TensorFlow was released in Nov 2015. It focus on speed, and put ‘easy to program’ on its secondary goal.
Trax – bottom-up clean re-design; easy to debug, you can read the code. But it does not have backwards compatibility.
Trax is built up with layers
Activation layer is an activation function,.
Dense Layer is a metric, fully connects two layers.
Embedding layer embed words into vectors.
Mean layer calculate the metric means.
Training in Trax
Trax has the trax.math.grad function which calculates the grad for the function.