This repository is used for creating a multimodal LLM-based model used for generating music captions given an audio file (currently only supported by .wav files). Note that the code itself is under the Apache 2.0 license, but any models trained using the YouTube8M-MusicTextClips dataset are under the Research-only, non-commercial Adobe Research License. However, the MusicCaps dataset is under the CC BY-SA 4.0 license and can be used for commercial purposes.
Documentation page is found here.
For developer installation, run the following command in the terminal:
git clone https://github.com/jako4295/j2a.git
pip install -e j2a/To train (and then evaluate) a model you can follow the following steps:
- Extract data from the databases using the data_handler module described in Data Extractor. This is split into two commands, one for each dataset:
python j2a/data_handler/youtube_download.py --data_enum musiccaps --output <path/to/output/folder>python j2a/data_handler/youtube_download.py --data_enum youtube_8m --output <path/to/output/folder>- Encode the data using the encoder module described in Encoder. This is done with the following command:
python j2a/encoder/encode.py --audio_folder <path/to/data_folder> --output_folder <path/to/output/folder>Note
Both the path to the data and the path to the encoded data is stored in the j2a/data_handler/.json_files/data_summary.json. This file is used for parsing all information to the train test split script. Hence, if the data is moved the paths needs to be updated. The procedure to change the paths is the same as when changing the path to a different encoding.
If you implement a different encoder and want to use this in the training process, you need to update the j2a/data_handler/.json_files/data_summary.json with the path to the new encoded data. This can be done as follows:
from j2a.data_handler.json_handler import JsonInteractor
inter = JsonInteractor() # loads the data_summary.json
inter.clean_json_param("encoded_path")
inter.update_json_paths("path/to/new/encoding", "encoded_path")
inter.save(ask=False, silent=True)- Create train and test splits using the
train_test_split.py. This script splits the data into 80% training data, 10% test data, and 10% validation data. The script can be run with the following command:
python j2a/data_handler/train_test_split.py --prompt_path <path/to/prompt_txt_file> --output_dir <path/to/output_dir>This will create a train/test/validation split in a separate folder inside the output_dir.
- A new model can be trained using the
j2a/train.py. This file is run with the following arguments:
python j2a/train.py --train_path <path/to/train.csv> --eval_path <path/to/validation.csv> --save_path <path/to/save_location>Note that the eval_path argument is optional. If it is desired to change more hyperparameters we refer to the j2a/train.py file. In the save_path a model_info will be created with relevant files.
Additional optional arguments are --load_projector_path (path to a pretrained projection path) and --update_llm (boolean - when true then the llm weights and biases are also updated).
- Evaluating the data is currently set to handle J2A-2.0 and J2A-2.1 and the type should be specified since they have a different audio projection layer. The command for evaluating a J2A-2.0 type model is seen below
python j2a/evaluation/j2a_2x.py --projector_path <path/to/model.pth> --test_csv_path <path/to/test.csv> --model_name "j2a-2.0"This will create a pred.csv with responses for all data in the test.csv and will be located in the same folder as the projector_path. If you have a trained llm then the argument --llm_path can be used to point to the folder with the language model.
There are currently two databases used in this project, MusicCaps and YouTube8M-MusicTextClips. In the data_handler folder, there is a method to extract data from both databases.
The data is stored in the specified folder and an additional .json file is created to store metadata about the data. This metadata is stored in j2a/data_handler/.json_files/data_summary.json.
To extract the data create a Python environment and install the following packages:
pip install -r requirements_datagen.txtNote
If you can't import j2a you need to run the pip install -e j2a/
Go to the llm_tolls/data_handler folder. To extract the data run
python <path/to/youtube_download.py> --data_enum <database> --output <path/to/output/folder>where the database specify the database you want to extract data from (options are musiccaps and youtube_8m). For the output folder please use absolute path. An example of this is shown below:
python j2a/data_handler/youtube_download.py --data_enum musiccaps --output <path/to/output_folder>This will create j2a/data_handler/.json_files/data_summary.json and store the data in the specified folder. Note that some files are age-restricted and will not be downloaded.
Once the data is extracted the data_summary.json file can be interacted with through the JsonInteractor class.
Warning
If the sound files are moved the path in j2a/data_handler/.json_files/data_summary.json will be wrong and errors can occur if the paths are not updated in the file. How to update the path is show below.
Base functionalities are:
add_feature: Add a feature to elements in thedata_summary.jsonfile.add_sound_path: Add a sound path to elements in thedata_summary.jsonfile. This is automatically done when extracting data through the youtube_download.py.get_info: Get information about elements in thedata_summary.jsonfile.save: Save the changes to thedata_summary.jsonfile.
To get the names of the elements in the data_summary.json file use the MetaDataEnum class. An example of how to use the MetaDataEnum class and the JsonInteractor class is shown below:
from j2a.data_handler.json_handler import JsonInteractor, MetaDataEnum
database_name = MetaDataEnum.musiccaps.value + "_0"
json_interactor = JsonInteractor()
json_interactor.get_info(database_name)This will print the following information about "MusicCaps_0" from the data_summary.json file:
+------------+-----------------------------------------------------------------+
| Info: | MusicCaps_0 |
+------------+-----------------------------------------------------------------+
| Sound file | /usr/local/path/to/data/MusicCaps_0.mp3 |
+------------+-----------------------------------------------------------------+
| Metadata | +--------------------------+----------------------------------+ |
| | | ytid | -0Gj8-vB1q4 | |
| | +--------------------------+----------------------------------+ |
| | | start_s | 30 | |
| | +--------------------------+----------------------------------+ |
| | | end_s | 40 | |
| | +--------------------------+----------------------------------+ |
| | | audioset_positive_labels | /m/0140xf,/m/02cjck,/m/04rlf | |
| | +--------------------------+----------------------------------+ |
| | | aspect_list | ['low quality', 'sustained strin | |
| | | | gs melody', 'soft female vocal', | |
| | | | 'mellow piano melody', 'sad', ' | |
| | | | soulful', 'ballad'] | |
| | +--------------------------+----------------------------------+ |
| | | caption | The low quality recording featur | |
| | | | es a ballad song that contains s | |
| | | | ustained strings, mellow piano m | |
| | | | elody and soft female vocal sing | |
| | | | ing over it. It sounds sad and s | |
| | | | oulful, like something you would | |
| | | | hear at Sunday services. | |
| | +--------------------------+----------------------------------+ |
| | | author_id | 4 | |
| | +--------------------------+----------------------------------+ |
| | | is_balanced_subset | 0 | |
| | +--------------------------+----------------------------------+ |
| | | is_audioset_eval | 1 | |
| | +--------------------------+----------------------------------+ |
+------------+-----------------------------------------------------------------+
If we move the data to a new folder and also want to add other features to the data_summary.json file we can use the add_feature method. An example of this is shown below:
json_interactor.add_sound_path(
database_name, "/usr/local/new_path/to/data"
)
json_interactor.add_feature([database_name], [{"bpm": 60, "duration": 10}])
json_interactor.get_info(database_name)The get_info method will now print the following information about "MusicCaps_0" from the json_interactor object file:
+------------+-----------------------------------------------------------------+
| Info: | MusicCaps_0 |
+------------+-----------------------------------------------------------------+
| Sound file | /usr/local/new_path/to/data/MusicCaps_0.mp3 |
+------------+-----------------------------------------------------------------+
| Metadata | +--------------------------+----------------------------------+ |
| | | ytid | -0Gj8-vB1q4 | |
| | +--------------------------+----------------------------------+ |
| | | start_s | 30 | |
| | +--------------------------+----------------------------------+ |
| | | end_s | 40 | |
| | +--------------------------+----------------------------------+ |
| | | audioset_positive_labels | /m/0140xf,/m/02cjck,/m/04rlf | |
| | +--------------------------+----------------------------------+ |
| | | aspect_list | ['low quality', 'sustained strin | |
| | | | gs melody', 'soft female vocal', | |
| | | | 'mellow piano melody', 'sad', ' | |
| | | | soulful', 'ballad'] | |
| | +--------------------------+----------------------------------+ |
| | | caption | The low quality recording featur | |
| | | | es a ballad song that contains s | |
| | | | ustained strings, mellow piano m | |
| | | | elody and soft female vocal sing | |
| | | | ing over it. It sounds sad and s | |
| | | | oulful, like something you would | |
| | | | hear at Sunday services. | |
| | +--------------------------+----------------------------------+ |
| | | author_id | 4 | |
| | +--------------------------+----------------------------------+ |
| | | is_balanced_subset | 0 | |
| | +--------------------------+----------------------------------+ |
| | | is_audioset_eval | 1 | |
| | +--------------------------+----------------------------------+ |
+------------+-----------------------------------------------------------------+
| Features | +----------+----+ |
| | | bpm | 60 | |
| | +----------+----+ |
| | | duration | 10 | |
| | +----------+----+ |
+------------+-----------------------------------------------------------------+
Important
If you want to save the changes made to j2a/data_handler/.json_files/data_summary.json you have to use the save method.
To save the changes to the data_summary.json file use the save method. If there are any changes to the files this will be printed to the console. An example of this is shown below:
json_interactor.save()This will print the following to the console:
('change', 'MusicCaps_0.sound_path', ('/usr/local/path/to/data/MusicCaps_0.mp3', '/usr/local/new_path/to/data/MusicCaps_0.mp3'))
('add', 'MusicCaps_0.features', [('bpm', 60), ('duration', 10)])
Are you sure you want to make the changes to the data_summary.json? [y/n]If you are sure you want to make the changes to the data_summary.json file type y and press enter. If you are not sure type n and press enter - in this case, the changes will still be on the json_interactor object, but will not be saved to the file. To discard all changes to the json_interactor just create a new JsonInteractor object.
Features, such as key and tempo, can be extracted from a song by using this module. To extract every supported feature you should access the extract_all function found in the extract_all.py file. Below is an example of how this is done.
from j2a.feature_extractor.extract_all import extract_all
path = "data/youtube/MusicCaps-8.mp3"
features = extract_all(path)This returns a typed dictionary, making it easy to see which features will be extracted.
Arguments can be passed to the extract_all function to enable/disable cuda functionality, or change the model used for extracting the genre of a song. Make sure you disable cuda if this is not available on your system, since it is enabled by default. All models available to the genre extractor are HuggingFace models, and an Enum class that lists these is defined in this file.
You could use the individual extractors, but this is not advised since the extractors using madmom do not produce a predictable output.
The encoder used is based on Masked Autoencoders with Multi-Window Local-Global Attention Are Better Audio Learners. To encode all data within a folder (only works with .wav files) the j2a.encoder.encode.py can be run with the following arguments:
python j2a/encoder/encode.py --audio_folder <path/to/data_folder> --output_folder <path/to/output_folder>This will encode all .wav files in the audio_path folder and store the encoded data in the output_path folder. The encoded data is stored in .pt files (this is the format used for torch tensors, so it can be loaded with torch.load()). By default, the encoding will create encoding to the .json file described in Data Extractor. However, if you want to encode the data without the .json file being updated then --save_to_json can be set to False.
If the weights and biases from the mwmae-jax-official encoder are not already downloaded, then this is done automatically. The weights and biases are stored in j2a/encoder/.weights_biases.
Important
If there are issues with the submodule mwmae-jax-official remember to use the git pull --recurse-submodules command when pulling the repository. If the submodule is not downloaded, then use the git submodule update --init --recursive command.