app.py

import gradio as gr
import spaces
import torch
import base64
from PIL import Image, ImageDraw
from io import BytesIO
import re

from deepseek_vl2.models import DeepseekVLV2Processor, DeepseekVLV2ForCausalLM
from deepseek_vl2.utils.io import load_pil_images


from transformers import AutoModelForCausalLM



models = {
    "deepseek-ai/deepseek-vl2-tiny": AutoModelForCausalLM.from_pretrained("deepseek-ai/deepseek-vl2-tiny", trust_remote_code=True),
    #"deepseek-ai/deepseek-vl2-small": AutoModelForCausalLM.from_pretrained("deepseek-ai/deepseek-vl2-small", trust_remote_code=True),
    #"deepseek-ai/deepseek-vl2": AutoModelForCausalLM.from_pretrained("deepseek-ai/deepseek-vl2", trust_remote_code=True)
}

processors = {
    "deepseek-ai/deepseek-vl2-tiny": DeepseekVLV2Processor.from_pretrained("deepseek-ai/deepseek-vl2-tiny",),
    #"deepseek-ai/deepseek-vl2-small": DeepseekVLV2Processor.from_pretrained("deepseek-ai/deepseek-vl2-small",),
    #"deepseek-ai/deepseek-vl2": DeepseekVLV2Processor.from_pretrained("deepseek-ai/deepseek-vl2",),
}


def image_to_base64(image):
    buffered = BytesIO()
    image.save(buffered, format="PNG")
    img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")
    return img_str


def draw_bounding_boxes(image, bounding_boxes, outline_color="red", line_width=2):
    draw = ImageDraw.Draw(image)
    for box in bounding_boxes:
        xmin, ymin, xmax, ymax = box
        draw.rectangle([xmin, ymin, xmax, ymax], outline=outline_color, width=line_width)
    return image


def rescale_bounding_boxes(bounding_boxes, original_width, original_height, scaled_width=1000, scaled_height=1000):
    x_scale = original_width / scaled_width
    y_scale = original_height / scaled_height
    rescaled_boxes = []
    for box in bounding_boxes:
        xmin, ymin, xmax, ymax = box
        rescaled_box = [
            xmin * x_scale,
            ymin * y_scale,
            xmax * x_scale,
            ymax * y_scale
        ]
        rescaled_boxes.append(rescaled_box)
    return rescaled_boxes


def deepseek(image, text_input, model_id):
    # specify the path to the model
    vl_chat_processor: DeepseekVLV2Processor = processors[model_id]
    tokenizer = vl_chat_processor.tokenizer

    vl_gpt: DeepseekVLV2ForCausalLM = models[model_id]
    vl_gpt = vl_gpt.to(torch.bfloat16).cuda().eval()

    ## single image conversation example
    conversation = [
        {
            "role": "<|User|>",
            "content": f"<image><|ref|>{text_input}<|/ref|>.",
            "images": ["./images/visual_grounding_1.jpeg"],
        },
        {"role": "<|Assistant|>", "content": ""},
    ]

    # load images and prepare for inputs
    #pil_images = load_pil_images(conversation)
    prepare_inputs = vl_chat_processor(
        conversations=conversation,
        images=[image],
        force_batchify=True,
        system_prompt=""
    ).to(vl_gpt.device)

    # run image encoder to get the image embeddings
    inputs_embeds = vl_gpt.prepare_inputs_embeds(**prepare_inputs)

    # run the model to get the response
    outputs = vl_gpt.language.generate(
        inputs_embeds=inputs_embeds,
        attention_mask=prepare_inputs.attention_mask,
        pad_token_id=tokenizer.eos_token_id,
        bos_token_id=tokenizer.bos_token_id,
        eos_token_id=tokenizer.eos_token_id,
        max_new_tokens=512,
        do_sample=False,
        use_cache=True
    )

    answer = tokenizer.decode(outputs[0].cpu().tolist(), skip_special_tokens=False)

    print(f"{prepare_inputs['sft_format'][0]}", answer)
    det_pattern = r"<\|det\|>\[\[(.+)]]<\|\/det\|>"

    det_match = re.search(det_pattern, answer)
    if det_match is None:
        return text_input, [], image
    
    det_content = det_match.group(1)
    bbox = [int(v.strip()) for v in det_content.split(",")]

    scaled_boxes = rescale_bounding_boxes([bbox], image.width, image.height)
    return answer, scaled_boxes, draw_bounding_boxes(image, scaled_boxes)


@spaces.GPU
def run_example(image, text_input, model_id="eepseek-ai/deepseek-vl2-tiny"):
    return deepseek(image, text_input, model_id)
    
css = """
  #output {
    height: 500px; 
    overflow: auto; 
    border: 1px solid #ccc; 
  }
"""
with gr.Blocks(css=css) as demo:
    gr.Markdown(
    """
    # Demo for Deepseek-VL2: Mixture-of-Experts Vision-Language Models for Advanced Multimodal Understanding
    """)
    with gr.Row():
        with gr.Column():
            input_img = gr.Image(label="Input Image", type="pil")
            model_selector = gr.Dropdown(choices=list(models.keys()), label="Model", value="deepseek-ai/deepseek-vl2-tiny")
            text_input = gr.Textbox(label="User Prompt")
            submit_btn = gr.Button(value="Submit")
        with gr.Column():
            model_output_text = gr.Textbox(label="Model Output Text")
            model_output_box = gr.Textbox(label="Model Output Box")
            annotated_image = gr.Image(label="Annotated Image")

    gr.Examples(
        examples=[
            ["assets/web_6f93090a-81f6-489e-bb35-1a2838b18c01.png", "select search textfield"],
            ["assets/web_6f93090a-81f6-489e-bb35-1a2838b18c01.png", "switch to discussions"],
        ],
        inputs=[input_img, text_input],
        outputs=[model_output_text, model_output_box, annotated_image],
        fn=run_example,
        cache_examples=True,
        label="Try examples"
    )

    submit_btn.click(run_example, [input_img, text_input, model_selector], [model_output_text, model_output_box, annotated_image])

demo.launch(debug=True)