params-flow 0.6.2

Tensorflow Keras utilities for reducing boilerplate code.

params-flow provides an alternative style for defining your Keras model or layer configuration in order to reduce the boilerplate code related to passing and (de)serializing your model/layer configuration arguments.

params-flow encourages this:

import params_flow

class MyDenseLayer(params_flow.Layer):      # using params_flow Layer/Model instead of Keras ones
  class Params(params_flow.Layer.Params):   # extend one or more base Params configurations
    num_outputs = None                      # declare all configuration arguments
    activation = "gelu"                     #   provide or override super() defaults
                                            # do not define an __init__()

  def build(self, in_shape):
    self.kernel = self.add_variable("kernel",
                                    [int(in_shape[-1]),
                                     self.params.num_outputs])     # access config arguments

which would be sufficient to pass the right configuration arguments to the super layer/model, as well as take care of (de)serialization, so you can concentrate on the build() or call() implementations, instead of writing boilerplate code like this:

from tf.keras.layers import Layer

class MyDenseLayer(Layer):
  def __init__(self,
               num_outputs,            # put all of the layer configuration in the constructor
               activation = "gelu",    #     provide defaults
               **kwargs):              # allow base layer configuration to be passed to super
    self.num_outputs = num_outputs
    self.activation = activation
    super().__init__(**kwargs)

  def build(self, in_shape):
    self.kernel = self.add_variable("kernel",
                                    [int(in_shape[-1]),
                                     self.num_outputs])      # access config arguments

  def get_config(self):                # serialize layer configuration, __init__() is the deserializer
    config = {
      'num_outputs': self.num_outputs,
      'activation': self.activation
    }
    base_config = super().get_config()
    return dict(list(base_config.items())) + list(config.items())

LICENSE

MIT. See License File.

Install

params-flow is on the Python Package Index (PyPI):

pip install params-flow

Usage

params-flow provides a Layer and Model base classes that help reducing common boilerplate code in your custom Keras layers and models.

When subclassing a Keras Model or Layer, each configuration parameter has to be provided as an argument in __init__(). Keras relies on both __init__() and get_config() to make a model/layer serializable.

While python idiomatic this style of defining your Keras models/layers results in a lot of boilerplate code. params-flow provides an alternative by encapsulating all those __init__() configuration arguments in a dedicated Params instance (Params is kind of a “type-safe” python dict - see kpe/py-params). The model/layer specific configuration needs to be declared as a nested Model.Params/Layer.Params subclass, and your model/layer have to subclass params_flow.Model/params_flow.Layer instead of the Keras ones:

class BertEmbeddingsLayer(Layer):
  class Params(PositionEmbeddingLayer.Params):
    vocab_size              = None
    token_type_vocab_size   = 2
    hidden_size             = 768
    use_position_embeddings = True

class TransformerEncoderLayer(Layer):
  class Params(TransformerSelfAttentionLayer.Params,
               ProjectionLayer.Params):
    intermediate_size       = 3072
    intermediate_activation = "gelu"

this allows you to declare the model’s configuration by simply extending the Params of the underlying layers:

class BertModel(Model):
  class Params(BertEmbeddingsLayer.Params,
               TransformerEncoderLayer.Params):
    pass

N.B. The two code excerpts above are taken from kpe/bert-for-tf2, so check there for the details of a non-trivial params-flow based implementation (of BERT).

Resources

• kpe/py-params - A “type-safe” dict class for python.
• kpe/bert-for-tf2 - BERT implementation using the TensorFlow 2 Keras API with the help of params-flow for reducing some of the common Keras boilerplate code needed when passing parameters to custom layers.