one-hot encoding not working properly with logistic regression

Question

I am trying to train a logistic regression model to recognize handwritten English letters. In my test data I have 74880 images. Each image has 784 pixels. The labels corresp