TensorFlow Multi-Model AI Suite

A collection of multiple machine-learning models built using TensorFlow, covering genres such as language classification, sentiment analysis, depression prediction, and mask detection.

All models come with:

• Full training scripts
• Data preprocessing utilities
• Saved tokenizers/encoders
• A unified inference interface (model_tester.py)
• Clean modular folder structure

⚠️ Important: This repository does not include datasets or trained weights. You must train your own models using the provided scripts, then place them into the ./Models folder.

---

📂 Project Structure

tensorflow/
│
├── depression_predictor/
│   ├── training scripts
│   ├── encoders (after training)
│   └── scaler.pkl
│
├── lang_classifier/
│   ├── training scripts
│   └── char_tokenizer.json
│
├── mask_detector/
│   ├── normalization scripts
│   ├── training scripts
│   └── evaluation utilities
│
├── sentiment_analysis/
│   ├── training scripts
│   └── word_tokenizer.json
│
├── tumor_detection/
│   ├── normalization scripts
│   ├── training scripts
│   └── class_indices.json
│
├── leaf_disease/
│   ├── normalization scripts
│   ├── training scripts
│   └── class_indices.json
│
├── toxic_classifier/
│   ├── training scripts
│   └── toxic_tokenizer.json
│
├── clustering/
│   └── cat_dog/
│       ├── semi_trainer.py
│       ├── semi_cluster.py
│       ├── normalization.py
│       └── documentation.ipynb
│
├── .gitignore
└── model_tester.py   ← Unified inference for all models

---

🧠 Available Models

Below is an overview of each model included in this repository.

---

1️⃣ Language Classifier

Goal: Detect whether a sentence is written in English, Hindi, or Punjabi. Techniques Used:

• Character-level tokenizer
• Bi-directional LSTM
• Multi-class softmax

Training Output:

• char_tokenizer.json
• language_classifier.h5

---

2️⃣ Sentiment Analysis Model

Goal: Classify text as positive, neutral, or negative. Techniques Used:

• Word-level tokenizer
• BiLSTM text classifier
• 30k+ dataset support
• ≥100k trainable parameters

Training Output:

• word_tokenizer.json
• sentiment_model.keras

---

3️⃣ Depression Predictor (Tabular ML Model)

Goal: Predict whether a student shows signs of depression using these features:

• Age
• Gender
• Department
• CGPA
• Sleep Duration
• Study Hours
• Social Media Hours
• Physical Activity
• Stress Level

Techniques Used:

• LabelEncoder for categorical columns
• StandardScaler for numerical columns
• Fully-connected neural network
• ≥100k trainable parameters

Training Output:

• categorical_encoders.pkl
• scaler.pkl
• depression_model.keras

---

4️⃣ Mask Detector (CNN)

Goal: Detect whether a person is wearing a mask in an image. Techniques Used:

• OpenCV preprocessing (256×256 normalization)
• CNN with Conv2D + MaxPooling
• Binary classification
• Confusion matrix

Training Output:

• mask_detector.keras

---

5️⃣ Brain Tumor Detection (Multi-Class CNN)

Goal: Classify MRI brain scans into 4 categories:

• No Tumor (notumor)
• Glioma (glioma)
• Meningioma (meningioma)
• Pituitary (pituitary)

Techniques Used:

• OpenCV preprocessing (256×256 normalization)
• 3-layer CNN architecture
• Multi-class classification with softmax
• Confusion matrix for detailed analysis
• ≥100k trainable parameters

Training Output:

• brain_tumor_model.keras
• class_indices.json

---

6️⃣ Leaf Disease Classification (CNN)

Goal: Classify plant leaf diseases into 38 different categories (e.g., Apple Scab, Tomato Blight, Potato Late Blight, etc.). Techniques Used:

• OpenCV preprocessing (256×256 normalization)
• 3-layer CNN architecture with Conv2D + MaxPooling
• Multi-class classification with softmax
• Plant + Disease mapping system
• Confusion matrix analysis

Training Output:

• leaf_disease_model.keras
• class_indices.json

---

7️⃣ Toxic Comments Classifier (BiLSTM)

Goal: Classify comments into 6 toxicity categories:

• Toxic
• Severe Toxic
• Obscene
• Threat
• Insult
• Identity Hate

Techniques Used:

• Word-level tokenizer (20k vocabulary)
• BiLSTM architecture
• Multi-label classification with sigmoid activation
• Binary Crossentropy loss
• ≥50k trainable parameters

Training Output:

• toxic_model.keras
• toxic_tokenizer.json

---

🐾 Clustering (Cats vs Dogs)

Goal: Demonstrate unsupervised and semi-supervised workflows to separate cat and dog images using dimensionality reduction and clustering.

Key items in clustering/cat_dog/:

• semi_trainer.py — Semi-supervised embedding/trainer used to generate image embeddings (WORKING).
• semi_cluster.py — Clustering pipeline that runs PCA/t-SNE and KMeans to produce the notebook's results and visualizations (WORKING).
• normalization.py — Image normalization (224×224 resize, BGR→RGB) used to prepare normalized_dataset/.
• documentation.ipynb — Full walkthrough of the normalization → PCA → KMeans → evaluation → visualization flow.

Experimental scripts (for exploration only):

• trainer.py — Early/trial trainer (EXPERIMENTAL)
• autoencoder_train.py — Autoencoder experiments for dimensionality reduction (EXPERIMENTAL)
• cluster.py — Alternate clustering prototype (EXPERIMENTAL)

Training / Run order (recommended):

1. Normalize images: python clustering/cat_dog/normalization.py
2. (Optional) Generate embeddings: python clustering/cat_dog/semi_trainer.py
3. Run clustering pipeline: python clustering/cat_dog/semi_cluster.py

---

🎯 Unified Inference System — model_tester.py

This script allows you to test any of the trained models from a single entry point.

Usage:

python model_tester.py

Then choose:

0 → Language Classifier
1 → Sentiment Analysis
2 → Depression Predictor
3 → Mask Detector
4 → Brain Tumor Detection
5 → Leaf Disease Classifier
6 → Toxic Comments Classifier
7 → Clustering (Cats vs Dogs)

---

Training Your Own Models

Each subfolder contains:

• Training script
• Preprocessing utilities
• Encoders/tokenizers
• Graph generation
• Evaluation logic

🔧 Steps to Train:

1. Prepare your dataset
2. Run the training script inside the appropriate module
3. After training, move the generated model file to:

./Models/

Examples:

./Models/language_classifier.keras
./Models/sentiment_model.keras
./Models/depression_model.keras
./Models/mask_detector.keras
./Models/brain_tumor_model.keras
./Models/leaf_disease_model.keras
./Models/toxic_model.keras
# Clustering pipeline outputs (examples)
./clustering/cat_dog/normalized_dataset/
./clustering/cat_dog/embeddings.npy
./clustering/cat_dog/cluster_labels.csv
./clustering/cat_dog/visualizations/cluster_tsne.png

4. Now you can use model_tester.py to run inference.

---

💡 Recommended Folder for Your Own Models

./Models/
│
├── language_classifier.h5
├── sentiment_model.keras
├── depression_model.keras
├── mask_detector.keras
├── brain_tumor_model.keras
├── leaf_disease_model.keras
└── toxic_model.keras

This keeps all inference handling consistent with model_tester.py.

---

🔁 Recommended Clustering Artifacts Layout

Keep clustering outputs alongside the clustering module to simplify debugging and reproducibility. Example structure:

clustering/cat_dog/
│
├── normalized_dataset/        # preprocessed images used for embedding extraction
├── embeddings.npy            # numpy array of image embeddings produced by semi_trainer
├── cluster_labels.csv        # mapping of image filename -> cluster id produced by semi_cluster
└── visualizations/           # PCA/t-SNE/cluster plots (PNGs)

---

📦 Dependencies

See requirements.txt:

tensorflow
numpy
pandas
scikit-learn
opencv-python
matplotlib
seaborn

---

📜 License

This project is open-source under the MIT License.