Spam-Classifier

Description

• New Updates:
  • Migrated from TF v2.6 to TF v2.11
  • Made use of TF Datasets to load and process data
  • Integrated tokenizer and model into a single end-to-end model
• Extended Dataset: Downloaded the dataset "sms spam classifier" from kaggle and extended the dataset using spam sms from my personal mobile phone, Thus making it a personalized dataset.
• Analysed the text using text cloud.
• Standardization: Lower text and removing punctuations & special chars.
• Data split: Train 70% & Test 30%.
• Models Trained: Conv-1D model & LSTM model
• Epochs Trained: 10 epochs
• K-Folds Trained: 10 folds
• Metrics Computed: Precision | Recall | F1-Score | ROC Curve
• Model Selection: LSTM model has an overall higher mean F-1 score than Conv model for 10-Fold CV. However, on looking at the individual F-1 score of spam & not-spam classes, Conv model gives a better score.
  Check the Notebook for detailed evaluation metrics. The module is containerized via Docker
• Model Enhancements: Pruning was done to add sparsity in the model to reduce the size of the model without losing on accuracy

Input and Output:

Input: { "text": string }
Output: { "prediction": "spam/not-spam", "confidence": range (0 to 1) }

Tech Stack:

Python 3.8, Tensorflow 2.11, Pandas, Numpy, Docker, AWS Lambda, AWS ECR, and GitHub CI/CD.

Deployment:

This project has been deployed as a AWS Lambda function using container image from AWS ECR service and made available as an API. AWS Lambda functions are cost-effective solutions for personal projects that have minimal network traffic.
In order to facilitate for a higher network traffic, it can easily be done so by pulling the ECR image into an EC2 container and deploying the container.