My DNN model to compare MNIST dataset using tensorflow not making progress

Question

I am doing exercise of DNN chapter from "Hands on machine learning using scikit and tensorflow" book. Below are the steps I am taking

Steps

• Generate new data set from MNIST such that each row has two images and its label is 1 if the images represent same digit else 0. 50% rows are of same digit and other 50% are of different digits
• create 2 DNNs dnn1 and dnn2 with 5 hidden layers and not output layer. Each DNN takes one image as input
• Join output of dnn1 and dnn2 and feed it to a layer with 10 neurons
• add last layer with single layer

Here is the link of google colab where I implemented it. It is not making any progress when training. But the solution provided by book is working, here is the google colab link. My implementation and the books implementation is slightly different but they are similar with same steps. I have spent two days trying to figure out what is different but I am not able to find it.

May be I have done a silly mistake, It would be great help to me if someone can point me what it is. I am new to tensorflow, someone with experience in it may easily see it.

My implementation in google colab
Books solution in google colab

Answer 1

It doesn't look like you normalized your data anywhere. The book solution does this right at the beginning:

X_train = X_train.astype(np.float32).reshape(-1, 28*28) / 255.0
X_test = X_test.astype(np.float32).reshape(-1, 28*28) / 255.0