In the previous video, you saw how you could stack
a convolutional layer with LSTMs and bidirectional LSTMs
to do sequence prediction. In this video, we'll go
through a workbook for this, which you can then try
for yourself later. As always, let's check if we
have tensorflow installed. If it isn't, install
the latest either nightly or the latest
release from tensorflow.org. Once you know you
have version two, then this code will generate the synthetic time
series for you, and this code will
turn the array of data into a dataset for training. Note that we've expanded the dimensions on the first line. This helper function can perform the forecasting for
us after training. This first copy of the neural network has
run for a quick 100 Epoch run to try and pick the optimum learning rate
for the optimizer. When it's done, we'll plot
the results and we'll see the 10 to the minus
5 is the optimum value, so we'll set that on the SGD
for the next training run. We'll train for
500 Epochs now and keep an eye on
our loss and our mae. When it's done, we'll run our forecast and
plot the results. Already the curve is
looking much better and the plateau that we'd seen in previous training has vanished, and our mae is low, it's just above five
at the validation set. Finally, if we plot
our training loss and mae, we also see a healthy
curve downwards.