app.py

import os
import random
import uuid
import json
import time
import asyncio
from threading import Thread

import gradio as gr
import spaces
import torch
import numpy as np
from PIL import Image, ImageOps
import cv2
import pymupdf
import io

from transformers import (
    Qwen2VLForConditionalGeneration,
    VisionEncoderDecoderModel,
    AutoModelForVision2Seq,
    AutoProcessor,
    TextIteratorStreamer,
)
from transformers.image_utils import load_image

from docling_core.types.doc import DoclingDocument, DocTagsDocument

import re
import ast
import html

# Constants for text generation
MAX_MAX_NEW_TOKENS = 2048
DEFAULT_MAX_NEW_TOKENS = 1024
MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

# Global variables for Dolphin model
model_k = None
processor_k = None
tokenizer_k = None

# Load models
def initialize_models():
    global model_k, processor_k, tokenizer_k
    # Load olmOCR-7B-0225-preview
    MODEL_ID_M = "allenai/olmOCR-7B-0225-preview"
    processor_m = AutoProcessor.from_pretrained(MODEL_ID_M, trust_remote_code=True)
    model_m = Qwen2VLForConditionalGeneration.from_pretrained(
        MODEL_ID_M, trust_remote_code=True, torch_dtype=torch.float16
    ).to(device).eval()

    # Load ByteDance's Dolphin
    MODEL_ID_K = "ByteDance/Dolphin"
    processor_k = AutoProcessor.from_pretrained(MODEL_ID_K, trust_remote_code=True)
    if model_k is None:
        model_k = VisionEncoderDecoderModel.from_pretrained(
            MODEL_ID_K, trust_remote_code=True, torch_dtype=torch.float16
        ).to(device).eval()
        tokenizer_k = processor_k.tokenizer

    # Load SmolDocling-256M-preview
    MODEL_ID_X = "ds4sd/SmolDocling-256M-preview"
    processor_x = AutoProcessor.from_pretrained(MODEL_ID_X, trust_remote_code=True)
    model_x = AutoModelForVision2Seq.from_pretrained(
        MODEL_ID_X, trust_remote_code=True, torch_dtype=torch.float16
    ).to(device).eval()

    return processor_m, model_m, processor_x, model_x

processor_m, model_m, processor_x, model_x = initialize_models()

# Preprocessing functions for SmolDocling-256M
def add_random_padding(image, min_percent=0.1, max_percent=0.10):
    """Add random padding to an image based on its size."""
    image = image.convert("RGB")
    width, height = image.size
    pad_w_percent = random.uniform(min_percent, max_percent)
    pad_h_percent = random.uniform(min_percent, max_percent)
    pad_w = int(width * pad_w_percent)
    pad_h = int(height * pad_h_percent)
    corner_pixel = image.getpixel((0, 0))
    padded_image = ImageOps.expand(image, border=(pad_w, pad_h, pad_w, pad_h), fill=corner_pixel)
    return padded_image

def normalize_values(text, target_max=500):
    """Normalize numerical values in text to a target maximum."""
    def normalize_list(values):
        max_value = max(values) if values else 1
        return [round((v / max_value) * target_max) for v in values]

    def process_match(match):
        num_list = ast.literal_eval(match.group(0))
        normalized = normalize_list(num_list)
        return "".join([f"<loc_{num}>" for num in normalized])

    pattern = r"\[([\d\.\s,]+)\]"
    normalized_text = re.sub(pattern, process_match, text)
    return normalized_text

def downsample_video(video_path):
    """Downsample a video to evenly spaced frames, returning PIL images with timestamps."""
    vidcap = cv2.VideoCapture(video_path)
    total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))
    fps = vidcap.get(cv2.CAP_PROP_FPS)
    frames = []
    frame_indices = np.linspace(0, total_frames - 1, 10, dtype=int)
    for i in frame_indices:
        vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)
        success, image = vidcap.read()
        if success:
            image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
            pil_image = Image.fromarray(image)
            timestamp = round(i / fps, 2)
            frames.append((pil_image, timestamp))
    vidcap.release()
    return frames

# Dolphin-specific functions
@spaces.GPU
def model_chat(prompt, image, is_batch=False):
    """Use Dolphin model for inference, supporting both single and batch processing."""
    global model_k, processor_k, tokenizer_k
    if model_k is None:
        initialize_models()

    if not is_batch:
        images = [image]
        prompts = [prompt]
    else:
        images = image
        prompts = prompt if isinstance(prompt, list) else [prompt] * len(images)

    inputs = processor_k(images, return_tensors="pt", padding=True).to(device)
    pixel_values = inputs.pixel_values.half()

    prompts = [f"<s>{p} <Answer/>" for p in prompts]
    prompt_inputs = tokenizer_k(
        prompts, add_special_tokens=False, return_tensors="pt", padding=True
    ).to(device)

    outputs = model_k.generate(
        pixel_values=pixel_values,
        decoder_input_ids=prompt_inputs.input_ids,
        decoder_attention_mask=prompt_inputs.attention_mask,
        min_length=1,
        max_length=4096,
        pad_token_id=tokenizer_k.pad_token_id,
        eos_token_id=tokenizer_k.eos_token_id,
        use_cache=True,
        bad_words_ids=[[tokenizer_k.unk_token_id]],
        return_dict_in_generate=True,
        do_sample=False,
        num_beams=1,
        repetition_penalty=1.1
    )
    sequences = tokenizer_k.batch_decode(outputs.sequences, skip_special_tokens=False)
    
    results = []
    for i, sequence in enumerate(sequences):
        cleaned = sequence.replace(prompts[i], "").replace("<pad>", "").replace("</s>", "").strip()
        results.append(cleaned)

    return results[0] if not is_batch else results

@spaces.GPU
def process_element_batch(elements, prompt, max_batch_size=16):
    """Process a batch of elements with the same prompt."""
    results = []
    batch_size = min(len(elements), max_batch_size)

    for i in range(0, len(elements), batch_size):
        batch_elements = elements[i:i + batch_size]
        crops_list = [elem["crop"] for elem in batch_elements]
        prompts_list = [prompt] * len(crops_list)
        
        batch_results = model_chat(prompts_list, crops_list, is_batch=True)
        
        for j, result in enumerate(batch_results):
            elem = batch_elements[j]
            results.append({
                "label": elem["label"],
                "bbox": elem["bbox"],
                "text": result.strip(),
                "reading_order": elem["reading_order"],
            })
    
    return results

def process_elements(layout_results, image):
    """Parse layout results and extract elements from the image."""
    try:
        elements = ast.literal_eval(layout_results)
    except:
        elements = []
    
    text_elements = []
    table_elements = []
    figure_results = []
    reading_order = 0
    
    for bbox, label in elements:
        try:
            x1, y1, x2, y2 = map(int, bbox)
            cropped = image.crop((x1, y1, x2, y2))
            if cropped.size[0] > 0 and cropped.size[1] > 0:
                element_info = {
                    "crop": cropped,
                    "label": label,
                    "bbox": [x1, y1, x2, y2],
                    "reading_order": reading_order,
                }
                if label == "text":
                    text_elements.append(element_info)
                elif label == "table":
                    table_elements.append(element_info)
                elif label == "figure":
                    figure_results.append({
                        "label": label,
                        "bbox": [x1, y1, x2, y2],
                        "text": "[Figure]",
                        "reading_order": reading_order
                    })
            reading_order += 1
        except Exception as e:
            print(f"Error processing element: {e}")
            continue
    
    recognition_results = figure_results.copy()
    
    if text_elements:
        text_results = process_element_batch(text_elements, "Read text in the image.")
        recognition_results.extend(text_results)
    
    if table_elements:
        table_results = process_element_batch(table_elements, "Parse the table in the image.")
        recognition_results.extend(table_results)
    
    recognition_results.sort(key=lambda x: x["reading_order"])
    return recognition_results

def generate_markdown(recognition_results):
    """Generate markdown from extracted elements."""
    markdown = ""
    for element in recognition_results:
        if element["label"] == "text":
            markdown += f"{element['text']}\n\n"
        elif element["label"] == "table":
            markdown += f"**Table:**\n{element['text']}\n\n"
        elif element["label"] == "figure":
            markdown += f"{element['text']}\n\n"
    return markdown.strip()

def convert_to_image(image):
    """Convert uploaded file to PIL Image, handling PDFs by extracting the first page."""
    if isinstance(image, str):  # File path from Gradio
        if image.lower().endswith('.pdf'):
            doc = pymupdf.open(image)
            page = doc[0]
            pix = page.get_pixmap()
            img_data = pix.tobytes("png")
            pil_image = Image.open(io.BytesIO(img_data)).convert("RGB")
            doc.close()
            return pil_image
        else:
            return Image.open(image).convert("RGB")
    elif isinstance(image, Image.Image):  # Already a PIL Image
        return image.convert("RGB")
    return None

def process_image_with_dolphin(image):
    """Process a single image with Dolphin model."""
    pil_image = convert_to_image(image)
    if pil_image is None:
        return "Error: Unable to process the uploaded file."
    layout_output = model_chat("Parse the reading order of this document.", pil_image)
    elements = process_elements(layout_output, pil_image)
    markdown_content = generate_markdown(elements)
    return markdown_content

@spaces.GPU
def generate_image(model_name: str, text: str, image: Image.Image,
                   max_new_tokens: int = 1024, temperature: float = 0.6,
                   top_p: float = 0.9, top_k: int = 50, repetition_penalty: float = 1.2):
    """Generate responses for image input using the selected model."""
    if model_name == "ByteDance-s-Dolphin":
        if image is None:
            yield "Please upload an image or PDF (first page will be processed)."
            return
        markdown_content = process_image_with_dolphin(image)
        yield markdown_content
    else:
        if model_name == "olmOCR-7B-0225-preview":
            processor = processor_m
            model = model_m
        elif model_name == "SmolDocling-256M-preview":
            processor = processor_x
            model = model_x
        else:
            yield "Invalid model selected."
            return

        if image is None:
            yield "Please upload an image."
            return

        images = [convert_to_image(image)]
        if images[0] is None:
            yield "Error: Unable to process the uploaded file."
            return

        if model_name == "SmolDocling-256M-preview":
            if "OTSL" in text or "code" in text:
                images = [add_random_padding(img) for img in images]
            if "OCR at text at" in text or "Identify element" in text or "formula" in text:
                text = normalize_values(text, target_max=500)

        messages = [
            {
                "role": "user",
                "content": [{"type": "image"} for _ in images] + [
                    {"type": "text", "text": text}
                ]
            }
        ]
        prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
        inputs = processor(text=prompt, images=images, return_tensors="pt").to(device)

        streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
        generation_kwargs = {
            **inputs,
            "streamer": streamer,
            "max_new_tokens": max_new_tokens,
            "temperature": temperature,
            "top_p": top_p,
            "top_k": top_k,
            "repetition_penalty": repetition_penalty,
        }
        thread = Thread(target=model.generate, kwargs=generation_kwargs)
        thread.start()

        buffer = ""
        full_output = ""
        for new_text in streamer:
            full_output += new_text
            buffer += new_text.replace("<|im_end|>", "")
            yield buffer

        if model_name == "SmolDocling-256M-preview":
            cleaned_output = full_output.replace("<end_of_utterance>", "").strip()
            if any(tag in cleaned_output for tag in ["<doctag>", "<otsl>", "<code>", "<chart>", "<formula>"]):
                if "<chart>" in cleaned_output:
                    cleaned_output = cleaned_output.replace("<chart>", "<otsl>").replace("</chart>", "</otsl>")
                    cleaned_output = re.sub(r'(<loc_500>)(?!.*<loc_500>)<[^>]+>', r'\1', cleaned_output)
                doctags_doc = DocTagsDocument.from_doctags_and_image_pairs([cleaned_output], images)
                doc = DoclingDocument.load_from_doctags(doctags_doc, document_name="Document")
                markdown_output = doc.export_to_markdown()
                yield f"**MD Output:**\n\n{markdown_output}"
            else:
                yield cleaned_output

@spaces.GPU
def generate_video(model_name: str, text: str, video_path: str,
                   max_new_tokens: int = 1024, temperature: float = 0.6,
                   top_p: float = 0.9, top_k: int = 50, repetition_penalty: float = 1.2):
    """Generate responses for video input using the selected model."""
    if model_name == "ByteDance-s-Dolphin":
        if video_path is None:
            yield "Please upload a video."
            return
        frames = downsample_video(video_path)
        markdown_contents = []
        for idx, (frame, _) in enumerate(frames):
            markdown_content = process_image_with_dolphin(frame)
            markdown_contents.append(f"**Frame {idx + 1}:**\n{markdown_content}")
        combined_markdown = "\n\n---\n\n".join(markdown_contents)
        yield combined_markdown
    else:
        if model_name == "olmOCR-7B-0225-preview":
            processor = processor_m
            model = model_m
        elif model_name == "SmolDocling-256M-preview":
            processor = processor_x
            model = model_x
        else:
            yield "Invalid model selected."
            return

        if video_path is None:
            yield "Please upload a video."
            return

        frames = downsample_video(video_path)
        images = [frame for frame, _ in frames]

        if model_name == "SmolDocling-256M-preview":
            if "OTSL" in text or "code" in text:
                images = [add_random_padding(img) for img in images]
            if "OCR at text at" in text or "Identify element" in text or "formula" in text:
                text = normalize_values(text, target_max=500)

        messages = [
            {
                "role": "user",
                "content": [{"type": "image"} for _ in images] + [
                    {"type": "text", "text": text}
                ]
            }
        ]
        prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
        inputs = processor(text=prompt, images=images, return_tensors="pt").to(device)

        streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
        generation_kwargs = {
            **inputs,
            "streamer": streamer,
            "max_new_tokens": max_new_tokens,
            "temperature": temperature,
            "top_p": top_p,
            "top_k": top_k,
            "repetition_penalty": repetition_penalty,
        }
        thread = Thread(target=model.generate, kwargs=generation_kwargs)
        thread.start()

        buffer = ""
        full_output = ""
        for new_text in streamer:
            full_output += new_text
            buffer += new_text.replace("<|im_end|>", "")
            yield buffer

        if model_name == "SmolDocling-256M-preview":
            cleaned_output = full_output.replace("<end_of_utterance>", "").strip()
            if any(tag in cleaned_output for tag in ["<doctag>", "<otsl>", "<code>", "<chart>", "<formula>"]):
                if "<chart>" in cleaned_output:
                    cleaned_output = cleaned_output.replace("<chart>", "<otsl>").replace("</chart>", "</otsl>")
                    cleaned_output = re.sub(r'(<loc_500>)(?!.*<loc_500>)<[^>]+>', r'\1', cleaned_output)
                doctags_doc = DocTagsDocument.from_doctags_and_image_pairs([cleaned_output], images)
                doc = DoclingDocument.load_from_doctags(doctags_doc, document_name="Document")
                markdown_output = doc.export_to_markdown()
                yield f"**MD Output:**\n\n{markdown_output}"
            else:
                yield cleaned_output

# Define examples
image_examples = [
    ["Convert this page to docling", "images/1.png"],
    ["OCR the image", "images/2.jpg"],
    ["Convert this page to docling", "images/3.png"],
]

video_examples = [
    ["Explain the ad in detail", "example/1.mp4"],
    ["Identify the main actions in the coca cola ad...", "example/2.mp4"]
]

css = """
.submit-btn {
    background-color: #2980b9 !important;
    color: white !important;
}
.submit-btn:hover {
    background-color: #3498db !important;
}
"""

# Create Gradio Interface
with gr.Blocks(css=css, theme="bethecloud/storj_theme") as demo:
    gr.Markdown("# **[Docling-VLMs](https://huggingface.co/collections/prithivMLmods/multimodal-implementations-67c9982ea04b39f0608badb0)**")
    gr.Markdown("**Note:** For Dolphin model, the text query is ignored, and PDFs are processed by parsing the first page.")
    with gr.Row():
        with gr.Column():
            with gr.Tabs():
                with gr.TabItem("Image Inference"):
                    image_query = gr.Textbox(label="Query Input", placeholder="Enter your query here...")
                    image_upload = gr.Image(type="pil", label="Image or PDF")
                    image_submit = gr.Button("Submit", elem_classes="submit-btn")
                    gr.Examples(examples=image_examples, inputs=[image_query, image_upload])
                with gr.TabItem("Video Inference"):
                    video_query = gr.Textbox(label="Query Input", placeholder="Enter your query here...")
                    video_upload = gr.Video(label="Video")
                    video_submit = gr.Button("Submit", elem_classes="submit-btn")
                    gr.Examples(examples=video_examples, inputs=[video_query, video_upload])
            with gr.Accordion("Advanced options", open=False):
                max_new_tokens = gr.Slider(label="Max new tokens", minimum=1, maximum=MAX_MAX_NEW_TOKENS, step=1, value=DEFAULT_MAX_NEW_TOKENS)
                temperature = gr.Slider(label="Temperature", minimum=0.1, maximum=4.0, step=0.1, value=0.6)
                top_p = gr.Slider(label="Top-p (nucleus sampling)", minimum=0.05, maximum=1.0, step=0.05, value=0.9)
                top_k = gr.Slider(label="Top-k", minimum=1, maximum=1000, step=1, value=50)
                repetition_penalty = gr.Slider(label="Repetition penalty", minimum=1.0, maximum=2.0, step=0.05, value=1.2)
        with gr.Column():
            output = gr.Textbox(label="Output", interactive=False, lines=3, scale=2)
            model_choice = gr.Radio(
                choices=["olmOCR-7B-0225-preview", "SmolDocling-256M-preview", "ByteDance-s-Dolphin"],
                label="Select Model",
                value="olmOCR-7B-0225-preview"
            )
            
    image_submit.click(
        fn=generate_image,
        inputs=[model_choice, image_query, image_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],
        outputs=output
    )
    video_submit.click(
        fn=generate_video,
        inputs=[model_choice, video_query, video_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],
        outputs=output
    )

if __name__ == "__main__":
    demo.queue(max_size=30).launch(share=True, mcp_server=True, ssr_mode=False, show_error=True)