TensorFlow

---

Index

• Module 1: Basics
• Module 2: Classification
• Module 3: Convolutional Neural Networks
• Module 4: Reinforcement Learning
• Projects
• Models
• Setup

This repository is my personal TensorFlow learning sandbox. Each module represents a concrete step forward - from understanding tensors and preprocessing data to building CNNs, NLP models, and reinforcement learning agents.

This is not a polished library or a tutorial series. It’s a transparent, hands-on log of learning TensorFlow by building real models, saving checkpoints, and measuring results.

---

---

Module 1

Basics

Key learnings:

• Difference between tf.Tensor (immutable) and tf.Variable (mutable, required for trainable weights)
• Tensor slicing behavior vs reassignment limitations
• Practical use of reshape when adapting data to model input requirements
• Data quality issues belong to preprocessing, not model or architecture choice

---

Module 2

Classification

Key learnings:

• Built end-to-end classification pipelines for:
  • Titanic (binary classification)
  • Iris (multiclass classification)
• Proper train/test splitting (x_train, y_train, x_test, y_test)
• Handling missing values using fillna
• Feature preprocessing:
  • Dropped non-informative features (e.g., fare)
  • Scaled numerical features with StandardScaler
  • One-hot encoded categorical features
• Model design:
  • Sigmoid output + binary crossentropy for binary classification
  • Softmax output + sparse categorical crossentropy for multiclass classification
• Used TensorFlow’s Normalization layer instead of external scalers for Iris
• Saved trained models as .keras files
• Achieved ~81% accuracy (Titanic) and ~70% accuracy (Iris)

---

Module 3

Convolutional Neural Networks (CNN)

Key learnings:

• Built CNNs from scratch for image classification using CIFAR-10
• Normalized image inputs to the 0–1 range
• Stacked Conv2D, MaxPooling, and Dense layers for feature extraction and classification
• Achieved ~72% accuracy on CIFAR-10 with a baseline CNN
• Applied image augmentation (zoom, shift, rotation) to improve generalization
• Observed trade-offs between augmentation and raw accuracy (~61%)
• Implemented transfer learning using MobileNetV2 (weights="imagenet", include_top=False)
• Preprocessed inputs to the [-1, 1] range as required by MobileNetV2
• Used GlobalAveragePooling and a custom classification head
• Achieved ~94% accuracy on Dogs vs Cats, highlighting the effectiveness of transfer learning

---

Module 4

Reinforcement Learning

Key learnings:

• Implemented Q-learning on the FrozenLake-v1 environment using Gymnasium
• Initialized and updated a Q-table using the Bellman equation
• Applied an epsilon-greedy strategy to balance exploration and exploitation
• Tuned hyperparameters: learning rate (α), discount factor (γ), exploration rate (ε)
• Implemented epsilon decay to shift from exploration to exploitation
• Evaluated agent performance by averaging rewards across episodes
• Achieved ~72% success rate on FrozenLake

---

Projects:

Sentiments

The IMDB dataset was directory-based (train/pos, train/neg, test/pos, test/neg), with labels inferred from folder names and an unused unsup folder. I loaded it using text_dataset_from_directory with a validation split, seed, and batch size. Reviews were preprocessed using a custom standardization function (lowercasing, removing punctuation, stripping <br /> tags) and a TextVectorization layer with max_tokens=10000 and sequence_length=250 to map words into integers. The model architecture included an Embedding layer, Dropout, GlobalAveragePooling1D, and a Dense output layer with sigmoid activation, compiled with Adam and binary crossentropy. After training with validation split, the model achieved ~81.9% accuracy on the test set and was saved as sentiments.keras.

Fuel Efficiency

I prepared the fuel efficiency dataset by dropping irrelevant columns, handling missing values, and one-hot encoding categorical origin data. After splitting into train/test sets and separating features from the target (MPG), I normalized the numerical features using TensorFlow’s Normalization layer. A simple regression model was built with a Normalizer input and a single Dense output neuron, compiled with Adam and mean absolute error. The model achieved ~ 1.81 MAE on the test set, showing good baseline performance, and was saved as regression.keras.

Fashion MNIST

I trained a CNN on the FashionMNIST dataset, normalizing 28×28 grayscale images to [0,1] and using Conv2D, MaxPooling, Dropout, and Dense layers with softmax for classification. Compiled with Adam and sparse categorical crossentropy, the model trained for 10 epochs with a validation split and achieved ~92.5% accuracy on the test set. Predictions correctly mapped review images to their classes, and the model was saved as fashion_mnist.keras.

Models

This folder contains the trained .keras models from different TensorFlow projects in this repository.
They are saved checkpoints of my experiments - ready to be reloaded for evaluation, predictions, or fine-tuning.

Model Index

Model File	Task / Dataset	Metric Achieved
titanic.keras	Binary Classification (Titanic survival)	~81% Accuracy
iris_species.keras	Multiclass Classification (Iris dataset)	~70% Accuracy
cifar10.keras	Image Classification (CIFAR-10)	~72% Accuracy
cifar_augmented.keras	CIFAR-10 with Data Augmentation	~61% Accuracy
dogsvscat.keras	Transfer Learning (Dogs vs Cats, MobileNetV2)	~94% Accuracy
sentiments.keras	Text Vectorization (IMDb reviews)	85.80 % Accuracy
fuel_efficiency.keras	Regression (Auto MPG dataset)	1.81 MAE
fashion_mnist.keras	Image Classification (Fashion MNIST)	~92.5% Accuracy

---

Setup

Install dependencies:

pip install -r requirements.txt

---

Usage

To load and use any model:

import tensorflow as tf

# Load the trained model
model = tf.keras.models.load_model("models/fashion_mnist.keras")

# Evaluate or predict
loss, acc = model.evaluate(x_test, y_test)
print(f"Accuracy: {acc:.2f}")

Tech Stack

• TensorFlow / Keras
• Python 3

---

Note

This repository documents learning through direct experimentation. The focus is on building, breaking, debugging, and measuring - not on theoretical completeness or production readiness.

The repo is intentionally iterative and unfinished. If you’re exploring TensorFlow, feel free to fork it, experiment, and adapt the ideas into your own projects.

---

Author

Created and maintained by  

---

License

This project is licensed under the .