|
--- |
|
dataset_info: |
|
features: |
|
- name: image |
|
dtype: image |
|
- name: text |
|
dtype: string |
|
splits: |
|
- name: train |
|
num_bytes: 3322131399.86 |
|
num_examples: 9604 |
|
download_size: 3282715107 |
|
dataset_size: 3322131399.86 |
|
license: mit |
|
task_categories: |
|
- text-to-image |
|
- text-to-audio |
|
language: |
|
- en |
|
tags: |
|
- music |
|
pretty_name: Mel-Spectrograms for Irish Traditional Music |
|
size_categories: |
|
- 1K<n<10K |
|
--- |
|
# Dataset Card for "irish-traditional-tunes" |
|
|
|
[More Information needed](https://github.com/huggingface/datasets/blob/main/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards) |
|
# Dataset Card for "irish-tunes-spectrograms" |
|
|
|
## 1. Dataset Description |
|
Dataset is used for the following project |
|
- **Homepage:** [Trad-fusion](https://github.com/hdparmar/Tradi-fusion) |
|
|
|
### 1.1 Dataset Summary |
|
This dataset contains 9604 Mel spectrograms that represent Traditional Irish Music. |
|
This dataset is smaller compared to [hdparmar/irish-tunes-spectrogram](https://huggingface.co/datasets/hdparmar/irish-tunes-spectrograms), to reduce the training time and increase the possibilty to train for longer steps/batch. |
|
Each spectrogram image is a 5 second split of audio resulting in dimensions 512x512 and includes 3 channels (mimicking, RGB) because most of the text-to-image models are trained on 3 channels. |
|
|
|
Although, I can find publications which says that having 3 channels for Mel Spectrogram can improve generalisation, since the other 2 channel are just the copy of first. |
|
The simple trick I used is to use cv2 to convert a grayscale into RGB, since most of the models are trained on 3 channels. |
|
|
|
The primary objective of this dataset is to serve as an abundant resource for those venturing into the fields of music analysis, machine learning, and artificial intelligence. |
|
|
|
### 1.2 Languages |
|
The dataset's metadata and documentation are all in English, ensuring accessibility and comprehension. |
|
|
|
## 2. Dataset Structure |
|
|
|
### 2.1 Data Instances |
|
Each data instance in this dataset is composed of two main elements: an image and a text caption. |
|
The image is a mel spectrogram that reflects a snippet of a traditional Irish tune. Accompanying it is a text field that serves as its caption. |
|
|
|
#### Example: |
|
The metadata.csv file the dataset is in this format |
|
``` |
|
{"file_name": "path/to/the/image.png", |
|
"text": "An Irish Traditional Tune"} |
|
``` |
|
### 2.2 Data Fields |
|
- **file_name**: This is the field that contains the path leading to the image file. It's the specific location where you can find each piece of the dataset. |
|
- **text**: This is the caption accompanying each image. For the sake of uniformity and ease, the caption for every image is "An Irish Traditional Tune." |
|
|
|
### 2.3 Data Splits |
|
As of the current version, the dataset consists solely of a training split. Additional data splits like validation or testing may be introduced in future iterations of the dataset. |
|
|
|
### 2.4 Uniform Captions: A Special Note |
|
All the spectrograms in this dataset come labeled with a uniform caption: "An Irish Traditional Tune." |
|
|
|
This consistency can be perhaps advantageous, especially in text-to-image tasks that focus primarily on image-based features, with the caption acting as a generalized label. |
|
|
|
## NOTE |
|
Furthur imformation to follow and same caption for all the mel-spectrograms are for ease of work put into producing the dataset |
