HWT / app.py
ankankbhunia's picture
Update app.py
f8ad616 verified
import gradio as gr
from PIL import Image
import numpy as np
from io import BytesIO
import glob
import os
import time
from data.dataset import load_itw_samples, crop_
import torch
import cv2
import os
import numpy as np
from models.model import TRGAN
from params import *
from torch import nn
from data.dataset import get_transform
import pickle
from PIL import Image
import tqdm
import shutil
from datetime import datetime
wellcomingMessage = """
<h1>πŸ’₯ Handwriting Synthesis - Generate text in anyone's handwriting πŸ’₯ </h1>
<p>πŸš€ This app is a demo for the ICCV'21 paper "Handwriting Transformer". Visit our github paper for more information - <a href="https://github.com/ankanbhunia/Handwriting-Transformers" target="_blank">https://github.com/ankanbhunia/Handwriting-Transformers</a></p>
<p>πŸš€ You can either choose from an existing style gallery or upload your own handwriting. If you choose to upload, please ensure that you provide a sufficient number of (~15) cropped handwritten word images for the model to work effectively. The demo is made available for research purposes, and any other use is not intended.</p>
<p>πŸš€ Some examples of cropped handwritten word images can be found <a href="https://huggingface.co./spaces/ankankbhunia/HWT/tree/main/files/example_data/style-1" target="_blank">here</a>.
"""
model_path = 'files/iam_model.pth'
batch_size = 1
print ('(1) Loading model...')
model = TRGAN(batch_size = batch_size)
model.netG.load_state_dict(torch.load(model_path, map_location=torch.device('cpu')) )
print (model_path+' : Model loaded Successfully')
model.eval()
# Define a function to generate an image based on text and images
def generate_image(text,folder, _ch3, images):
# Your image generation logic goes here (replace with your actual implementation)
# For demonstration purposes, we'll just concatenate the uploaded images horizontally.
try:
text_copy = text
if images:
style_log = images
style_inputs, width_length = load_itw_samples(images)
elif folder:
style_log = folder
style_inputs, width_length = load_itw_samples(folder)
else:
return None
# Load images
text = text.replace("\n", "").replace("\t", "")
text_encode = [j.encode() for j in text.split(' ')]
eval_text_encode, eval_len_text = model.netconverter.encode(text_encode)
eval_text_encode = eval_text_encode.to(DEVICE).repeat(batch_size, 1, 1)
input_styles, page_val = model._generate_page(style_inputs.to(DEVICE).clone(), width_length, eval_text_encode, eval_len_text, no_concat = True)
page_val = crop_(page_val[0]*255)
input_styles = crop_(input_styles[0]*255)
max_width = max(page_val.shape[1],input_styles.shape[1])
if page_val.shape[1]!=max_width:
page_val = np.concatenate([page_val, np.ones((page_val.shape[0],max_width-page_val.shape[1]))*255], 1)
else:
input_styles = np.concatenate([input_styles, np.ones((input_styles.shape[0],max_width-input_styles.shape[1]))*255], 1)
upper_pad = np.ones((45,input_styles.shape[1]))*255
input_styles = np.concatenate([upper_pad, input_styles], 0)
page_val = np.concatenate([upper_pad, page_val], 0)
page_val = Image.fromarray(page_val).convert('RGB')
input_styles = Image.fromarray(input_styles).convert('RGB')
current_datetime = datetime.now()
formatted_datetime = current_datetime.strftime("%Y-%m-%d %H:%M:%S")
print (f'{formatted_datetime}: input_string - {text_copy}, style_input - {style_log}\n')
return input_styles, page_val
except:
print ('ERROR! Try again.')
return None, None
input_text_string = "In the quiet hum of everyday life, the dance of existence unfolds. Time, an ever-flowing river, carries the stories of triumph and heartache. Each fleeting moment is a brushstroke on the canvas of our memories."
# Define Gradio Interface
iface = gr.Interface(
fn=generate_image,
inputs=[
gr.Textbox(value = input_text_string, label = "Input text"),
gr.Dropdown(value = "files/example_data/style-30", choices=glob.glob('files/example_data/*'), label="Choose from provided writer styles"),
gr.Markdown("### OR"),
gr.File(label="Upload multiple word images", file_count="multiple")
],
outputs=[#gr.Markdown("## Output"),
gr.Image(type="pil", label="Style Image"),
gr.Image(type="pil", label="Generated Image")],
description = wellcomingMessage,
thumbnail = "Handwriting Synthesis - Mimic anyone's handwriting!",
# examples = [["The sun dipped below the horizon, painting the sky in hues of orange and pink. A gentle breeze rustled the leaves, whispering secrets to the ancient trees. In that fleeting moment, nature's beauty spoke louder than words ever could.", 'files/example_data/style-30', None, None],
# ["Lorem ipsum dolor sit amet, consectetur adipiscing elit. Vestibulum eget lectus eu ex iaculis tristique. Nullam vestibulum, odio vel tincidunt aliquet, sapien quam efficitur risus, nec hendrerit justo quam eget elit. Sed vel augue a lacus facilisis venenatis. ", 'files/example_data/style-31', None, None],
# ["The sun dipped below the horizon, painting the sky in hues of orange and pink. A gentle breeze rustled the leaves, whispering secrets to the ancient trees. In that fleeting moment, nature's beauty spoke louder than words ever could.", None, None, glob.glob('files/example_data/style-1/*')],
# ["Lorem ipsum dolor sit amet, consectetur adipiscing elit. Vestibulum eget lectus eu ex iaculis tristique. Nullam vestibulum, odio vel tincidunt aliquet, sapien quam efficitur risus, nec hendrerit justo quam eget elit. Sed vel augue a lacus facilisis venenatis. ", None, None, glob.glob('files/example_data/style-2/*')]]
)
# Launch the Gradio Interface
iface.launch(debug=True, share=True)