import gradio as gr
from diffusers import AudioLDMControlNetPipeline, ControlNetModel
from pretty_midi import PrettyMIDI
import torch

if torch.cuda.is_available():
    device = "cuda"
    torch_dtype = torch.float16
else:
    device = "cpu"
    torch_dtype = torch.float32

controlnet = ControlNetModel.from_pretrained(
    "lauraibnz/midi-audioldm", torch_dtype=torch_dtype)
pipe = AudioLDMControlNetPipeline.from_pretrained(
    "cvssp/audioldm-m-full", controlnet=controlnet, torch_dtype=torch_dtype)
pipe = pipe.to(device)
generator = torch.Generator(device)


def predict(midi_file=None, prompt="", negative_prompt="", audio_length_in_s=5, random_seed=0, controlnet_conditioning_scale=1, num_inference_steps=20, guess_mode=False):
    midi_file = midi_file.name
    midi = PrettyMIDI(midi_file)
    audio = pipe(
        prompt,
        negative_prompt=negative_prompt,
        midi=midi,
        audio_length_in_s=audio_length_in_s,
        num_inference_steps=num_inference_steps,
        controlnet_conditioning_scale=float(controlnet_conditioning_scale),
        guess_mode=guess_mode,
        generator=generator.manual_seed(int(random_seed)),
    )
    return (16000, audio.audios.T)


demo = gr.Interface(
        fn=predict, inputs=[
            gr.File(label="midi file", file_types=[".mid"]),
            "text",
            gr.Textbox(label="negative prompt"),
            gr.Slider(0, 30, value=5, step=5, label="duration (seconds)"),
            gr.Number(value=42, label="seed"),
            gr.Slider(0.0, 1.0, value=1.0, step=0.1, label="conditioning scale"),
            gr.Slider(0, 50, value=20, step=0.1, label="inference steps"),
            gr.Checkbox(label="guess mode")
            ], 
        outputs="audio", 
        examples=[["S00.mid", "piano", "", 10, 25, 1.0, 20, False]], 
        cache_examples=True, 
        title="🎹 MIDI-AudioLDM",
        description="MIDI-AudioLDM is a MIDI-conditioned text-to-audio model based on the project [AudioLDM](https://huggingface.co/spaces/haoheliu/audioldm-text-to-audio-generation). The model has been conditioned using the ControlNet architecture and has been developed within Hugging Face’s [🧨 Diffusers](https://huggingface.co/docs/diffusers/) framework. Once trained, MIDI-AudioLDM accepts a MIDI file and a text prompt as inputs and returns an audio file, which is an interpretation of the MIDI based on the given text description. This enables detailed control over different musical aspects such as notes, mood and timbre.", 
        theme=gr.themes.Base(text_size=gr.themes.sizes.text_md, font=[gr.themes.GoogleFont("Nunito Sans")])
    )
demo.launch()