app.py

import gradio as gr
import torch
import numpy as np
from modules.models import *
from util import get_prompt_template
from torchvision import transforms as vt
import torchaudio
from PIL import Image


def greet(audio, image):
    device = "cuda:0" if torch.cuda.is_available() else "cpu"

    # Get model
    model_conf_file = f'./config/model/ACL_ViT16.yaml'
    model = ACL(model_conf_file, device)
    model.train(False)
    model.load('./pretrain/Param_best.pth')

    # Get placeholder text
    prompt_template, text_pos_at_prompt, prompt_length = get_prompt_template()

    # Input pre processing
    sample_rate, audio = audio
    audio = audio.astype(np.float32, order='C') / 32768.0
    desired_sample_rate = 16000
    set_length = 10

    audio_file = torch.from_numpy(audio)

    if desired_sample_rate != sample_rate:
        audio_file = torchaudio.functional.resample(audio_file, sample_rate, desired_sample_rate)

    if audio_file.shape[0] == 2:
        audio_file = torch.concat([audio_file[0], audio_file[1]], dim=0)  # Stereo -> mono (x2 duration)

    audio_file.squeeze(0)

    if audio_file.shape[0] > (desired_sample_rate * set_length):
        audio_file = audio_file[:desired_sample_rate * set_length]

    # zero padding
    if audio_file.shape[0] < (desired_sample_rate * set_length):
        pad_len = (desired_sample_rate * set_length) - audio_file.shape[0]
        pad_val = torch.zeros(pad_len)
        audio_file = torch.cat((audio_file, pad_val), dim=0)

    audio_file = audio_file.unsqueeze(0)

    image_transform = vt.Compose([
        vt.Resize((352, 352), vt.InterpolationMode.BICUBIC),
        vt.ToTensor(),
        vt.Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)),  # CLIP
    ])

    image_file = image_transform(image)

    # Inference
    placeholder_tokens = model.get_placeholder_token(prompt_template.replace('{}', ''))
    audio_driven_embedding = model.encode_audio(audio_file.to(model.device), placeholder_tokens, text_pos_at_prompt,
                                                prompt_length)

    # Localization result
    out_dict = model(image_file.to(model.device), audio_driven_embedding, 352)
    seg = out_dict['heatmap'][j:j + 1]
    seg_image = ((1 - seg.squeeze().detach().cpu().numpy()) * 255).astype(np.uint8)
    seg_image = Image.fromarray(seg_image)
    heatmap_image = cv2.applyColorMap(np.array(seg_image), cv2.COLORMAP_JET)
    overlaid_image = cv2.addWeighted(np.array(image), 0.5, heatmap_image, 0.5, 0)

    return overlaid_image


description = 'hello world'

demo = gr.Interface(
    fn=greet,
    inputs=[gr.Image(type='pil'), gr.Audio()],
    outputs=gr.Image(type="pil"),
    title='AudioToken',
    description=description,
)

demo.launch(debug=True)