app.py

import gradio as gr
import cv2
import numpy as np
from PIL import Image
import base64
from io import BytesIO
from models.image_text_transformation import ImageTextTransformation
import argparse
import torch

parser = argparse.ArgumentParser()
parser.add_argument('--gpt_version', choices=['gpt-3.5-turbo', 'gpt4'], default='gpt-3.5-turbo')
parser.add_argument('--image_caption', action='store_true', dest='image_caption', default=True, help='Set this flag to True if you want to use BLIP2 Image Caption')
parser.add_argument('--dense_caption', action='store_true', dest='dense_caption', default=True, help='Set this flag to True if you want to use Dense Caption')
parser.add_argument('--semantic_segment', action='store_true', dest='semantic_segment', default=True, help='Set this flag to True if you want to use semantic segmentation')
parser.add_argument('--sam_arch', choices=['vit_b', 'vit_l', 'vit_h'], dest='sam_arch', default='vit_b', help='vit_b is the default model (fast but not accurate), vit_l and vit_h are larger models')
parser.add_argument('--captioner_base_model', choices=['blip', 'blip2'], dest='captioner_base_model', default='blip', help='blip2 requires 15G GPU memory, blip requires 6G GPU memory')
parser.add_argument('--region_classify_model', choices=['ssa', 'edit_anything'], dest='region_classify_model', default='edit_anything', help='Select the region classification model: edit anything is ten times faster than ssa, but less accurate.')
parser.add_argument('--image_caption_device', choices=['cuda', 'cpu'], default='cuda', help='Select the device: cuda or cpu, gpu memory larger than 14G is recommended')
parser.add_argument('--dense_caption_device', choices=['cuda', 'cpu'], default='cuda', help='Select the device: cuda or cpu, < 6G GPU is not recommended>')
parser.add_argument('--semantic_segment_device', choices=['cuda', 'cpu'], default='cuda', help='Select the device: cuda or cpu, gpu memory larger than 14G is recommended. Make sue this model and image_caption model on same device.')
parser.add_argument('--contolnet_device', choices=['cuda', 'cpu'], default='cpu', help='Select the device: cuda or cpu, <6G GPU is not recommended>')

args = parser.parse_args()

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

if device == "cuda":
    args.image_caption_device = "cpu"
    args.dense_caption_device = "cuda"
    args.semantic_segment_device = "cuda"
    args.contolnet_device = "cuda"
else:
    args.image_caption_device = "cpu"
    args.dense_caption_device = "cpu"
    args.semantic_segment_device = "cpu"
    args.contolnet_device = "cpu"

def pil_image_to_base64(image):
    buffered = BytesIO()
    image.save(buffered, format="JPEG")
    img_str = base64.b64encode(buffered.getvalue()).decode()
    return img_str

def add_logo():
    with open("examples/logo.png", "rb") as f:
        logo_base64 = base64.b64encode(f.read()).decode()
    return logo_base64

def process_image(image_src, options=None, processor=None):
    print(options)
    if options is None:
        options = []
    processor.args.semantic_segment = "Semantic Segment" in options
    image_generation_status = "Image Generation" in options
    image_caption, dense_caption, region_semantic, gen_text = processor.image_to_text(image_src)
    if image_generation_status:
        gen_image = processor.text_to_image(gen_text)
        gen_image_str = pil_image_to_base64(gen_image)
    # Combine the outputs into a single HTML output
    custom_output = f'''
    <h2>Image->Text:</h2>
    <div style="display: flex; flex-wrap: wrap;">
        <div style="flex: 1;">
            <h3>Image Caption</h3>
            <p>{image_caption}</p>
        </div>
        <div style="flex: 1;">
            <h3>Dense Caption</h3>
            <p>{dense_caption}</p>
        </div>
        <div style="flex: 1;">
            <h3>Region Semantic</h3>
            <p>{region_semantic}</p>
        </div>
    </div>
    <div style="display: flex; flex-wrap: wrap;">
        <div style="flex: 1;">
            <h3>GPT4 Reasoning:</h3>
            <p>{gen_text}</p>
        </div>
    </div>
    '''
    if image_generation_status:
        custom_output += f'''
        <h2>Text->Image:</h2>
        <div style="display: flex; flex-wrap: wrap;">
            <div style="flex: 1;">
                <h3>Generated Image</h3>
                <img src="data:image/jpeg;base64,{gen_image_str}" width="400" style="vertical-align: middle;">
            </div>
        </div>
        '''
    return custom_output

processor = ImageTextTransformation(args)

# Create Gradio input and output components
image_input = gr.inputs.Image(type='filepath', label="Input Image")
semantic_segment_checkbox = gr.inputs.Checkbox(label="Semantic Segment", default=False)
image_generation_checkbox = gr.inputs.Checkbox(label="Image Generation", default=False)


extra_title = r'![vistors](https://visitor-badge.glitch.me/badge?page_id=fingerrec.Image2Paragraph)' + '\n' + \
              r'[![Duplicate this Space](https://huggingface.co/datasets/huggingface/badges/raw/main/duplicate-this-space-md-dark.svg)](https://huggingface.co/spaces/Awiny/Image2Paragraph?duplicate=true)' + '\n\n'



logo_base64 = add_logo()
# Create the title with the logo
title_with_logo = \
    f'<img src="data:image/jpeg;base64,{logo_base64}" width="400" style="vertical-align: middle;"> Understanding Image with Text'

examples = [
    ["examples/test_4.jpg"],
]

# Create Gradio interface
interface = gr.Interface(
    fn=lambda image, options: process_image(image, options, processor),
    inputs=[image_input,        
            gr.CheckboxGroup(
            label="Options",
            choices=["Image Generation", "Semantic Segment"],
            ),
            ],
    outputs=gr.outputs.HTML(),
    title=title_with_logo,
    examples=examples,
    description=extra_title +"""
    Image.txt. This code support image to text transformation. Then the generated text can do retrieval, question answering et al to conduct zero-shot.
    \n Github: https://github.com/showlab/Image2Paragraph
    \n Twitter: https://twitter.com/awinyimgprocess/status/1646225454599372800?s=46&t=HvOe9T2n35iFuCHP5aIHpQ
    \n Since GPU is expensive, we use CPU for demo and not include semantic segment anything. Run code local with gpu or google colab we provided for fast speed.
    \n Ttext2image model is controlnet ( very slow in cpu(~2m)), which used canny edge as reference.
    \n To speed up, we generate image with small size 384, run the code local for high-quality sample.
    """
)

# Launch the interface
interface.launch()