Maram-almasary's picture
Update app.py
3f0629a verified
import gradio as gr
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline
import matplotlib.pyplot as plt
import networkx as nx
import io
from PIL import Image
import torch
import os
print("Installation complete. Loading models...")
# Load models once at startup
model_name = "csebuetnlp/mT5_multilingual_XLSum"
tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=False)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
# If you have a GPU, use it
device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Using device: {device}")
model = model.to(device)
# Load question generator once
question_generator = pipeline(
"text2text-generation",
model="valhalla/t5-small-e2e-qg",
device=device if device == "cuda" else -1
)
def summarize_text(text, src_lang):
inputs = tokenizer(text, return_tensors="pt", max_length=512, truncation=True).to(device)
# Use more efficient generation parameters
summary_ids = model.generate(
inputs["input_ids"],
max_length=150,
min_length=30,
length_penalty=2.0,
num_beams=4,
early_stopping=True
)
summary = tokenizer.decode(summary_ids[0], skip_special_tokens=True)
return summary
def generate_questions(summary):
# Generate questions one at a time with beam search
questions = []
for _ in range(3): # Generate 3 questions
result = question_generator(
summary,
max_length=64,
num_beams=4,
do_sample=True,
top_k=30,
top_p=0.95,
temperature=0.7
)
questions.append(result[0]['generated_text'])
# Remove duplicates
questions = list(set(questions))
return questions
def generate_concept_map(summary, questions):
# Use NetworkX and matplotlib for rendering
G = nx.DiGraph()
# Add summary as central node
summary_short = summary[:50] + "..." if len(summary) > 50 else summary
G.add_node("summary", label=summary_short)
# Add question nodes and edges
for i, question in enumerate(questions):
q_short = question[:30] + "..." if len(question) > 30 else question
node_id = f"Q{i}"
G.add_node(node_id, label=q_short)
G.add_edge("summary", node_id)
# Create the plot directly in memory
plt.figure(figsize=(10, 8))
pos = nx.spring_layout(G, seed=42) # Fixed seed for consistent layout
nx.draw(G, pos, with_labels=False, node_color='skyblue',
node_size=1500, arrows=True, connectionstyle='arc3,rad=0.1',
edgecolors='black', linewidths=1)
# Add labels with better font handling
# FIX: Removed 'wrap' parameter which is not supported in this version of NetworkX
labels = nx.get_node_attributes(G, 'label')
nx.draw_networkx_labels(G, pos, labels=labels, font_size=9,
font_family='sans-serif')
# Save to memory buffer
buf = io.BytesIO()
plt.savefig(buf, format='png', dpi=100, bbox_inches='tight')
buf.seek(0)
plt.close()
return Image.open(buf)
def analyze_text(text, lang):
if not text.strip():
return "Please enter some text.", "No questions generated.", None
# Process the text
try:
print("Generating summary...")
summary = summarize_text(text, lang)
print("Generating questions...")
questions = generate_questions(summary)
print("Creating concept map...")
concept_map_image = generate_concept_map(summary, questions)
# Format questions as a list
questions_text = "\n".join([f"- {q}" for q in questions])
return summary, questions_text, concept_map_image
except Exception as e:
import traceback
print(f"Error processing text: {str(e)}")
print(traceback.format_exc())
return f"Error processing text: {str(e)}", "", None
# Alternative simpler concept map function in case the above still has issues
def generate_simple_concept_map(summary, questions):
"""Fallback concept map generator with minimal dependencies"""
plt.figure(figsize=(10, 8))
# Create a simple radial layout
n_questions = len(questions)
# Draw the central node (summary)
plt.scatter([0], [0], s=1000, color='skyblue', edgecolors='black')
plt.text(0, 0, summary[:50] + "..." if len(summary) > 50 else summary,
ha='center', va='center', fontsize=9)
# Draw the question nodes in a circle around the summary
radius = 5
for i, question in enumerate(questions):
angle = 2 * 3.14159 * i / max(n_questions, 1)
x = radius * 0.8 * -1 * (max(n_questions, 1) - 1) * ((i / max(n_questions - 1, 1)) - 0.5)
y = radius * 0.6 * (i % 2 * 2 - 1)
# Draw node
plt.scatter([x], [y], s=800, color='lightgreen', edgecolors='black')
# Draw edge from summary to question
plt.plot([0, x], [0, y], 'k-', alpha=0.6)
# Add question text
plt.text(x, y, question[:30] + "..." if len(question) > 30 else question,
ha='center', va='center', fontsize=8)
plt.axis('equal')
plt.axis('off')
# Save to memory buffer
buf = io.BytesIO()
plt.savefig(buf, format='png', dpi=100, bbox_inches='tight')
buf.seek(0)
plt.close()
return Image.open(buf)
examples = [
["الذكاء الاصطناعي هو فرع من علوم الكمبيوتر يهدف إلى إنشاء آلات ذكية تعمل وتتفاعل مثل البشر. بعض الأنشطة التي صممت أجهزة الكمبيوتر الذكية للقيام بها تشمل: التعرف على الصوت، التعلم، التخطيط، وحل المشاكل.", "ar"],
["Artificial intelligence is a branch of computer science that aims to create intelligent machines that work and react like humans. Some of the activities computers with artificial intelligence are designed for include: Speech recognition, learning, planning, and problem-solving.", "en"]
]
print("Creating Gradio interface...")
# Modify the analyze_text function to use the fallback concept map if needed
def analyze_text_with_fallback(text, lang):
if not text.strip():
return "Please enter some text.", "No questions generated.", None
try:
print("Generating summary...")
summary = summarize_text(text, lang)
print("Generating questions...")
questions = generate_questions(summary)
print("Creating concept map...")
try:
# Try the main concept map generator first
concept_map_image = generate_concept_map(summary, questions)
except Exception as e:
print(f"Main concept map failed: {e}, using fallback")
# If it fails, use the fallback generator
concept_map_image = generate_simple_concept_map(summary, questions)
# Format questions as a list
questions_text = "\n".join([f"- {q}" for q in questions])
return summary, questions_text, concept_map_image
except Exception as e:
import traceback
print(f"Error processing text: {str(e)}")
print(traceback.format_exc())
return f"Error processing text: {str(e)}", "", None
iface = gr.Interface(
fn=analyze_text_with_fallback, # Use the function with fallback
inputs=[gr.Textbox(lines=10, placeholder="Enter text here..."), gr.Dropdown(["ar", "en"], label="Language")],
outputs=[gr.Textbox(label="Summary"), gr.Textbox(label="Questions"), gr.Image(label="Concept Map")],
examples=examples,
title="AI Study Assistant",
description="Enter a text in Arabic or English and the model will summarize it and generate questions and a concept map."
)
# For Colab, we need to use a public URL
iface.launch(share=True)