Skip to content

cnn_with_python.md

Latest commit

 

History

History
116 lines (83 loc) · 3.73 KB

cnn_with_python.md

File metadata and controls

116 lines (83 loc) · 3.73 KB

🧠 CNN on CIFAR‑10 with Keras/TensorFlow — Preprocess → Define → Train

This file, following with python code, walks through a minimal, reproducible pipeline to build a Convolutional Neural Network (CNN) on the CIFAR‑10 dataset using Keras/TensorFlow:

Steps:

  1. Preparing to build a CNN in Python (data loading, preview, normalization, one‑hot encoding)
  2. Defining a CNN in Python (Sequential model with Conv/BatchNorm/MaxPool/Dropout blocks → Dense head)
  3. Training a CNN in Python (compile with Adam + categorical cross‑entropy, fit for 20 epochs, visualize curves, evaluate on test set)

📚 Dataset: CIFAR‑10

  • 10 classes: airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck
  • Images: 50,000 train / 10,000 test
  • Resolution: 32×32 RGB
  • Loaded via tf.keras.datasets.cifar10 (channels‑last input: (32, 32, 3))

🗂️ Project Structure

  • .
  • ├─ README.md
  • ├─ requirements.txt
  • └─ cnn_cifar10_full.py # Single script: preprocess + model + training + evaluation

✅ Requirements

tensorflow>=2.9.0
numpy>=1.21.0
matplotlib>=3.5.0
scikit-learn>=1.0.0   # optional; only needed for confusion matrix

## Install

``
pip install -r requirements.txt
``

OR

``
pip install tensorflow numpy matplotlib scikit-learn
``

## 🚀 Quick Start (End‑to‑End)
- Run the full pipeline (preprocessing → define → train → evaluate):

python cnn_cifar10_full.py --model compact --epochs 20 --batch-size 64 --val-split 0.1 --save-model --plot-cm
``

### Outputs:
- Console: model summary, training progress, final test accuracy and loss
- Images: outputs/accuracy_loss_curves.png (+ optional outputs/confusion_matrix.png)
- Model: outputs/cifar10_cnn.h5 (if --save-model is used)

## 🧰 Preparing to Build a CNN in Python (Preprocessing)

We:
- Load CIFAR‑10: x_train, y_train, x_test, y_test
- Preview a few samples to sanity‑check inputs
- Normalize pixels by dividing by 255 → values in [0, 1] to help model convergence
- One‑hot encode labels (0..9 → 10‑dim vectors) for multi‑class softmax

## 🏗️ Defining a CNN in Python (Model Architecture)
- We use a Keras Sequential model with channels‑last input (32, 32, 3):

### Conv Block ×3 (filters 32 → 64 → 128):
- Conv2D(3×3, ReLU, padding="same")
- BatchNormalization() (faster training, stable gradients)
- MaxPooling2D(2×2)
- Dropout(0.25) (reduces overfitting)

### Dense Head:

- Flatten()
- Dense(512, ReLU)
- BatchNormalization()
- Dropout(0.5)

### Output:
- Dense(10, Softmax) (multi‑class probabilities)
- I provide two implementations: step‑by‑step (.add) and compact (single layer list). Choose with --model stepwise or --model compact.

## 🏋️ Training a CNN in Python

- Compile: optimizer="adam", loss="categorical_crossentropy", metrics=["accuracy"]
- Fit: epochs=20, batch_size=64, validation_split=0.1
- Visualize: plot training vs validation accuracy/loss across epochs
- Evaluate: report final test accuracy & loss; optionally create a confusion matrix

**Healthy training behavior:**

- Validation curves track training curves closely → limited overfitting
- Expect ~70–75% test accuracy for this baseline (varies by run) without augmentation

## 🛠️ Troubleshooting
- Matplotlib on headless servers:

import matplotlib matplotlib.use("Agg") # before importing pyplot

- GPU/CPU control:

CUDA_VISIBLE_DEVICES="" python cnn_cifar10_full.py

- Shape errors: Ensure input shape is (32, 32, 3) and labels are one‑hot encoded with 10 classes.
- Overfitting: Increase dropout, add data augmentation (tf.keras.layers.RandomFlip/RandomRotation), or reduce epochs.

 Following code shows all these in detail. Refer cnn_cifar10_full.py