fashion_mnist_tf

Fashion-MNIST Image Classifier

Model Description

This is a neural network model trained to classify Fashion-MNIST images into 10 different clothing categories. The model uses a simple feedforward architecture with dense layers and achieves good performance on the Fashion-MNIST benchmark dataset.

Model Architecture

• Input Layer: 28x28 grayscale images
• Preprocessing: Pixel normalization (rescaling to [0,1])
• Hidden Layers:
  • Dense layer with 128 neurons (ReLU activation)
  • Dense layer with 64 neurons (ReLU activation)
• Output Layer: Dense layer with 10 neurons (Softmax activation)
• Total Parameters: Approximately 109K parameters

Training Details

• Dataset: Fashion-MNIST (60,000 training images, 10,000 test images)
• Optimizer: Stochastic Gradient Descent (SGD)
• Loss Function: Categorical Crossentropy
• Batch Size: 64
• Max Epochs: 100 (with early stopping)
• Validation Split: 20% of training data
• Callbacks: Early stopping and learning rate reduction on plateau

Performance

The model was evaluated on the Fashion-MNIST test set with the following metrics:

• Accuracy: 0.8386
• F1 Score: 0.8397

Intended Use

This model is designed for:

• Educational purposes and learning about image classification
• Classifying Fashion-MNIST images into 10 categories:
  • 0: T-shirt/top
  • 1: Trouser
  • 2: Pullover
  • 3: Dress
  • 4: Coat
  • 5: Sandal
  • 6: Shirt
  • 7: Sneaker
  • 8: Bag
  • 9: Ankle boot

Usage

import tensorflow as tf

# Load the model
model = tf.keras.models.load_model('fashion_mnist_model.h5')

# Prepare your image (28x28 grayscale, values 0-255)
# image = your_preprocessed_image

# Make prediction
prediction = model.predict(image.reshape(1, 28, 28))
predicted_class = tf.argmax(prediction, axis=1).numpy()[0]

Limitations

• Only works with 28x28 grayscale images
• Trained specifically on Fashion-MNIST dataset
• May not generalize well to real-world clothing images
• Simple architecture may not capture complex patterns

Training Procedure

The model was trained with:

1. Data normalization (pixel values scaled to [0,1])
2. Train/validation split (80/20)
3. Early stopping based on validation F1 score
4. Learning rate scheduling