9  Running the AutoEncoder for the first time

Above is pretty much my reaction when I trained for the first time the AE. The loss did converges but it wasn’t near zero at all so the model was not learning properly.

To help visualizing what was going on I added a piece of code that would print the current batch, synthetic and then real data.

1if np.random.uniform(size=1) > 0.99:
2  print("synthetic")
  print(tf.round(synthetic).numpy())
3  print("real")
  print(real.numpy())

1

Every 1 out 100 batches do

2

Print the reconstructed data

3

Print the original data

I then noticed how the generated data initially had more frequent events as 1 for all of the records then after some time they suddenly become 0 one after the other. This was clearly not ideal.