Classifying Digits with Logistic Regression

Personal Projects #Data Science#Python

Logistic Regression Class#

Create a logistic regression class that takes pytorch.nn.Module as its input. It has 2 functions:

init: A constructor that takes self, inputSize, and numClasses, calls the superclass constructor, and then sets its linear to nn.Linear(inputSize, numClasses).
forward: Calculates softmax with forward pass using self.linear(x), where x is an image.

Logistic Regression Model#

Split and load the MINST digit dataset into training and testing data. Next, define the hyperparameters as follows:

inputSize=784 - since images are 28x28 which is 784 pixels
numClasses=10 - since there are 10 possible labels for each image (0-9)
epochs-5 - run training 5 times
batchSize=100 - train on 100 image batches to prevent memory overflow
learningRate=0.001 Then create the Logistic Regression model using Pytorch, the inputSize and numClasses. Set the loss function to use cross entropy with crit=nn.CrossEntropyLoss() and set the optimizer to use stochastic gradient descent with learningRate with optimize=torch.optim.SGD(model.parameters(),lr=learningRate).

Training the Model#

In each epoch, enumerate through the training data and store the index and (image,label) tuple. Set the images with img=Variable(img.view(-1,28*28)) and labels with label=Variable(labels). Next, reset the gradients to 0 with optimize.zero_grad(). Complete the forward pass with output=model(img) and record the output. Calculate the loss with loss=crit(output,labels), where crit is the cross entropy loss defined above. Then perform the backwards step with loss.backwards(). Lastly, update the weights of the model with optimize.step().

Results#

Track the number of correct labels and the number of total attempts to classify the test data. For each image and label in the test data, get the image (same as in training the model) and run the model with output=model(img). Next, find the predicted values using _,pred = torch.max(output.data) and increment the total images processed. Increment the correct predictions if pred==labels.

Digit Classification Results