app.py

import torch
import numpy as np
import gradio as gr
from PIL import Image
from omegaconf import OmegaConf
from pathlib import Path
from vocoder.bigvgan.models import VocoderBigVGAN
from ldm.models.diffusion.ddim import DDIMSampler
from ldm.util import instantiate_from_config

SAMPLE_RATE = 16000

torch.set_grad_enabled(False)
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")


def initialize_model(config, ckpt=None):
    config = OmegaConf.load(config)
    model = instantiate_from_config(config.model)
    if ckpt == None:
        print("not load state dict")
    else:
        model.load_state_dict(torch.load(ckpt,map_location='cpu')["state_dict"], strict=False)

    model = model.to(device)
    model.cond_stage_model.to(model.device)
    model.cond_stage_model.device = model.device
    print(model.device,device,model.cond_stage_model.device)
    sampler = DDIMSampler(model)

    return sampler

sampler = initialize_model('configs/img_to_audio/img2audio_args.yaml', 'useful_ckpts/ta54_epoch=000216.ckpt')
vocoder = VocoderBigVGAN('vocoder/logs/bigv16k53w',device=device)


def img2audio(sampler,vocoder,image, seed, scale, ddim_steps, W=624, H=80):
    # print(type(image))# np.ndarray
    
    n_samples = 1 # only support 1 sample
    prng = np.random.RandomState(seed)
    start_code = prng.randn(n_samples, sampler.model.first_stage_model.embed_dim, H // 8, W // 8)
    start_code = torch.from_numpy(start_code).to(device=device, dtype=torch.float32)
    
    uc = None
    if scale != 1.0:
        uc = sampler.model.get_learned_conditioning(n_samples * [""])
    
    image = Image.fromarray(image)
    image = sampler.model.cond_stage_model.preprocess(image).unsqueeze(0)
    image_embedding = sampler.model.cond_stage_model.forward_img(image)
    c = image_embedding.repeat(n_samples, 1, 1)# shape:[1,77,1280],即还没有变成句子embedding，仍是每个单词的embedding
    
    shape = [sampler.model.first_stage_model.embed_dim, H//8, W//8]  # (z_dim, 80//2^x, 848//2^x)
    samples_ddim, _ = sampler.sample(S=ddim_steps,
                                        conditioning=c,
                                        batch_size=n_samples,
                                        shape=shape,
                                        verbose=False,
                                        unconditional_guidance_scale=scale,
                                        unconditional_conditioning=uc,
                                        x_T=start_code)

    x_samples_ddim = sampler.model.decode_first_stage(samples_ddim)
    x_samples_ddim = torch.clamp((x_samples_ddim+1.0)/2.0, min=0.0, max=1.0) # [0, 1]

    wav_list = []
    for idx,spec in enumerate(x_samples_ddim):
        wav = vocoder.vocode(spec)
        wav_list.append((SAMPLE_RATE,wav))
    best_wav = wav_list[0]
    return best_wav


def predict(image, ddim_steps, scale, seed):# 经过试验，这个input_image需要是256x256、512x512的大小效果才正常，实际应该resize一下，输出再resize回去，但是他们使用的是pad，不知道为什么
    melbins,mel_len = 80,624
    with torch.no_grad():
        result = img2audio(
            sampler=sampler,
            vocoder=vocoder,
            image=image,
            seed=seed,
            scale=scale,
            ddim_steps=ddim_steps,
            H=melbins, W=mel_len
        )

    return result


with gr.Blocks() as demo:
    with gr.Row():
        gr.Markdown("## Make-An-Audio: Text-to-Audio Generation")

    with gr.Row():
        with gr.Column():
            image = gr.Image(label="Input Image: Select ome image to upload       ")
            run_button = gr.Button(label="Run")

            
            with gr.Accordion("Advanced options", open=False):
                # num_samples = 1
                ddim_steps = gr.Slider(label="Steps", minimum=1,
                                       maximum=1000, value=100, step=1)
                scale = gr.Slider(
                    label="Guidance Scale:(Large => more relevant to text but the quality may drop)", minimum=0.1, maximum=4.0, value=1.5, step=0.1
                )
                seed = gr.Slider(
                    label="Seed:Change this value (any integer number) will lead to a different generation result.",
                    minimum=0,
                    maximum=2147483647,
                    step=1,
                    value=44,
                )

        with gr.Column():
            # audio_list = []
            # for i in range(int(num_samples)):
            #     audio_list.append(gr.outputs.Audio())
            outaudio = gr.Audio()


    run_button.click(fn=predict, inputs=[
                    image,ddim_steps, scale, seed], outputs=[outaudio])# inputs的参数只能传gr.xxx
    with gr.Row():
        with gr.Column():
            gr.Examples(
                        examples = [['./example_imgs/cat.png',100,3,55],['./example_imgs/violin.png',100,3,55]],
                        inputs = [image,ddim_steps, scale, seed],
                        outputs = [outaudio]
                        )
        with gr.Column():
            pass

demo.launch()