import random
import subprocess
import torch

import numpy as np
import gradio as gr

from transformers import AutoProcessor, MusicgenForConditionalGeneration

COLORS = [
    ["#ff0000", "#00ff00"],
    ["#00ff00", "#0000ff"],
    ["#0000ff", "#ff0000"],
]

path = "facebook/musicgen-large"
processor = AutoProcessor.from_pretrained(path)
model = MusicgenForConditionalGeneration.from_pretrained(path, torch_dtype=torch.float16).to("cuda")

def predict(text):
    inputs = processor(
        text=[text],
        padding=True,
        return_tensors="pt",).to("cuda")
    
    with torch.autocast("cuda"):
        outputs = model.generate(**inputs, do_sample=True, guidance_scale=3, max_new_tokens=512)

        video_path = gr.make_waveform((32000, outputs[0].cpu().numpy().astype(np.float16).ravel()), bars_color=random.choice(COLORS), bar_count=75)
        command = f'ffmpeg -y -i {video_path} -vf "scale=250:150" result.mp4'
        subprocess.run(command, shell=True) 

        return (32000, outputs[0][0].cpu().numpy().astype(np.float16)), "result.mp4"

title = "MusicGen"

gr.Interface(
    fn=predict,
    inputs=[
        gr.Textbox(label="Text prompt"),
    ],
    outputs=["audio", "video"],
    title=title,
    theme="gradio/monochrome",
).queue(max_size=10).launch()