In my previous blog post I described how to Build a SAP Data Intelligence ML Scenario with TensorFlow. In this blog post I show how to improve it based on insights from TensorBoard.
When launching it for the first time, I get an error message, because there are no event files yet:
It takes only one small addition to my previous Python training code to add this though:
history = model.fit(dataset, epochs=64, callbacks=[tf.keras.callbacks.TensorBoard('/vhome/tf')])
As a result, I get my loss graphicly displayed by epoch:
As it appears, I seem to have over trained my model since it becomes unstable from around epoch 50 on. So how about reducing it to 32 epochs to be on the safe side:
model = tf.keras.Sequential( [ tf.keras.Input(name='input', batch_size=1, shape=(4,)), tf.keras.layers.Dense(8, name='hidden', activation=tf.nn.relu), tf.keras.layers.Dense(8, name='relu', activation=tf.nn.relu), tf.keras.layers.Dense(3, name='output') ] ) model.compile(tf.keras.optimizers.SGD(), tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)) model.summary() history = model.fit(dataset, epochs=32, callbacks=[tf.keras.callbacks.TensorBoard('/vhome/tf')])
That looks more stable and in fact, due to the added labels, I got a nice graphical representation of my model as a bonus as well:
While this is a simple example, I hope it gives you an idea how TensorBoard might help you improving your SAP Data Intelligence ML Scenario.