You have come a long way! You have learnt how to customise your layers, modes and the training function using the approach of Model subclassing . The most important points to re-iterate are:
tf. keras provides many built-in layers, for example:
- Convolutional layers: Conv1D, Conv2D, etc.
- Pooling layers: MaxPooling1D, MaxPooling2D,
- RNN layers:
GRU,LSTM, etc.
BatchNormalization,Dropout,Embedding, etc.
But if you don’t find what you need, it’s easy to extend the API by creating your own layers. All layers subclass the Layer class and implement:
callmethod, that specifies the computation done by the layer.
buildmethod, that creates the weights of the layer (this is just a style convention since you can create weights in__init__, as well).
- The outer container, the thing you want to train, is a
Model. AModelis just like aLayer, but with added training and serialization utilities.