FashionMNIST ConvNets Deep Learning Project

Using two Deep learning models to Classify the image from Fashion-MNIST dataset

• Basic Model (fully connected layer model)
• Convolutional neural networks (ConvNets)

Two models will use the Keras framework with R implementation.

Data Set

• 60000 images for training and 10000 images for testing
• Each example is a 28x28 gray-scale image, associated with a label from
  • 0 T-shirt/top
  • 1 Trouser
  • 2 Pullover
  • 3 Dress
  • 4 Coat
  • 5 Sandal
  • 6 Shirt
  • 7 Sneaker
  • 8 Bag
  • 9 Ankle boot 
  • 10 classes
• The dataset is CSV format. The detailed format is label, pixel1, pixel2, pixel3, … pixel784. Each image is 28 pixels in height and 28 pixels in width, for a total of 784 pixels in total

Problem Definition

• Train the 60000 images and test 10000 images to classify the image’s label

  Problem Type

• Multi-Class and single-label classification

  Model Configuration

• The softmax as the last-layer’s activation and the loss function will use the categorical_crossentropy.

  Hyper-parameters setting

• Use the dropout and L2 regularization to reduce over-fitting effects.

  Assembling Data Set

• dataset_fashion_mnist() function from Keras to download the dataset (simple model)
• Downloaded from the Kaggle.com, then use the read_csv() to manipulate the data (ConvNets Model)

Result

Simple Model Result

• Simple deep learning model achieves an accuracy of 88.11% and loss of 34.15%.

ConvNets Model Result

• ConvNets model achieves an accuracy of 91.85% and loss is 20%, up from the previous model’s accuracy of 88.11% and loss of 34.15%.

Summary

• Our ConvNets model achieved an accuracy of 91.85%. It turns out our classifier does better than the Kaggle’s best baseline reported here , which is an SVM classifier with mean accuracy of 89.7%.
• Comparing the simple model, ConvNets is the best model for the attacking the image classification problems.
• Tuning the model and hyper-parameters is very important to improve the accuracy.