In Development (as of 14 April 2021)¶
The fact that TensorDiffEq is built on top of Keras allows for some unique properties when training PINNs. So far, a (non-exhaustive) list includes:
easy modification of neural network architecture with ANY Keras layers (whether or not they are useful for your model is left up to you)
easy modification of optimizers using the
tf.keras.optimizersbank of solvers
exporting and re-importing a model for later use or transfer learning
Here we will discuss a few of these options and how to execute these ‘hacks’. Please contribute more if you discover any!