This repo contains deep learning implementations in pytorc for Fashion-MNIST dataset. Fashion-MNIST dataset contains 70,000 fashion imgaes with 10 class labels in total, which is devided into 60,000 and 10,000 examples for a training and test set, respectively. As a benchmarking purpose, MNIST dataset containing hand-written degits has been widely used. Fashion-MNIST was created as traditional MNIST dataset is too simple for benchmarking, For more details of Fashion-MNIST, please visit the original repo.

The implmented models are as follows.

• Logistic Regression
• Multilayer Perceptron (MLP)
  • Three layers neural networks with leaky relu activation
  • The default of hidden units are [256, 128, 64] (you can specify different configurations in cli).
• Convolutional Neural Networks (CNN)
  • Three convolution blcoks and 2 fully connected layers
  • Each convolution block contains Conv2d followed by batch normalization, leaky reu activation, and max pooling layer.
  • Dropout layer is added to fully connected layers.
  • Each CNN block dimension is
    • CNN layer 1: input 28 * 28 * 1, output 28 * 28 * 16
    • Max pooling 1: om[it 28 * 28 * 16, output 14 * 14 * 16
    • CNN layer 2: input 14 * 14 * 16, outpu 14 * 14 * 32
    • Max pooling 2: input 14 * 14 * 32, output 7 * 7 * 32
    • CNN layer 3: input 7 * 7 * 32, output 7 * 7 * 64

• dataset: contains Fashion-MNIST dataset
• models.py: model implementation with pytorch
• trainer.py: a training script for the given model
• test.py: a testing script fot the trained model

To replicate the model development, first you need to install pytorch.

Run the following command for training.

• Logistic Regression

python trainer.py --model 'logistic regression' --epochs 100 --batch_size 32 --optimizer adam --learning_rate 0.001 --weight_decay 0.0001 --model_save True

• MLP

python trainer.py --model mlp --epochs 300 --batch_size 64 --optimizer adam --learning_rate 0.0001 --weight_decay 0.00001 --model_save True

• CNN

python trainer.py --model 'cnn' --epochs 300 --batch_size 64 --optimizer adam --learning_rate 0.0001 --weight_decay 0.00001 --model_save True

You can refer to the trainer.py to see more options of cli flags.

• Logistic Regression

python test.py --model_path outputs/mnist_logistic\ regression.pt --model_name mlp --batch_size 64

• MLP

python test.py --model_path outputs/mnist_mlp.pt --model_name mlp --batch_size 64

• CNN

python test.py --model_path outputs/mnist_cnn.pt --model_name cnn --batch_size 64