webui_demo.py

import argparse
import os
import random

import numpy as np
import torch
import torch.backends.cudnn as cudnn
import gradio as gr
from PIL import Image

from transformers.generation import GenerationConfig

from lavis.common.config import Config
from lavis.common.dist_utils import get_rank
from lavis.common.registry import registry
from lavis.models import load_model_and_preprocess

from functools import partial
from copy import deepcopy

import cv2

def extract_frames(video_path, num_frames):
    cap = cv2.VideoCapture(video_path)
    total_num_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    sampling_interval = int(total_num_frames / num_frames)
        
    if sampling_interval == 0:  # total_frames < target_frames, 逐帧提取
        sampling_interval = 1

    images = []
    frame_count = 0
    while cap.isOpened():
        ret, frame = cap.read()
        if not ret:
            break
        if frame_count % sampling_interval == 0:
            frame = frame[:, :, ::-1]  # BGR to RGB
            images.append(Image.fromarray(frame).convert("RGB"))
        frame_count += 1
        if len(images) >= num_frames:
            break
    cap.release()

    if len(images) == 0:
        raise AssertionError(f"Video not found or no frames extracted: {video_path}")

    return images

def disable_torch_init():
    """
    Disable the redundant torch default initialization to accelerate model creation.
    """
    setattr(torch.nn.Linear, "reset_parameters", lambda self: None)
    setattr(torch.nn.LayerNorm, "reset_parameters", lambda self: None)

def _load_model_processor(args):
    if args.cpu_only:
        device_map = "cpu"
    else:
        device_map = 'cuda:{}'.format(args.gpu_id)
    
    global load_model_and_preprocess
    load_model_and_preprocess = partial(load_model_and_preprocess,is_eval=True,device=device_map)

    model, vis_processors, _ = load_model_and_preprocess("minigpt4qwen", args.model_type, llm_device_map=args.llm_device_map)
    model.load_checkpoint(args.checkpoint_path)

    model.llm_model.transformer.bfloat16()
    model.llm_model.lm_head.bfloat16()

    generation_config = {
    "chat_format": "chatml",
    "eos_token_id": 151643,
    "pad_token_id": 151643,
    "max_window_size": 6144,
    "max_new_tokens": 512,
    "transformers_version": "4.31.0"
    }

    return model, vis_processors, generation_config

def parse_args():
    parser = argparse.ArgumentParser(description="Demo")
    parser.add_argument("--model-type",type=str,default='qwen7b_chat',choices=['qwen7b_chat','qwen14b_chat'])
    parser.add_argument("-c", "--checkpoint-path", type=str,
                        help="Checkpoint name or path, default to %(default)r")
    parser.add_argument("-s", "--seed", type=int, default=42, help="Random seed")
    parser.add_argument("--cpu-only", action="store_true", help="Run demo with CPU only")
    parser.add_argument("--gpu_id", type=int, default=0, help="specify the gpu to load the model.")
    parser.add_argument("--llm_device_map", type=str, default="cpu")
    args = parser.parse_args()
    return args

# ========================================
#             Model Initialization
# ========================================

print('Initializing Chat')
args = parse_args()

if torch.cuda.is_available() and not args.cpu_only:
    device='cuda:{}'.format(args.gpu_id)
else:
    device=torch.device('cpu')

disable_torch_init()
model, vis_processors, default_generation_config = _load_model_processor(args)
vis_processor = vis_processors["eval"]

print('Initialization Finished')

# ========================================
#             Gradio Setting
# ========================================


def gradio_reset(history, img_list):
    if history is not None:
        history = []
    if img_list is not None:
        img_list = []
    return None, \
           gr.update(value=None, interactive=True, visible=True), \
           gr.update(value=None, interactive=True, visible=False), \
           gr.update(placeholder='Please upload your image first', interactive=False), \
           gr.update(value="Upload & Start Chat", interactive=True), \
           gr.update(value=None), \
           history, \
           img_list


def load_and_process_img(image,img_list):
        if isinstance(image, str):  # is a image path
            raw_image = Image.open(image).convert('RGB')
            image = vis_processor(raw_image)
        elif isinstance(image, Image.Image):
            raw_image = image
            raw_image = raw_image.convert('RGB')
            image = vis_processor(raw_image)
        elif isinstance(image, torch.Tensor):
            if len(image.shape) == 3:
                image = image
            else:
                assert False, "the `image.ndim` must be 3"

        img_list.append(image)
        msg = "Received."
        return msg

def load_and_process_video(video_path,img_list):
        assert isinstance(video_path, str), "Input must be a path of video"
        raw_images = extract_frames(video_path, num_frames=32) # hard-code the `num_frames`(32)
        images = [vis_processor(raw_image) for raw_image in raw_images]

        img_list.extend(images)
        msg = "Received."
        return msg

def upload_img(gr_img, text_input, history, img_list, img_prefix):
    if gr_img is None:
        return (
            gr.update(),
            gr.update(),
            gr.update(),
            history, img_list, img_prefix
        )

    llm_message = load_and_process_img(gr_img, img_list)
    img_prefix = '<Img><ImageHere></Img>'
    return gr.update(interactive=False), \
        gr.update(interactive=True, placeholder='Type and press Enter'), \
        gr.update(value="Start Chatting", interactive=False), \
        history, \
        img_list, \
        img_prefix

# def upload_imgs(gr_imgs, text_input, history, img_list, img_prefix):
#     if gr_imgs is None:
#         return (
#             gr.update(),
#             gr.update(),
#             gr.update(),
#             history, img_list, img_prefix
#         )
#     for gr_img in gr_imgs:
#         llm_message = load_and_process_img(gr_img, img_list)
#     img_prefix = '<Img>' + '<ImageHere>' * len(gr_imgs) + '</Img>'
#     return gr.update(interactive=False), \
#            gr.update(interactive=True, placeholder='Type and press Enter'), \
#            gr.update(value="Start Chatting", interactive=False), \
#            history, \
#            img_list, \
#            img_prefix

def upload_video(video, text_input, history, img_list, img_prefix):
    # Add your video processing logic here
    if video is None:
        return (
            gr.update(),
            gr.update(),
            gr.update(),
            history, img_list, img_prefix
        )
    llm_message = load_and_process_video(video, img_list)

    img_prefix = f"<Img>{'<ImageHere>' * len(img_list)}</Img>"
    return gr.update(interactive=False), \
           gr.update(interactive=True, placeholder='Type and press Enter'), \
           gr.update(value="Start Chatting", interactive=False), \
           history, \
           img_list, \
           img_prefix

def gradio_ask(user_message, chatbot, img_prefix):
    if len(user_message) == 0:
        return gr.update(interactive=True, placeholder='Input should not be empty!'), chatbot, history
    def get_ask(user_message, img_prefix):
        return img_prefix + user_message
    user_message = get_ask(user_message,img_prefix)
    chatbot = chatbot + [[user_message, None]]
    img_prefix = ""
    return '', chatbot, img_prefix


def gradio_answer(chatbot, history, img_list, do_sample,num_beams, temperature, top_k, top_p):
    generation_config = deepcopy(default_generation_config)
    generation_config.update(
        {
            "do_sample": do_sample=='True',
            "num_beams": num_beams,
            'temperature': temperature,
            "top_k": top_k,
            "top_p": top_p,
        }
    )
    global device
    image_tensor =  torch.stack(img_list).to(device)
    generation_config = GenerationConfig.from_dict(generation_config)
    global args
    if args.cpu_only:
        model.bfloat16()
        response, history = model.chat(query=chatbot[-1][0], history=history, image_tensor=image_tensor.bfloat16(), generation_config=generation_config,verbose=True)
    else:
        with torch.cuda.amp.autocast(enabled=True,dtype=torch.bfloat16):
            response, history = model.chat(query=chatbot[-1][0], history=history, image_tensor=image_tensor.bfloat16(), generation_config=generation_config,verbose=True)
    chatbot[-1][1] = response
    return chatbot, history, img_list

title = """<h1 align="center">Demo of MPPQwen</h1>"""
description = """<h3>This is the demo of MPPQwen, supporting {single-image/multi-image/video} {single-turn/multi-turn} conversation. Upload your images and start chatting! <br> To use
            example questions, click example image, hit upload, and press enter in the chatbox. </h3>"""

from transformers.trainer_utils import set_seed
set_seed(args.seed)
#TODO show examples below

with gr.Blocks() as demo:
    gr.Markdown(title)
    gr.Markdown(description)

    with gr.Row():
        with gr.Column(scale=0.5):
            mode = gr.Dropdown(choices=["Single Image", "Video"], label="Select Mode", value=None)
            image_single = gr.Image(type="filepath", label="Upload Image", value=None)
            video = gr.Video(label="Upload Video", value=None)
            upload_button = gr.Button(value="Upload & Start Chat", interactive=True, variant="primary")
            clear = gr.Button("Restart 🔄")
            do_sample = gr.components.Radio(['True', 'False'],
                            label='do_sample(If False, num_beams, temperature and so on cannot work!)',
                            value='False')

            num_beams = gr.Slider(
                minimum=1,
                maximum=10,
                value=1,
                step=1,
                interactive=True,
                label="beam search numbers)",
            )

            temperature = gr.Slider(
                minimum=0.1,
                maximum=2.0,
                value=1.0,
                step=0.1,
                interactive=True,
                label="Temperature",
            )

            top_k = gr.Slider(
                minimum=0,
                maximum=5,
                value=1,
                step=1,
                interactive=True,
                label="Top_k",
            )

            top_p = gr.Slider(
                minimum=0.1,
                maximum=1.0,
                value=1.0,
                step=0.05,
                interactive=True,
                label="Top_p",
            )

        with gr.Column():
            history = gr.State(value=[])
            img_list = gr.State(value=[])
            chatbot = gr.Chatbot(label='MPPQwen')
            img_prefix = gr.State(value="")
            text_input = gr.Textbox(label='User', placeholder='Please upload your image first', interactive=False)

            # gr.Examples(examples=[
            #     ["examples/minigpt4_image_3.jpg", "描述下这幅图片"],
            # ], inputs=[image, text_input])

    def update_inputs(mode):
        if mode == "Single Image":
            return gr.update(visible=True), gr.update(visible=False)
        elif mode == "Video":
            return gr.update(visible=False), gr.update(visible=True)
        else:
            return gr.update(visible=True), gr.update(visible=True)

    mode.change(update_inputs, inputs=mode, outputs=[image_single, video])
    # print('Mode\t', mode.value)

    upload_button.click(upload_img, [image_single, text_input, history, img_list, img_prefix], [image_single, text_input, upload_button, history, img_list, img_prefix])
    upload_button.click(upload_video, [video, text_input, history, img_list, img_prefix], [video, text_input, upload_button, history, img_list, img_prefix])

    # print(list(map(type,[text_input, chatbot, img_prefix])))
    # print(list(map(type,[chatbot, history, img_list, do_sample, num_beams, temperature, top_k, top_p])))
    text_input.submit(gradio_ask, [text_input, chatbot, img_prefix], [text_input, chatbot, img_prefix]).then(
        gradio_answer, [chatbot, history, img_list, do_sample, num_beams, temperature, top_k, top_p], [chatbot, history, img_list]
    )

    clear.click(gradio_reset, [history, img_list], [chatbot, image_single, video, text_input, upload_button, mode, history, img_list], queue=False)



demo.launch(share=True,inbrowser=True)