add: audiocraft spectrogram visualization enhancements

[mratanusarkar]

Description:

Enhancing our spectrogram to better capture the spectral signature of generated audio, enabling easier identification of anomalies and a deeper understanding of the model's auditory output.

Having a good spectrogram tweaked to human auditory perception helps to correctly understand the sound signature in human reference, and in turn, understand the model outputs.

Changes:

1. Applied logarithmic scaling for better sound perception alignment.
2. Dynamic range set to 5th-95th percentiles for improved audio focus.
3. Adjusted frequency range: Lower limit set at 20Hz; upper limit set dynamically based on audio energy concentration.
4. Switched to 'magma' colormap for perceptual uniformity.
5. Improved code structure for maintainability.

Impact:

Provides a clearer, more intuitive view of audio content, benefiting audio enthusiasts and general viewers.

To demonstrate, consider the following examples:

• In this (run), just by looking I can say the music will be beautiful with distinct notes and patterns
• whereas these runs (run, run), are rhythmic, following beat patterns and little intense and loud
• and this (run) turned out to be noisy.

Future Scope:

If @wandb implements issue#6224, it would be cool to replace these spectrogram images with an "WandB Interactive Spectrograms"

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[github-actions]

Thanks for contributing to wandb/examples!
We appreciate your efforts in opening a PR for the examples repository. Our goal is to ensure a smooth and enjoyable experience for you 😎.

Guidelines

The examples repo is regularly tested against the ever-evolving ML stack. To facilitate our work, please adhere to the following guidelines:

• Notebook naming: You can use a combination of snake_case and CamelCase for your notebook name. Avoid using spaces (replace them with _) and special characters (&%$?). For example:

Cool_Keras_integration_example_with_weights_and_biases.ipynb 

is acceptable, but

Cool Keras Example with W&B.ipynb

is not. Avoid spaces and the & character. To refer to W&B, you can use: weights_and_biases or just wandb (it's our library, after all!)

• Managing dependencies within the notebook: You may need to set up dependencies to ensure that your code works. Please avoid the following practices:

  • Docker-related activities. If Docker installation is required, consider adding a full example with the corresponding Dockerfile to the wandb/examples/examples folder (where non-Colab examples reside).
  • Using pip install as the primary method to install packages. When calling pip in a cell, avoid performing other tasks. We automatically filter these types of cells, and executing other actions might break the automatic testing of the notebooks. For example,

  pip install -qU wandb transformers gpt4
  

  is acceptable, but

  pip install -qU wandb
  import wandb

  is not.

  • Installing packages from a GitHub branch. Although it's acceptable 😎 to directly obtain the latest bleeding-edge libraries from GitHub, did you know that you can install them like this:

  !pip install -q git+https://github.com/huggingface/transformers

  You don't need to clone, then cd into the repo and install it in editable mode.

  • Avoid referencing specific Colab directories. Google Colab has a /content directory where everything resides. Avoid explicitly referencing this directory because we test our notebooks with pure Jupyter (without Colab). Instead, use relative paths to make the notebook reproducible.

• The Jupyter notebook file .ipynb is nothing more than a JSON file with primarily two types of cells: markdown and code. There is also a bunch of other metadata specific to Google Colab. We have a set of tools to ensure proper notebook formatting. These tools can be found at wandb/nb_helpers.

Before merging, wait for a maintainer to clean and format the notebooks you're adding. You can tag @tcapelle.

Before marking the PR as ready for review, please run your notebook one more time. Restart the Colab and run all. We will provide you with links to open the Colabs below

The following colabs were changed
-colabs/audiocraft/AudioCraft_MusicGen.ipynb

[soumik12345]

@mratanusarkar
Thanks for the PR.
Can you please change the 's to "?

[mratanusarkar]

@mratanusarkar Thanks for the PR. Can you please change the 's to "?

@soumik12345 let me know if any more changes are required.

Also FYI,
The png image size and quality can be improved from matplotlib side with:

fig, ax = plt.subplots(figsize=(10, 6))  # Adjust the numbers to your preference
...
plt.savefig(output_file, format='png', dpi=300, bbox_inches='tight', pad_inches=0) # add increased dpi

but I didn't include them to save image file size, and the current images look good enough in the wandb tables.
users can change them if needed, for offline download of spectrogram images.

[soumik12345]

The png image size and quality can be improved from matplotlib side with:

@mratanusarkar
Can you please add a config for that?

[mratanusarkar]

The png image size and quality can be improved from matplotlib side with:

@mratanusarkar Can you please add a config for that?

Almost all the variables & parameters (15+) in get_spectrogram() could be added as config.
But I feel that's unnecessary provided by the fact that it's a column data eventually.

The variable names are clear and as per DSP or Audio Engineering terms,
and anyone interested in it can play around with the function.

Adding configs will make it more complex for most of the common use cases. But still, let me know your views.

[soumik12345]

Thanks again for you contribution!