Creating deep learning models on MNIST database using Pytorch

MNIST database

http://yann.lecun.com/exdb/mnist/

LogSoftMax Model

Steps

Download the dataset

• use torchvision.dataset to download the MNIST digits dataset and transform to Tensors.
• generate train and test loaders of batch size 64 and assign it to variables trainloader and testloader respectively.
• define the method named model_buid() to build a sequential model.
  • parameters:
    • input_size int: number of input features
    • hidden_sizes array of size 3: array of number of nodes in each of the three hidden layers
    • output int: number of nodes at output layer (or number of classes to classify)
  • returns model: sequential model
    Apply relu activation between each layer and for the final layer out apply logSoftmax.
    Logsoftmax log transformation of softmax output which is given by :$$ \sigma(x_i) = \log \cfrac{e^{x_i}}{\sum_{k=1}^{N}{e^{x_k}}}$$ where N is the number of classes
    more information here
  • using the function we defined, initialize the model to have input size 784, three hidden layers to have 256, 128 and 64 nodes and finally an output layer of size 10 nodes.

Define criterion (loss function) and optimizer

• Define criterion to be negative likelihood loss since the network output is log transformed probabilities.
• Define optimizer to be Adam and set learning rate to 0.003 and pass the parameters of classifier model

Train the classifier

• flatten the image tensor from (1 x 28 x 28) to (1, 784)

Evaluate class probabilities

• transform log transformed softmax output log_ps to exponential and assign it to ps (torch.exp())
• return the top probability and its index from ps (ps.topk())