app.py

import gradio as gr
import spaces
import torch
import os
import uuid
import io
import numpy as np
from PIL import Image
import torchvision.transforms as T
from torchvision.transforms.functional import InterpolationMode
from transformers import AutoModel, AutoTokenizer
from decord import VideoReader, cpu

# =============================================================================
# InternVL 전처리/로딩 코드 (원본 예시에서 발췌)
# =============================================================================
IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)

def build_transform(input_size):
    MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
    transform = T.Compose([
        T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
        T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
        T.ToTensor(),
        T.Normalize(mean=MEAN, std=STD)
    ])
    return transform

def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
    best_ratio_diff = float('inf')
    best_ratio = (1, 1)
    area = width * height
    for ratio in target_ratios:
        target_aspect_ratio = ratio[0] / ratio[1]
        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
        if ratio_diff < best_ratio_diff:
            best_ratio_diff = ratio_diff
            best_ratio = ratio
        elif ratio_diff == best_ratio_diff:
            # 이미지 면적 기준으로 좀 더 큰 쪽 선택
            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
                best_ratio = ratio
    return best_ratio

def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
    orig_width, orig_height = image.size
    aspect_ratio = orig_width / orig_height

    target_ratios = set(
        (i, j) for n in range(min_num, max_num + 1) 
        for i in range(1, n + 1) 
        for j in range(1, n + 1) 
        if i * j <= max_num and i * j >= min_num
    )
    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
    target_aspect_ratio = find_closest_aspect_ratio(
        aspect_ratio, target_ratios, orig_width, orig_height, image_size
    )
    target_width = image_size * target_aspect_ratio[0]
    target_height = image_size * target_aspect_ratio[1]
    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]

    resized_img = image.resize((target_width, target_height))
    processed_images = []
    for i in range(blocks):
        box = (
            (i % (target_width // image_size)) * image_size,
            (i // (target_width // image_size)) * image_size,
            ((i % (target_width // image_size)) + 1) * image_size,
            ((i // (target_width // image_size)) + 1) * image_size
        )
        split_img = resized_img.crop(box)
        processed_images.append(split_img)

    if use_thumbnail and len(processed_images) != 1:
        thumbnail_img = image.resize((image_size, image_size))
        processed_images.append(thumbnail_img)
    return processed_images

def load_image(image_file, input_size=448, max_num=12):
    image = Image.open(image_file).convert('RGB')
    transform = build_transform(input_size=input_size)
    images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
    pixel_values = [transform(img) for img in images]
    pixel_values = torch.stack(pixel_values)
    return pixel_values

def get_index(bound, fps, max_frame, first_idx=0, num_segments=32):
    if bound:
        start, end = bound[0], bound[1]
    else:
        start, end = -100000, 100000
    start_idx = max(first_idx, round(start * fps))
    end_idx = min(round(end * fps), max_frame)
    seg_size = float(end_idx - start_idx) / num_segments
    frame_indices = np.array([
        int(start_idx + (seg_size / 2) + np.round(seg_size * idx))
        for idx in range(num_segments)
    ])
    return frame_indices

def load_video(video_path, bound=None, input_size=448, max_num=1, num_segments=8):
    """
    InternVL 예시 코드 참고: 여러 프레임을 추출하여 dynamic_preprocess 적용.
    여기서는 기본적으로 num_segments=8로 설정.
    """
    vr = VideoReader(video_path, ctx=cpu(0), num_threads=1)
    max_frame = len(vr) - 1
    fps = float(vr.get_avg_fps())

    pixel_values_list, num_patches_list = [], []
    transform = build_transform(input_size=input_size)
    frame_indices = get_index(bound, fps, max_frame, first_idx=0, num_segments=num_segments)

    for frame_index in frame_indices:
        frame = vr[frame_index]
        img = Image.fromarray(frame.asnumpy()).convert('RGB')
        processed_imgs = dynamic_preprocess(img, image_size=input_size, use_thumbnail=True, max_num=max_num)
        tile_values = [transform(tile) for tile in processed_imgs]
        tile_values = torch.stack(tile_values)
        num_patches_list.append(tile_values.shape[0])
        pixel_values_list.append(tile_values)

    # 여러 프레임을 이어 붙여 최종 pixel_values 생성
    pixel_values = torch.cat(pixel_values_list, dim=0)  # (sum(num_patches_list), 3, H, W)
    return pixel_values, num_patches_list


# =============================================================================
# InternVL 모델 로딩
# =============================================================================
MODEL_ID = "OpenGVLab/InternVL2_5-2B"

model = AutoModel.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
    use_flash_attn=True,
    trust_remote_code=True
).eval().cuda()

tokenizer = AutoTokenizer.from_pretrained(
    MODEL_ID,
    trust_remote_code=True,
    use_fast=False
)

# Gradio 상단에 표시할 설명 문구
DESCRIPTION = "[InternVL Demo](https://github.com/OpenGVLab/InternVL) - Using the 2.5B Model"

image_extensions = Image.registered_extensions()
video_extensions = ("avi", "mp4", "mov", "mkv", "flv", "wmv", "mjpeg", "wav", "gif", "webm", "m4v", "3gp")

def identify_and_save_blob(blob_path):
    """
    Qwen 예제 코드와 동일: blob을 열어보고 이미지인지 영상인지 확인 후,
    임시 파일로 저장하여 경로 리턴
    """
    try:
        with open(blob_path, 'rb') as file:
            blob_content = file.read()
            # Try to identify if it's an image
            try:
                Image.open(io.BytesIO(blob_content)).verify()  # Check if it's a valid image
                extension = ".png"  # Default to PNG for saving
                media_type = "image"
            except (IOError, SyntaxError):
                # If it's not a valid image, assume it's a video
                extension = ".mp4"  # Default to MP4 for saving
                media_type = "video"
            
            # Create a unique filename
            filename = f"temp_{uuid.uuid4()}_media{extension}"
            with open(filename, "wb") as f:
                f.write(blob_content)
            return filename, media_type
    except FileNotFoundError:
        raise ValueError(f"The file {blob_path} was not found.")
    except Exception as e:
        raise ValueError(f"An error occurred while processing the file: {e}")

def process_file_upload(file_path):
    """
    파일 업로드 시 이미지/영상 미리보기 혹은 그대로 패스.
    """
    if isinstance(file_path, str):
        if file_path.endswith(tuple([i for i, f in image_extensions.items()])):
            # 이미지를 열어서 preview로 넘김
            return file_path, Image.open(file_path)
        elif file_path.endswith(video_extensions):
            # 영상은 preview를 None으로
            return file_path, None
        else:
            # blob 파일인 경우 처리
            try:
                media_path, media_type = identify_and_save_blob(file_path)
                if media_type == "image":
                    return media_path, Image.open(media_path)
                return media_path, None
            except Exception as e:
                print(e)
                raise ValueError("Unsupported media type. Please upload an image or video.")
    return None, None

@spaces.GPU
def internvl_inference(media_input, text_input=None):
    """
    Qwen 예제의 qwen_inference 대신 InternVL을 이용한 추론 함수.
    - 이미지/영상 파일을 InternVL에서 요구하는 pixel_values로 변환 후
      model.chat() 호출하여 답변 생성.
    """
    if isinstance(media_input, str):  # If it's a filepath
        media_path = media_input

        # 미디어 종류 식별
        if media_path.endswith(tuple([i for i, f in image_extensions.items()])):
            media_type = "image"
        elif media_path.endswith(video_extensions):
            media_type = "video"
        else:
            # blob인지 체크
            try:
                media_path, media_type = identify_and_save_blob(media_input)
            except Exception as e:
                print(e)
                raise ValueError("Unsupported media type. Please upload an image or video.")
    else:
        return "No media input found"

    # 이미지 vs 영상 처리
    if media_type == "image":
        # 단일 이미지만 처리한다고 가정 (멀티-이미지도 확장 가능)
        pixel_values = load_image(media_path, max_num=12)
        pixel_values = pixel_values.to(torch.bfloat16).cuda()  # (N, 3, H, W)
        # InternVL 대화
        question = f"<image>\n{text_input}" if text_input else "<image>\n"
        generation_config = dict(max_new_tokens=1024, do_sample=True)

        response = model.chat(
            tokenizer,
            pixel_values,
            question,
            generation_config
        )
        return response

    elif media_type == "video":
        # 영상: 예시로 첫 8프레임에 대해 처리
        pixel_values, num_patches_list = load_video(
            media_path,
            num_segments=8,
            max_num=1
        )
        pixel_values = pixel_values.to(torch.bfloat16).cuda()
        question_prefix = "".join([f"Frame{i+1}: <image>\n" for i in range(len(num_patches_list))])
        question = question_prefix + (text_input if text_input else "")
        generation_config = dict(max_new_tokens=1024, do_sample=True)

        # 영상에서도 동일한 chat() 함수 사용
        response = model.chat(
            tokenizer,
            pixel_values,
            question,
            generation_config,
            num_patches_list=num_patches_list
        )
        return response

    return "Unsupported media type"

# 간단한 CSS
css = """
  #output {
    height: 500px; 
    overflow: auto; 
    border: 1px solid #ccc; 
  }
"""

# Gradio 데모 구성
with gr.Blocks(css=css) as demo:
    gr.Markdown(DESCRIPTION)

    with gr.Tab(label="Image/Video Input"):
        with gr.Row():
            with gr.Column():
                input_media = gr.File(
                    label="Upload Image or Video", type="filepath"
                )
                preview_image = gr.Image(label="Preview", visible=True)
                text_input = gr.Textbox(label="Question")
                submit_btn = gr.Button(value="Submit")
            with gr.Column():
                output_text = gr.Textbox(label="Output Text")

        input_media.change(
            fn=process_file_upload,
            inputs=[input_media],
            outputs=[input_media, preview_image]
        )
        
        submit_btn.click(
            internvl_inference, 
            [input_media, text_input], 
            [output_text]
        )

demo.launch(debug=True)