Hi, you probably use a newer version of Keras. From 2.2.4 import was changed for layers. Try tensorflow.keras.layers instead. Also: https://stackoverflow.com/questions/55324762/the-added-layer-must-be-an-instance-of-class-layer-found-tensorflow-python-ke

First of all, you’re right that in the MNIST dataset we have digits from 0 to 9. However, in this exercise I wanted to perform binary classification, which means choosing between two classes. I chose 0s and 1s and eliminated other digits from the MNIST dataset.
Then, as for this line of code: keras.layers.Dense(1, input_shape=(784,), activation=’sigmoid’). I used Dense layer to get input vector of 784 data points and squeeze it to a range (0;1) using sigmoid and then output one value only. The first parameter of Dense method is described as “dimensionality of the output space” – in our case is a single number, hence we have “Dense(1, …)”.
You may also wonder why we use Dense layer for logistic regression. Well, while performing a logistic regression we want to first calculate the polynomial value of: x1 * w1 + x2 * w2 + … + x784 * w784 and this is exactly what Dense layer does.