---
datasets:
- AnyaSchen/image2music_abc
tags:
- music
- image
---

This repo contains model for music generation from images. The generated music returns in ABC format and it can be sound for example [here](https://editor.drawthedots.com/). Note, that you need to correct BPM (this is speed) to make music more logical and natural.
The model is fune-tuned concatecation of two pre-trained models: [google/vit-base-patch16-224](https://huggingface.co./google/vit-base-patch16-224) as encoder and [sander-wood/text-to-music](sander-wood/text-to-music) as decoder.
To use this model you can write this:

```
from PIL import Image
import requests
from transformers import AutoTokenizer, VisionEncoderDecoderModel, ViTImageProcessor

def generate_music(model, image, tokenizer):
    pixel_values = feature_extractor(images=image, return_tensors="pt").pixel_values
    pixel_values = pixel_values.to(device)

    generated_tokens = model.generate(
        pixel_values,
        max_length=300,
        num_beams=5,
        top_p=0.8,
        temperature=2.0,
        do_sample=True,
        pad_token_id=tokenizer.pad_token_id,
        eos_token_id=tokenizer.eos_token_id,
    )

    generated_music = tokenizer.decode(generated_tokens[0], skip_special_tokens=True)
    return generated_music

path = 'AnyaSchen/image2music'
fine_tuned_model = VisionEncoderDecoderModel.from_pretrained(path).to(device)
feature_extractor = ViTImageProcessor.from_pretrained(path)
tokenizer = AutoTokenizer.from_pretrained(path)

url = 'https://anandaindia.org/wp-content/uploads/2018/12/happy-man.jpg'
image = Image.open(requests.get(url, stream=True).raw)

generated_music = generate_music(fine_tuned_model, image, tokenizer)
print(generated_music)

```