Model1 improved - ipynb

.gitignore

@@ -0,0 +1,27 @@
+# macOS
+.DS_Store
+
+# Python cache
+__pycache__/
+*.py[cod]
+.ipynb_checkpoints/
+
+# Local environments
+.venv/
+venv/
+env/
+
+# Large/generated model files
+*.h5
+*.keras
+*.pkl
+*.joblib
+*.ckpt
+checkpoints/
+models/
+
+# Logs and temporary outputs
+logs/
+*.log
+tmp/
+temp/

README.md

@@ -1,70 +1,130 @@
-# Image Classification with CNN
-
-Build a Convolutional Neural Network (CNN) model to classify images from the CIFAR-10 dataset into 10 predefined categories through iterative architectural improvements and advanced training techniques.
-
-[Repository link](https://github.com/pavithraaiengineer/CNN-DL-ImageClassification)
-
-## Project Results
-In this project, we processed the CIFAR-10 dataset and explored the results of 10 different classifier iterations:
-* Models #1 - #4: Building sequential CNN models from scratch.
-* Models #5 - #8: Deepening architecture and implementing Global Average Pooling.
-* Model #9: Transfer learning and fine-tuning using EfficientNetB0.
-* Model #10: Ensemble,
-* Deployment for final prediction.
-
-## Repository Folders and Files
-
-### Documents
-* **Presentation Slides**
-* **Read me file**
-* **10 notebooks with diferent models**
-
----
-
-## Technical Evolution & Performance
-
-```diff
-# Training Progress Log
-- Model #1: Baseline (2 Conv layers) -> Test Acc: 55.70%
-- Model #2: Added BN and Dropout -> Test Acc: 62.63%
-+ Model #3: Optimized 32x32 + BN -> Test Acc: 82.15%
-+ Model #4: 6 Conv layers + L2 Reg -> Test Acc: 81.00%
-+ Model #6: 13 Conv layers + GAP -> Test Acc: 93.03%
-+ Model #7: Contrast Aug + AdamW -> Test Acc: 93.12%
-+ Model #8: Brightnes augmentation. + AdamW -> Test Acc: 93.12%
-+ Model #9: Transfer Learning (EfficientNetB0) -> Test Acc: 91.35%
-+ Model #10: Final Ensemble for Maximum Stability 93.21%
-
----
-## PROBLEMS & CHALLENGES
-
-```diff
-- Overfitting: High training accuracy but failed to generalize to the test set.
-- Class Weakness: "Cat" and "Dog" categories had the lowest performance.
-- Recall Drop: Cat/Dog recall declined continuously during training iterations.
-- Feature Extraction: Baseline model was too shallow to capture complex image patterns.
-- Data Quality: 32x32 resolution was too low and blurry for distinct classification.
-
----
+# CIFAR-10 Image Classification with CNNs
+
+This project builds and compares deep learning models for image classification on the CIFAR-10 dataset. The goal was to improve model performance step by step, starting from a simple Convolutional Neural Network and ending with stronger CNN, transfer learning, and ensemble experiments.
+
+## Dataset
+
+The project uses the CIFAR-10 image dataset:
+
+- 60,000 color images
+- 10 image classes
+- 32x32 pixel image size
+- 50,000 training images
+- 10,000 test images
+
+Classes:
+
+```text
+airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck
+```
+
+The dataset is small and low-resolution, which makes classification harder for visually similar classes such as cats and dogs.
+
+## Project Goal
+
+The main goal was to train a CNN model that can classify CIFAR-10 images accurately while reducing overfitting. We tested multiple model versions to understand how architecture changes and training strategies affect performance.
+
+## Repository Structure
+
+```text
+.
+├── README.md
+├── notebooks/
+│   ├── Model1.ipynb
+│   ├── Model1_v2.ipynb
+│   ├── Model2.ipynb
+│   ├── Model2_4.ipynb
+│   ├── Model2_6.ipynb
+│   ├── Model3.ipynb
+│   ├── Model4.ipynb
+│   ├── Model5.ipynb
+│   ├── Model6.ipynb
+│   ├── Model7.ipynb
+│   ├── Model8.ipynb
+│   ├── Model9.ipynb
+│   └── model10.ipynb
+└── presentations/
+    └── CIFAR-10 CNN Classification Project (1).key
+```
+
+All model notebooks are stored in the `notebooks/` folder so the repository is easier to navigate.
+
+## Methods Tested
+
+We compared several CNN and deep learning strategies:
+
+- Baseline CNN built from scratch
+- Deeper CNN architectures
+- Batch Normalization
+- Dropout regularization
+- L2 regularization
+- Data augmentation
+- Global Average Pooling
+- AdamW optimizer
+- Transfer learning with EfficientNetB0
+- Ensemble prediction for final stability
+
+## Model Evolution
+
+| Model | Main Change | Test Accuracy |
+|---|---|---:|
+| Model 1 | Baseline CNN with 2 convolution layers | 55.70% |
+| Model 2 | Added Batch Normalization and Dropout | 62.63% |
+| Model 3 | Optimized CNN for 32x32 images with BatchNorm | 82.15% |
+| Model 4 | 6 convolution layers with L2 regularization | 81.00% |
+| Model 6 | Deeper CNN with 13 convolution layers and GAP | 93.03% |
+| Model 7 | Contrast augmentation and AdamW | 93.12% |
+| Model 8 | Brightness augmentation and AdamW | 93.12% |
+| Model 9 | Transfer learning with EfficientNetB0 | 91.35% |
+| Model 10 | Final ensemble model | 93.21% |
+
+## Best Result
+
+The best final result came from the ensemble model:
+
+```text
+Best test accuracy: 93.21%
+```
+
+Model 7 was the strongest single model, while Model 10 gave the most stable final prediction through ensembling.
+
+## Main Challenges
+
+- Overfitting appeared when training accuracy improved faster than test accuracy.
+- Cat and dog images were often confused because CIFAR-10 images are very small.
+- The 32x32 resolution limits how much visual detail the model can learn.
+- Deeper models helped, but regularization and augmentation were needed to improve generalization.
+
 ## Recommended Notebooks
-To see the best performing results and the final architecture, we recommend starting with these notebooks:
 
-```diff
-+ Recommended: model_7_training (Best Single Model)
-+ Recommended: model_10_ensemble (Final Ensemble Strategy & Deployment)
-! Note: Model 7 provides the highest individual accuracy, while Model 10 
-! offers the most stable predictions on unseen data and highest accuracy.
+Start with these notebooks if you want to understand the strongest results:
+
+- `notebooks/Model7.ipynb`: best single-model strategy
+- `notebooks/model10.ipynb`: final ensemble strategy
 
-## Installation & Setup
+## How to Run
 
-To replicate the environment and run the notebooks, follow these steps:
+Create and activate a Python environment:
 
 ```bash
-# 1. Create a virtual environment
 python -m venv .venv
-
-# 2. Activate the environment
-# On Windows:
-.venv\Scripts\activate
-# On macOS/Linux:
 source .venv/bin/activate
+```
+
+Install the main libraries:
+
+```bash
+pip install tensorflow keras numpy pandas matplotlib seaborn scikit-learn jupyter
+```
+
+Open Jupyter:
+
+```bash
+jupyter notebook
+```
+
+Then run the notebooks from the `notebooks/` folder.
+
+## Notes
+
+This project was completed as a deep learning image classification project. This repository version is organized to make the model experiments, results, and presentation easier to review.