A subset of a growing collection of productivity tools that I use when working with Keras layers. Useful for manipulating namespaces (classes with only static methods). Has methods for converting functions to lazily evaluated and/or cached methods. Has methods to convert functions to a few sorts of factory methods. Useful for e.g., converting a namespace of neural network topologies to objects that mimic the behavior of Keras layers.
Example snippets for each module.
import lazy
@lazy.lazy_evaluation(lazy=True, cached=True)
def your_method...
The decorated method returns a Data wrapper object. Lazily evaluated and/or cached data can be accessed by the value property of the returned Data object.
import namespaces
@namespaces.namespace_to_callable_factory
class MyNetworks:
def network1(*args, **kwargs):
...
return keras.layers.some_layer(some_args)(input)
def network2(*args, **kwargs):
...
return keras.layers.some_layer(some_args)(input)
The decorated namespace methods will mimic the behavior of Keras layers. Each method should return a model defined by using the functional API.
import layer_group
from tensorflow import Tensor
from tensorflow.keras import Model
def op(*args, **kwargs) -> Tensor:
...
MyLayer = LayerGroup("MyLayer", op)
model = Model(inputs=[inp], outputs=[MyLayer(inp)])
model.summary()
Model: "model"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) [(None, 5)] 0
_________________________________________________________________
my_layer (MyLayer) (None, 50) 300
=================================================================
Total params: 300
Trainable params: 300
Non-trainable params: 0
_________________________________________________________________
Your op will appear as a compact layer group, and will no longer pollute summaries and graph plots.