Image2Paragraph / app_w_sam.py
Awiny's picture
updata sam version
b510b75
raw
history blame
5.91 kB
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('--image_caption_device', choices=['cuda', 'cpu'], default='cpu', help='Select the device: cuda or cpu, gpu memory larger than 14G is recommended')
parser.add_argument('--dense_caption_device', choices=['cuda', 'cpu'], default='cpu', help='Select the device: cuda or cpu, < 6G GPU is not recommended>')
parser.add_argument('--semantic_segment_device', choices=['cuda', 'cpu'], default='cpu', help='Select the device: cuda or cpu, gpu memory larger than 14G is recommended')
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()