app.py

import os
import argparse
import numpy as np
from PIL import Image
import torch
import torchvision.transforms as T
from transformers import AutoTokenizer
import gradio as gr
from resnet50 import build_model
# from utils import generate_similiarity_map, post_process, load_tokenizer, build_transform_R50
from utils import generate_similiarity_map, get_transform, post_process, load_tokenizer, build_transform_R50
from utils import IMAGENET_MEAN, IMAGENET_STD
from internvl.train.dataset import dynamic_preprocess
from internvl.model.internvl_chat import InternVLChatModel
import spaces

# 模型配置
CHECKPOINTS = {
    "TokenFD_4096_English_seg": "TongkunGuan/TokenFD_4096_English_seg",
    "TokenFD_2048_Bilingual_seg": "TongkunGuan/TokenFD_2048_Bilingual_seg",
}

# 全局变量
HF_TOKEN = os.getenv("HF_TOKEN")



def load_model(check_type):
    # device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    device = torch.device("cuda")
    if check_type == 'R50':
        tokenizer = load_tokenizer('tokenizer_path')
        model = build_model(argparse.Namespace()).eval()
        model.load_state_dict(torch.load(CHECKPOINTS['R50'], map_location='cpu')['model'])
        transform = build_transform_R50(normalize_type='imagenet')

    elif check_type == 'R50_siglip':
        tokenizer = load_tokenizer('tokenizer_path')
        model = build_model(argparse.Namespace()).eval()
        model.load_state_dict(torch.load(CHECKPOINTS['R50_siglip'], map_location='cpu')['model'])
        transform = build_transform_R50(normalize_type='imagenet')

    elif 'TokenFD' in check_type:
        model_path = CHECKPOINTS[check_type]
        tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True, use_fast=False, use_auth_token=HF_TOKEN)
        # model = InternVLChatModel.from_pretrained(model_path, torch_dtype=torch.bfloat16).eval()
        model = InternVLChatModel.from_pretrained(model_path, low_cpu_mem_usage=True, torch_dtype=torch.bfloat16 ,load_in_8bit=False, load_in_4bit=False).eval()
        transform = get_transform(is_train=False, image_size=model.config.force_image_size)
        
    return model.to(device), tokenizer, transform, device

def process_image(model, tokenizer, transform, device, check_type, image, text):
    
    src_size = image.size
    if 'TokenFD' in check_type:
        images, target_ratio = dynamic_preprocess(image, min_num=1, max_num=12, 
                                                  image_size=model.config.force_image_size,
                                                  use_thumbnail=model.config.use_thumbnail,
                                                  return_ratio=True)
        pixel_values = torch.stack([transform(img) for img in images]).to(device)
    else:
        pixel_values = torch.stack([transform(image)]).to(device)
        target_ratio = (1, 1)

    # 文本处理
    text_input = text

    if text_input[0] in '!"#$%&\'()*+,-./0123456789:;<=>?@^_{|}~0123456789':
        input_ids = tokenizer(text_input)['input_ids'][1:]
    else:
        input_ids = tokenizer(' '+text_input)['input_ids'][1:]

    input_ids = torch.tensor(input_ids, device=device)
    
    # 获取嵌入
    with torch.no_grad():
        if 'R50' in check_type:
            text_embeds = model.language_embedding(input_ids)
        else:
            text_embeds = model.tok_embeddings(input_ids).clone()
        
        vit_embeds, size1 = model.forward_tokenocr(pixel_values.to(torch.bfloat16).to(device))
        
        print("vit_embeds",vit_embeds)
        print("vit_embeds,shape",vit_embeds.shape)
        print("target_ratio",target_ratio)
        print("check_type",check_type)
        vit_embeds, size2 = post_process(vit_embeds, target_ratio, check_type)
        
        # 计算相似度
        text_embeds = text_embeds / text_embeds.norm(dim=-1, keepdim=True)
        vit_embeds = vit_embeds / vit_embeds.norm(dim=-1, keepdim=True)
        similarity = text_embeds @ vit_embeds.T
        resized_size = size1 if size1 is not None else size2

    # print(f"text_embeds shape: {text_embeds.shape}, numel: {text_embeds.numel()}") # text_embeds shape: torch.Size([4, 2048]), numel: 8192
    # print(f"vit_embeds shape: {vit_embeds.shape}, numel: {vit_embeds.numel()}") # vit_embeds shape: torch.Size([9728, 2048]), numel: 19922944
    # print(f"similarity shape: {similarity.shape}, numel: {similarity.numel()}")# similarity shape: torch.Size([4, 9728]), numel: 38912


    # 生成可视化
    attn_map = similarity.reshape(len(text_embeds), resized_size[0], resized_size[1])
    # attn_map = similarity.reshape(len(text_embeds), *target_ratio)
    all_bpe_strings = [tokenizer.decode(input_id) for input_id in input_ids]
    current_vis = generate_similiarity_map(images, attn_map, 
                                           [tokenizer.decode([i]) for i in input_ids], 
                                           [], target_ratio, src_size)
    
    current_bpe = [tokenizer.decode([i]) for i in input_ids]
    # current_bpe[-1] = 'Input text'

    
    # current_bpe.append(text)

    return image, current_vis, current_bpe

# 事件处理函数
# 上一项和下一项按钮
def update_index(direction, current_vis, current_bpe, current_index):
    # 计算新的索引
    new_index = max(0, min(current_index + direction, len(current_vis) - 1))

    # 更新可视化内容
    return (
        current_vis[new_index], 
        format_bpe_display(current_bpe[new_index]), 
        new_index  # 更新索引
    )

def format_bpe_display(bpe):
    # 使用HTML标签来设置字体大小、颜色，加粗，并居中
    return f"<div style='text-align:center; font-size:20px;'><strong>Current BPE: <span style='color:red;'>{bpe}</span></strong></div>"



# Gradio界面
with gr.Blocks(title="BPE Visualization Demo") as demo:
    gr.Markdown("## BPE Visualization Demo - TokenFD基座模型能力可视化")
    
    with gr.Row():
        with gr.Column(scale=0.5):
            model_type = gr.Dropdown(
                choices=["TokenFD_4096_English_seg", "TokenFD_2048_Bilingual_seg"],
                label="Select model type",
                value="TokenFD_4096_English_seg"  # 设置默认值为第一个选项
            )
            image_input = gr.Image(label="Upload images", type="pil")
            text_input = gr.Textbox(label="Input text")

            run_btn = gr.Button("RUN")
            
            gr.Examples(
                examples=[
                    [os.path.join("examples", "examples0.jpg"), "Veterans and Benefits"],
                    [os.path.join("examples", "examples1.jpg"), "Refreshers"],
                    [os.path.join("examples", "examples2.png"), "Vision Transformer"]
                ],
                inputs=[image_input, text_input],
                label="Sample input"
            )
        
        with gr.Column(scale=2):
            gr.Markdown("<p style='font-size:20px;'><span style='color:red;'>If the input text is not included in the image</span>, the attention map will show a lot of noise (the actual response value is very low), since we normalize the attention map according to the relative value.</p>")

            with gr.Row():
                orig_img = gr.Image(label="Original picture", interactive=False)
                heatmap = gr.Image(label="BPE visualization", interactive=False)
            
            with gr.Row() as controls:
                prev_btn = gr.Button("⬅ Last", visible=False)
                next_btn = gr.Button("⮕ Next", visible=False)
            
            bpe_display = gr.Markdown("Current BPE: ")

    current_vis_state = gr.State([])
    current_bpe_state = gr.State([])
    current_index_state = gr.State(0)
    
    # 事件处理
    @spaces.GPU
    def on_run_clicked(model_type, image, text):

        current_index = 0  # Reset index when new image is processed
        image, current_vis, current_bpe = process_image(*load_model(model_type), model_type, image, text)
        
        bpe_text = format_bpe_display(current_bpe)
        
        print("current_vis",len(current_vis))
        print("current_bpe",len(current_bpe))
        
        # return image, current_vis[0],f"<div style='text-align:center; font-size:20px;'><strong>Current BPE: <span style='color:red;'>{current_bpe[0]}</span></strong></div>", gr.update(visible=True), gr.update(visible=True)
        return (
            image, 
            current_vis[current_index], 
            format_bpe_display(current_bpe[current_index]), 
            gr.update(visible=True), 
            gr.update(visible=True), 
            current_vis,  # 存储整个列表
            current_bpe,  # 存储整个列表
            current_index  # 存储当前索引
        )

    
    
    run_btn.click(
        on_run_clicked,
        inputs=[model_type, image_input, text_input],
        outputs=[orig_img, heatmap, bpe_display, prev_btn, next_btn, current_vis_state, current_bpe_state, current_index_state]
    )
    
    prev_btn.click(
        update_index,
        inputs=[gr.State(-1), current_vis_state, current_bpe_state, current_index_state],
        outputs=[heatmap, bpe_display, current_index_state]
    )
    
    next_btn.click(
        update_index,
        inputs=[gr.State(1), current_vis_state, current_bpe_state, current_index_state],
        outputs=[heatmap, bpe_display, current_index_state]
    )






if __name__ == "__main__":
    demo.launch()