision==0.15.2
...

3. if you want to use CUDA, install PyTorch and torchaudio with CUDA support before python -m pip install -r requirements.txt

It is recommended to install pytorch 2.0.1 on account of we pinned the version of torchaudio to 2.0.x .
< PyTorch 2.0.1 Install Guide >

4. run python -m pip install -r requirements.txt

Result:

It works very well

---

Thank you for reading this long issue and make a great project.
I hope you all have a great and peaceful Christmas season. Merry Christmas🎄

[J4BEZ]

now(GMT+9 | December 15th, 2023 13:00), the stable version of torchAudio is 2.1.2 and It also doesn't have
keyword arguments that named 'max_iter', 'torlerance_loss', 'tolerance_change', 'sgdargs'
in torchaudio.transforms.InverseMelScale either.

< InverseMelScale - Torchaudio 2.1.2 documentation >

Therefore It could be a simple solution to pin torchaudio version to 2.0.1
until we refactor the project to be compatible with torchaudio 2.1.2

[CyberT33N]

I just deleted it related to:

• This project is still working but some stuff is deprecated riffusion-app-hobby#43

But would be awesome to set the versions to a fixed number in the requirements.txt file