Fashion-MNIST is an alternative to the more commonly used MNIST dataset. It contains a set of 28x28 grayscale images, each of which is assigned with one of the following labels:
- T-shirt/top
- Trouser
- Pullover
- Dress
- Coat
- Sandal
- Shirt
- Sneaker
- Bag
- Ankle boot
Since fashion-MNIST is a more challenging image recognition task compared to MNIST, models will generally perform worse on this dataset. Below is a summary of a VGG-like CNN model that achieves 93.8% accuracy on the test set (note that the number of trainable parameters are decreased and dropout layers are added to avoid overfitting):
