app.py

import os
import re
import json
import argparse

# get arguments
parser = argparse.ArgumentParser(description="VITS model for Blue Archive characters")
parser.add_argument('--device', type=str, default='cpu', choices=["cpu","cuda"], help="inference method (default: cpu)")
parser.add_argument("--share", action="store_true", default=False, help="share gradio app")
parser.add_argument("--queue", type=int, default=1, help="max number of concurrent workers for the queue (default: 1, set to 0 to disable)")
parser.add_argument('--api', action="store_true", default=False, help="enable REST routes (allows to skip queue)")
parser.add_argument("--limit", type=int, default=100, help="prompt words limit (default: 100, set to 0 to disable)")
args = parser.parse_args()
#limit_text = os.getenv("SYSTEM") == "spaces"  # limit text and audio length in huggingface spaces

import torch
import utils
import commons
from models import SynthesizerTrn
from text import text_to_sequence

import gradio as gr

metadata_path = "./models/metadata.json"
config_path = "./models/config.json"
default_sid = 10

# load each model's definition
with open(metadata_path, "r", encoding="utf-8") as f:
    metadata = json.load(f)

##################### MODEL LOADING #####################
# loads one model and returns a container with all references to the model
device = torch.device(args.device)
def load_checkpoint(model_path):    
    hps = utils.get_hparams_from_file(config_path)
    net_g = SynthesizerTrn(
        len(hps.symbols),
        hps.data.filter_length // 2 + 1,
        hps.train.segment_size // hps.data.hop_length,
        n_speakers=hps.data.n_speakers,
        **hps.model)
    model = net_g.eval().to(device)
    pythomodel = utils.load_checkpoint(model_path, net_g, None)
    
    return {"hps": hps, "net_g": net_g}

# loads all models on system memory and returns a dict with all references
def load_models():
    models = {}
    for name, model_info in metadata.items():
        model = load_checkpoint(model_info["model"])
        models[name] = model

    return models

models = load_models()

##################### INFERENCE #####################
def get_text(text, hps):
    text_norm = text_to_sequence(text, hps.data.text_cleaners)
    if hps.data.add_blank:
        text_norm = commons.intersperse(text_norm, 0)
    text_norm = torch.LongTensor(text_norm)
    return text_norm

def prepare_text(text):
    prepared_text = text.replace('\n', ' ').replace('\r', '').replace(" ", "")
    if args.limit > 0:
        text_len = len(re.sub("\[([A-Z]{2})\]", "", text))
        max_len = args.limit
        if text_len > max_len:
            return None
    return prepared_text

# inferes a model
def speak(selected_index, text, noise_scale, noise_scale_w, length_scale, speaker_id):
    # get selected model
    model = list(models.items())[selected_index][1]
    # clean and truncate text
    text = prepare_text(text)
    if text is None:
        return ["Error: text is too long", None]
    
    stn_tst = get_text(text, model["hps"])
    with torch.no_grad():
        x_tst = stn_tst.unsqueeze(0).to(device)
        x_tst_lengths = torch.LongTensor([stn_tst.size(0)]).to(device)
        sid = torch.LongTensor([speaker_id]).to(device)
        audio = model["net_g"].infer(x_tst, x_tst_lengths, sid=sid, noise_scale=noise_scale, noise_scale_w=noise_scale_w, length_scale=length_scale)[0][0,0].data.cpu().float().numpy()

    return ["Success", (22050, audio)]

##################### GRADIO UI #####################
# javascript function for creating a download link, reads the audio file from the DOM and returns it with the prompt as the filename
download_audio_js = """
() =>{{
    let root = document.querySelector("body > gradio-app");
    if (root.shadowRoot != null)
        root = root.shadowRoot;
    let audio = root.querySelector("#output-audio").querySelector("audio");
    let text = root.querySelector("#input-text").querySelector("textarea");
    if (audio == undefined)
        return;
    text = text.value;
    if (text == undefined)
        text = Math.floor(Math.random()*100000000);
    audio = audio.src;
    let oA = document.createElement("a");
    oA.download = text.substr(0, 20)+'.wav';
    oA.href = audio;
    document.body.appendChild(oA);
    oA.click();
    oA.remove();
}}
"""
def make_gradio_ui():
    with gr.Blocks(css="#avatar {width:auto;height:200px;}") as demo:
        # make title
        gr.Markdown(
            """
            # <center> VITS for Blue Archive Characters
            ## <center> This is for educational purposes, please do not generate harmful or inappropriate content\n
            [![HitCount](https://hits.dwyl.com/tovaru/vits-for-ba.svg?style=flat-square&show=unique)](http://hits.dwyl.com/tovaru/vits-for-ba)\n
            This is based on [zomehwh's implementation](https://huggingface.co/spaces/zomehwh/vits-models), please visit their space if you want to finetune your own models!
            """
        )

        first = list(metadata.items())[0][1]
        with gr.Row():
            # setup column
            with gr.Column():
                # set text input field
                text = gr.Textbox(
                    lines=5,
                    interactive=True,
                    label=f"Text ({args.limit} words limit)" if args.limit > 0 else "Text",
                    value=first["example_text"],
                    elem_id="input-text")                

                # button for loading example texts
                example_btn = gr.Button("Example", size="sm")
                
                # character selector
                names = []
                for _, model in metadata.items():
                    names.append(model["name_en"])
                character = gr.Radio(label="Character", choices=names, value=names[0], type="index")

                # inference parameters
                with gr.Row():
                    ns = gr.Slider(label="noise_scale", minimum=0.1, maximum=1.0, step=0.1, value=0.6, interactive=True)
                    nsw = gr.Slider(label="noise_scale_w", minimum=0.1, maximum=1.0, step=0.1, value=0.668, interactive=True)
                with gr.Row():
                    ls = gr.Slider(label="length_scale", minimum=0.1, maximum=2.0, step=0.1, value=1, interactive=True)
                    sid = gr.Number(value=default_sid, label="Speaker ID (increase or decrease to change the intonation)")

                # generate button
                generate_btn = gr.Button(value="Generate", variant="primary")
                
            # results column
            with gr.Column():
                avatar = gr.Image(first["avatar"], type="pil", elem_id="avatar")
                output_message = gr.Textbox(label="Result Message")
                output_audio = gr.Audio(label="Output Audio", interactive=False, elem_id="output-audio")
                download_btn = gr.Button("Download Audio")

            # set character selection updates
            def update_selection(index):
                avatars = list(metadata.items())
                new_avatar = avatars[index][1]["avatar"]
                return gr.Image.update(value=new_avatar)
            
            character.change(update_selection, character, avatar)

        # set generate button action
        generate_btn.click(fn=speak, inputs=[character, text, ns, nsw, ls, sid], outputs=[output_message, output_audio], api_name="generate")

        # set download butto naction
        download_btn.click(None, [], [], _js=download_audio_js)

        def load_example(index):
            example = list(metadata.items())[index][1]["example_text"]
            return example
        example_btn.click(fn=load_example, inputs=[character], outputs=[text])
    
    return demo

if __name__ == "__main__":   
    demo = make_gradio_ui()
    if args.queue > 0:
        demo.queue(concurrency_count=args.queue, api_open=args.api)
    demo.launch(share=args.share, show_api=args.api)