import torch
from transformers import GPT2LMHeadModel, GPT2Tokenizer
import gradio as gr
# Load pre-trained model and tokenizer
model_name = "PleIAs/OCRonos-Vintage"
model = GPT2LMHeadModel.from_pretrained(model_name)
tokenizer = GPT2Tokenizer.from_pretrained(model_name)
# Set the pad token to be the same as the eos token
tokenizer.pad_token = tokenizer.eos_token
# Set the device to GPU if available, otherwise use CPU
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
# Function for generating text
def historical_generation(prompt, max_new_tokens=600):
prompt = f"### Text ###\n{prompt}"
inputs = tokenizer(prompt, return_tensors="pt", padding=True, truncation=True, max_length=1024)
input_ids = inputs["input_ids"].to(device)
attention_mask = inputs["attention_mask"].to(device)
# Generate text
output = model.generate(
input_ids,
attention_mask=attention_mask,
max_new_tokens=max_new_tokens,
pad_token_id=tokenizer.eos_token_id,
top_k=50,
temperature=0.3,
top_p=0.95,
do_sample=True,
repetition_penalty=1.5,
bos_token_id=tokenizer.bos_token_id,
eos_token_id=tokenizer.eos_token_id
)
# Decode the generated text
generated_text = tokenizer.decode(output[0], skip_special_tokens=True)
# Extract text after "### Correction ###"
if "### Correction ###" in generated_text:
generated_text = generated_text.split("### Correction ###")[1].strip()
# Tokenize the generated text
tokens = tokenizer.tokenize(generated_text)
# Create highlighted text output
highlighted_text = []
for token in tokens:
# Clean token and get token type
clean_token = token.replace("Ġ", "")
token_type = tokenizer.convert_ids_to_tokens([tokenizer.convert_tokens_to_ids(token)])[0]
highlighted_text.append((clean_token, token_type))
return highlighted_text
# Tokenizer information display
import os
os.system('python -m spacy download en_core_web_sm')
import spacy
from spacy import displacy
nlp = spacy.load("en_core_web_sm")
def text_analysis(text):
doc = nlp(text)
html = displacy.render(doc, style="dep", page=True)
html = (
"
"
+ html
+ "
"
)
pos_count = {
"char_count": len(text),
"token_count": len(list(doc)),
}
pos_tokens = [(token.text, token.pos_) for token in doc]
return pos_tokens, pos_count, html
# Gradio interface for text analysis
def full_interface(prompt, max_new_tokens):
generated_highlight = historical_generation(prompt, max_new_tokens)
tokens, pos_count, html = text_analysis(prompt)
return generated_highlight, pos_count, html
# Create Gradio interface
iface = gr.Interface(
fn=full_interface,
inputs=[
gr.Textbox(
label="Prompt",
placeholder="Enter a prompt for historical text generation...",
lines=3
),
gr.Slider(
label="Max New Tokens",
minimum=50,
maximum=1000,
step=50,
value=600
)
],
outputs=[
gr.HighlightedText(
label="Generated Historical Text",
combine_adjacent=True,
show_legend=True
),
gr.JSON(label="Tokenizer Info"),
gr.HTML(label="Dependency Parse Visualization")
],
title="Historical Text Generation with OCRonos-Vintage",
description="Generate historical-style text using OCRonos-Vintage and analyze the tokenizer output.",
theme=gr.themes.Base()
)
if __name__ == "__main__":
iface.launch()
# import torch
# from transformers import GPT2LMHeadModel, GPT2Tokenizer
# import gradio as gr
# Load pre-trained model and tokenizer
# model_name = "PleIAs/OCRonos-Vintage"
# model = GPT2LMHeadModel.from_pretrained(model_name)
# tokenizer = GPT2Tokenizer.from_pretrained(model_name)
# Set the pad token to be the same as the eos token
# tokenizer.pad_token = tokenizer.eos_token
# Set the device to GPU if available, otherwise use CPU
# device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# model.to(device)
# def historical_generation(prompt, max_new_tokens=600):
# prompt = f"### Text ###\n{prompt}"
# inputs = tokenizer(prompt, return_tensors="pt", padding=True, truncation=True, max_length=1024)
# input_ids = inputs["input_ids"].to(device)
# attention_mask = inputs["attention_mask"].to(device)
# Generate text
# output = model.generate(
# input_ids,
# attention_mask=attention_mask,
# max_new_tokens=max_new_tokens,
# pad_token_id=tokenizer.eos_token_id,
# top_k=50,
# temperature=0.3,
# top_p=0.95,
# do_sample=True,
# repetition_penalty=1.5,
# bos_token_id=tokenizer.bos_token_id,
# eos_token_id=tokenizer.eos_token_id
# )
# Decode the generated text
# generated_text = tokenizer.decode(output[0], skip_special_tokens=True)
# Remove the prompt from the generated text
# generated_text = generated_text.replace("### Text ###\n", "").strip()
# Tokenize the generated text
# tokens = tokenizer.tokenize(generated_text)
# Create highlighted text output
# highlighted_text = []
# for token in tokens:
# Remove special tokens and get the token type
# clean_token = token.replace("Ġ", "").replace("", "")
# token_type = tokenizer.convert_ids_to_tokens([tokenizer.convert_tokens_to_ids(token)])[0]
# highlighted_text.append((clean_token, token_type))
# return highlighted_text
# Create Gradio interface
# iface = gr.Interface(
# fn=historical_generation,
# inputs=[
# gr.Textbox(
# label="Prompt",
# placeholder="Enter a prompt for historical text generation...",
# lines=3
# ),
# gr.Slider(
# label="Max New Tokens",
# minimum=50,
# maximum=1000,
# step=50,
# value=600
# )
# ],
# outputs=gr.HighlightedText(
# label="Generated Historical Text",
# combine_adjacent=True,
# show_legend=True
# ),
# title="Historical Text Generation with OCRonos-Vintage",
# description="Generate historical-style text using the OCRonos-Vintage model. The output shows token types as highlights.",
# theme=gr.themes.Base()
# )
# if __name__ == "__main__":
# iface.launch()