Duck And Cover Genre Encoder

🚀 Duck and Cover - Genre Autoencoder

This model generates album covers based on multiple conditions such as release year, artist, album name, and genres, using different types of GANs.

🚀 Quick Start

This model is part of the duck_and_cover repository. The scope of this repository is to generate album covers based on several conditions like release year, artist & album name, and genre(s) using different types of GANs. The possible list of genres that this encoder covers can be found here.

🔧 Technical Details

For training, prajjwal1/bert-mini has been finetuned on a list of 466,045 albums with different genre combinations taken from the aforementioned list to embed genre information. Meanwhile, a simple Linear Layer was trained to decode and predict the given genre from the embeddings. The albums are real - world albums retrieved using the Spotify API.

The intention behind this model is that Hard Rock is somehow related to Rock, while Pop Rock is related to Rock as well. A BERT Tokenizer can capture this information as a lot of music genres are described by using pre - and suffixes.

The model was validated on 133,155 samples during training and tested on 66,578 samples. It yields a 98.29% Exact Match ratio on the testset and a 98.24% Exact Match Ratio on the validation set. This is extremely high given that the model can embed up to 3452 labels and most of the albums only had up to 5 labels.

💻 Usage Examples

Basic Usage

The model can be used to embed genres to a 256 dimensional space using the following input.

from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained("mnne/duck-and-cover-genre-encoder")
tokenizer = AutoTokenizer.from_pretrained("mnne/duck-and-cover-genre-encoder")

genres = " , ".join(["classic soul", "memphis soul", "soul", "soul blues", "southern soul"])

x = tokenizer([genres], return_tensors="pt")

output = model(**x)