import torch
import gradio as gr
from diffusers import AudioLDMPipeline
from transformers import AutoProcessor, ClapModel

# make Space compatible with CPU duplicates
if torch.cuda.is_available():
    device = "cuda"
    torch_dtype = torch.float16
else:
    device = "cpu"
    torch_dtype = torch.float32

# load the diffusers pipeline
repo_id = "cvssp/audioldm-m-full"
pipe = AudioLDMPipeline.from_pretrained(repo_id, torch_dtype=torch_dtype).to(device)
pipe.unet = torch.compile(pipe.unet)

# CLAP model (only required for automatic scoring)
clap_model = ClapModel.from_pretrained("sanchit-gandhi/clap-htsat-unfused-m-full").to(device)
processor = AutoProcessor.from_pretrained("sanchit-gandhi/clap-htsat-unfused-m-full")

generator = torch.Generator(device)

def score_waveforms(text, waveforms):
    inputs = processor(text=text, audios=list(waveforms), return_tensors="pt", padding=True)
    inputs = {key: inputs[key].to(device) for key in inputs}
    with torch.no_grad():
        logits_per_text = clap_model(**inputs).logits_per_text  # this is the audio-text similarity score
        probs = logits_per_text.softmax(dim=-1)  # we can take the softmax to get the label probabilities
        most_probable = torch.argmax(probs)  # and now select the most likely audio waveform
    waveform = waveforms[most_probable]
    return waveform

def text_to_music(text_input, negative_prompt, seed, duration, guidance_scale, n_candidates):
    waveforms = pipe(
        text_input,
        audio_length_in_s=duration,
        guidance_scale=guidance_scale,
        num_inference_steps=100,
        negative_prompt=negative_prompt,
        num_waveforms_per_prompt=n_candidates if n_candidates else 1,
        generator=generator.manual_seed(int(seed)),
    )["audios"]

    if waveforms.shape[0] > 1:
        waveform = score_waveforms(text_input, waveforms)
    else:
        waveform = waveforms[0]

    return waveform.detach().cpu().numpy()

iface = gr.Interface(
    fn=text_to_music,
    inputs=[
        gr.inputs.Textbox(label="Input text", default="A hammer is hitting a wooden surface"),
        gr.inputs.Textbox(label="Negative prompt", default="low quality, average quality"),
        gr.inputs.Number(label="Seed", default=45),
        gr.inputs.Slider(label="Duration (seconds)", minimum=2.5, maximum=10.0, default=5.0, step=0.1),
        gr.inputs.Slider(label="Guidance scale", minimum=0.0, maximum=4.0, default=2.5, step=0.1),
        gr.inputs.Slider(label="Number waveforms to generate", minimum=1, maximum=3, default=3, step=1),
    ],
    outputs=gr.outputs.Audio(label="Generated Audio", type="numpy"),
    live=True,
    title="Text to Music",
    description="Convert text into music using a pre-trained model.",
    theme="default",
)

iface.launch()