PlutoCon 2021 WYSIWYR Demo (MNIST)

In this notebook we will load a pretrained model for classifying MNIST handwritten digits from 28x28 greyscale images. However, this notebook is less about the model and more about interacting with it via "what you see is what you REST" features.

Loading Data

To start off we will download the MNIST dataset using the MLDatasets package.

Now we load a pre-trained model which has been serialized with Julia's native serialization library. The model is made up of 3 convolutional layers, 3 max pooling layers, and one dense layer.

To test our model we will only load in the test data. Our model was trained with training data from MNIST.traindata() in another notebook.

test_x shape: (28, 28, 10000), test_y shape: (10000,)

First we assign a variable input_images to a small slice of test data.

Start Index:

End Index:

For example, the first (and only) element in the sample is a 7

Passing our input_images through the model loaded earlier, we get a 10x1 matrix, where each column corresponds to an input image, and each row corresponds to the class which the model thinks the image corresponds to. For example, a high value in the first row corresponds to a high confidence that the image contains a 0 digit.

The highest value by far is in the 8th index, which corresponds to the model predicting a 7 digit.

The last step is to convert these predictions into numbers, then compare them to their true labels

Finally we can measure the accuracy of the model by comparing our predictions to the actual labels and finding the average.