nguyennp86's picture
Deploy Email Classifier API
1291f7a
from fastapi import FastAPI, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from fastapi.responses import HTMLResponse
from pydantic import BaseModel
import joblib
import pandas as pd
import numpy as np
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.naive_bayes import MultinomialNB
from sklearn.pipeline import Pipeline
from sklearn.model_selection import train_test_split
from sklearn.metrics import classification_report, accuracy_score
import re
import os
from typing import List, Dict, Any
import logging
# Configure logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
# Initialize FastAPI app
app = FastAPI(
title="Email Attachment Classifier API",
description="API to classify whether an email has attachments or not using Naive Bayes",
version="1.0.0",
docs_url="/docs",
redoc_url="/redoc"
)
# Add CORS middleware
app.add_middleware(
CORSMiddleware,
allow_origins=["*"],
allow_credentials=True,
allow_methods=["*"],
allow_headers=["*"],
)
# Pydantic models
class EmailInput(BaseModel):
message: str
class EmailBatchInput(BaseModel):
messages: List[str]
class PredictionResponse(BaseModel):
message: str
prediction: int
prediction_label: str
confidence: float
probabilities: Dict[str, float]
class BatchPredictionResponse(BaseModel):
predictions: List[PredictionResponse]
class ModelInfo(BaseModel):
model_type: str
accuracy: float
feature_count: int
training_samples: int
# Global variables
model_pipeline = None
model_info = None
def preprocess_text(text: str) -> str:
"""Preprocess email text"""
# Convert to lowercase
text = text.lower()
# Remove extra whitespace
text = re.sub(r'\s+', ' ', text)
# Remove special characters but keep basic punctuation
text = re.sub(r'[^\w\s,.\-!?]', ' ', text)
return text.strip()
def load_and_train_model():
"""Load data and train the Naive Bayes model"""
global model_pipeline, model_info
try:
# Load the dataset (assuming it's in the same directory)
if os.path.exists('Synthetic_Email_Dataset.csv'):
df = pd.read_csv('Synthetic_Email_Dataset.csv')
else:
logger.warning("Dataset file not found, creating sample data")
# Create sample data for demonstration
sample_data = {
'label': [0, 1, 0, 1] * 100,
'message': [
"Hello, You asked for it, so here is the notes. Warm wishes, David",
"Good morning, Just sharing the meeting agenda as requested. Cheers, Anna",
"Dear team, As discussed, I'm sending the manual. Regards, Emily",
"Hi all, Please find attached the project plan. Thanks, Michael"
] * 100
}
df = pd.DataFrame(sample_data)
# Preprocess messages
df['processed_message'] = df['message'].apply(preprocess_text)
# Split data
X = df['processed_message']
y = df['label']
X_train, X_test, y_train, y_test = train_test_split(
X, y, test_size=0.2, random_state=42, stratify=y
)
# Create pipeline
model_pipeline = Pipeline([
('tfidf', TfidfVectorizer(
max_features=1000,
ngram_range=(1, 2),
stop_words='english',
lowercase=True,
min_df=1,
max_df=0.95
)),
('classifier', MultinomialNB(alpha=1.0))
])
# Train model
logger.info("Training Naive Bayes model...")
model_pipeline.fit(X_train, y_train)
# Evaluate model
y_pred = model_pipeline.predict(X_test)
accuracy = accuracy_score(y_test, y_pred)
# Store model info
model_info = ModelInfo(
model_type="Multinomial Naive Bayes",
accuracy=round(accuracy, 4),
feature_count=model_pipeline.named_steps['tfidf'].vocabulary_.__len__(),
training_samples=len(X_train)
)
logger.info(f"Model trained successfully with accuracy: {accuracy:.4f}")
logger.info(f"Feature count: {model_info.feature_count}")
# Save model
joblib.dump(model_pipeline, 'email_classifier_model.pkl')
logger.info("Model saved successfully")
return True
except Exception as e:
logger.error(f"Error in training model: {str(e)}")
return False
def load_pretrained_model():
"""Load pretrained model if available"""
global model_pipeline, model_info
try:
if os.path.exists('email_classifier_model.pkl'):
model_pipeline = joblib.load('email_classifier_model.pkl')
logger.info("Pretrained model loaded successfully")
# Set default model info if not available
if model_info is None:
model_info = ModelInfo(
model_type="Multinomial Naive Bayes",
accuracy=0.92, # Default value
feature_count=len(model_pipeline.named_steps['tfidf'].vocabulary_),
training_samples=320 # Default value
)
return True
except Exception as e:
logger.error(f"Error loading pretrained model: {str(e)}")
return False
@app.on_event("startup")
async def startup_event():
"""Initialize model on startup"""
logger.info("Starting Email Classifier API...")
# Try to load pretrained model first
if not load_pretrained_model():
# If no pretrained model, train new one
if not load_and_train_model():
logger.error("Failed to initialize model")
@app.get("/", response_class=HTMLResponse)
async def root():
"""Root endpoint with API documentation"""
html_content = """
<!DOCTYPE html>
<html>
<head>
<title>Email Attachment Classifier API</title>
<style>
body { font-family: Arial, sans-serif; margin: 40px; }
.header { color: #2c3e50; }
.endpoint { background-color: #f8f9fa; padding: 15px; margin: 10px 0; border-radius: 5px; }
.method { color: #27ae60; font-weight: bold; }
code { background-color: #e9ecef; padding: 2px 4px; border-radius: 3px; }
</style>
</head>
<body>
<h1 class="header">📧 Email Attachment Classifier API</h1>
<p>This API classifies whether an email message indicates an attachment or not using Naive Bayes classifier.</p>
<h2>Available Endpoints:</h2>
<div class="endpoint">
<h3><span class="method">GET</span> /info</h3>
<p>Get model information and statistics</p>
</div>
<div class="endpoint">
<h3><span class="method">POST</span> /predict</h3>
<p>Predict single email message</p>
<p><strong>Body:</strong> <code>{"message": "Your email content here"}</code></p>
</div>
<div class="endpoint">
<h3><span class="method">POST</span> /predict-batch</h3>
<p>Predict multiple email messages</p>
<p><strong>Body:</strong> <code>{"messages": ["Email 1", "Email 2", ...]}</code></p>
</div>
<div class="endpoint">
<h3><span class="method">GET</span> /health</h3>
<p>Check API health status</p>
</div>
<h2>Interactive Documentation:</h2>
<p>Visit <a href="/docs">/docs</a> for Swagger UI or <a href="/redoc">/redoc</a> for ReDoc</p>
<h2>Labels:</h2>
<ul>
<li><strong>0:</strong> No attachment mentioned</li>
<li><strong>1:</strong> Attachment mentioned</li>
</ul>
</body>
</html>
"""
return HTMLResponse(content=html_content, status_code=200)
@app.get("/health")
async def health_check():
"""Health check endpoint"""
if model_pipeline is None:
return {"status": "unhealthy", "message": "Model not loaded"}
return {"status": "healthy", "message": "API is running"}
@app.get("/info", response_model=ModelInfo)
async def get_model_info():
"""Get model information"""
if model_info is None:
raise HTTPException(status_code=503, detail="Model not initialized")
return model_info
@app.post("/predict", response_model=PredictionResponse)
async def predict_single(email: EmailInput):
"""Predict single email message"""
if model_pipeline is None:
raise HTTPException(status_code=503, detail="Model not loaded")
try:
# Preprocess input
processed_message = preprocess_text(email.message)
# Make prediction
prediction = model_pipeline.predict([processed_message])[0]
probabilities = model_pipeline.predict_proba([processed_message])[0]
# Prepare response
prediction_label = "Has attachment" if prediction == 1 else "No attachment"
confidence = float(max(probabilities))
prob_dict = {
"no_attachment": float(probabilities[0]),
"has_attachment": float(probabilities[1])
}
return PredictionResponse(
message=email.message,
prediction=int(prediction),
prediction_label=prediction_label,
confidence=confidence,
probabilities=prob_dict
)
except Exception as e:
logger.error(f"Prediction error: {str(e)}")
raise HTTPException(status_code=500, detail=f"Prediction failed: {str(e)}")
@app.post("/predict-batch", response_model=BatchPredictionResponse)
async def predict_batch(emails: EmailBatchInput):
"""Predict multiple email messages"""
if model_pipeline is None:
raise HTTPException(status_code=503, detail="Model not loaded")
if len(emails.messages) > 100:
raise HTTPException(status_code=400, detail="Maximum 100 messages per batch")
try:
predictions = []
# Preprocess all messages
processed_messages = [preprocess_text(msg) for msg in emails.messages]
# Make batch predictions
batch_predictions = model_pipeline.predict(processed_messages)
batch_probabilities = model_pipeline.predict_proba(processed_messages)
# Prepare responses
for i, (message, prediction, probabilities) in enumerate(
zip(emails.messages, batch_predictions, batch_probabilities)
):
prediction_label = "Has attachment" if prediction == 1 else "No attachment"
confidence = float(max(probabilities))
prob_dict = {
"no_attachment": float(probabilities[0]),
"has_attachment": float(probabilities[1])
}
predictions.append(PredictionResponse(
message=message,
prediction=int(prediction),
prediction_label=prediction_label,
confidence=confidence,
probabilities=prob_dict
))
return BatchPredictionResponse(predictions=predictions)
except Exception as e:
logger.error(f"Batch prediction error: {str(e)}")
raise HTTPException(status_code=500, detail=f"Batch prediction failed: {str(e)}")
if __name__ == "__main__":
import uvicorn
uvicorn.run(app, host="0.0.0.0", port=7860)