Pclanglais's picture
Create app.py
750020e verified
raw
history blame
4.69 kB
import transformers
import re
from transformers import AutoConfig, AutoTokenizer, AutoModel, AutoModelForCausalLM
from vllm import LLM, SamplingParams
import torch
import gradio as gr
import json
import os
import shutil
import requests
import chromadb
import difflib
import pandas as pd
from chromadb.config import Settings
from chromadb.utils import embedding_functions
# Define the device
device = "cuda" if torch.cuda.is_available() else "cpu"
model_checkpoint = "PleIAs/Estienne"
token_classifier = pipeline(
"token-classification", model=editorial_model, aggregation_strategy="simple", device=device
)
tokenizer = AutoTokenizer.from_pretrained(editorial_model, model_max_length=512)
def split_text(text, max_tokens=500):
# Split the text by newline characters
parts = text.split("\n")
chunks = []
current_chunk = ""
for part in parts:
# Add part to current chunk
if current_chunk:
temp_chunk = current_chunk + "\n" + part
else:
temp_chunk = part
# Tokenize the temporary chunk
num_tokens = len(tokenizer.tokenize(temp_chunk))
if num_tokens <= max_tokens:
current_chunk = temp_chunk
else:
if current_chunk:
chunks.append(current_chunk)
current_chunk = part
if current_chunk:
chunks.append(current_chunk)
# If no newlines were found and still exceeding max_tokens, split further
if len(chunks) == 1 and len(tokenizer.tokenize(chunks[0])) > max_tokens:
long_text = chunks[0]
chunks = []
while len(tokenizer.tokenize(long_text)) > max_tokens:
split_point = len(long_text) // 2
while split_point < len(long_text) and not re.match(r'\s', long_text[split_point]):
split_point += 1
# Ensure split_point does not go out of range
if split_point >= len(long_text):
split_point = len(long_text) - 1
chunks.append(long_text[:split_point].strip())
long_text = long_text[split_point:].strip()
if long_text:
chunks.append(long_text)
return chunks
#Curtesy of claude
def generate_html_diff(old_text, new_text):
d = difflib.Differ()
diff = list(d.compare(old_text.split(), new_text.split()))
html_diff = []
for word in diff:
if word.startswith(' '):
html_diff.append(word[2:])
elif word.startswith('+ '):
html_diff.append(f'<span style="background-color: #90EE90;">{word[2:]}</span>')
# We're not adding anything for words that start with '- '
return ' '.join(html_diff)
# Class to encapsulate the Falcon chatbot
class MistralChatBot:
def __init__(self, system_prompt="Le dialogue suivant est une conversation"):
self.system_prompt = system_prompt
def predict(self, user_message):
#We drop the newlines.
editorial_text = re.sub("\n", " ¶ ", user_message)
# Tokenize the prompt and check if it exceeds 500 tokens
num_tokens = len(tokenizer.tokenize(prompt))
if num_tokens > 500:
# Split the prompt into chunks
batch_prompts = split_text(prompt, max_tokens=500)
else:
batch_prompts = [prompt]
out = token_classifier(batch_prompts)
out = "".join(out)
generated_text = '<h2 style="text-align:center">Réponse</h3>\n<div class="generation">' + html_diff + "</div>"
return generated_text
# Create the Falcon chatbot instance
mistral_bot = MistralChatBot()
# Define the Gradio interface
title = "Éditorialisation"
description = "Un outil expérimental d'identification de la structure du texte à partir d'un encoder (Deberta)"
examples = [
[
"Qui peut bénéficier de l'AIP?", # user_message
0.7 # temperature
]
]
additional_inputs=[
gr.Slider(
label="Température",
value=0.2, # Default value
minimum=0.05,
maximum=1.0,
step=0.05,
interactive=True,
info="Des valeurs plus élevées donne plus de créativité, mais aussi d'étrangeté",
),
]
demo = gr.Blocks()
with gr.Blocks(theme='JohnSmith9982/small_and_pretty', css=css) as demo:
gr.HTML("""<h1 style="text-align:center">Correction d'OCR</h1>""")
text_input = gr.Textbox(label="Votre texte.", type="text", lines=1)
text_button = gr.Button("Identifier les structures éditoriales")
text_output = gr.HTML(label="Le texte corrigé")
text_button.click(mistral_bot.predict, inputs=text_input, outputs=[text_output])
if __name__ == "__main__":
demo.queue().launch()