Embedding inside the model vs outside the model-CodePudding

What is the difference between using the embedding layer inside the model and outside the model? I can build the embedding layer into the model:

model = tf.keras.Sequential()
model.add(tf.keras.layers.Embedding(input_dim=1000, output_dim=64, input_length=10))
...
model.fit(features, target ...)

I can also use embdedding outside the model to generate embedded data and then feed it into the model:

embedding_encoder = tf.keras.layers.Embedding(input_dim=1000, output_dim=64)
embedded_features = embedding_encoder(features)
...

model.fit(embedded_features, target ...)

Does it mean that if I use embedding outside the model, embedding parameters are not learned during the training?

CodePudding user response：

Does it mean that if I use embedding outside the model, embedding parameters are not learned during the training?

The dense vector representations assigned from an Embedding layer are generally only trainable when setting trainable=True. It's entirely up to you how you want to preprocess your data yourself and how much you want to leave to the Embedding layer. Usually, if you are working on a NLP task, you can add a StringLookup or TextVectorization layer prior to adding an Embedding layer that allows you to preprocess your texts and train them in an elegant way without any "manual" steps.

Generally:

Each integer value fed to an Embedding layer is mapped to a unique N-dimensional vector representation, where N is chosen by you. These vector representations are, by default, drawn from a uniform distribution. The Embedding layer inherits from tf.keras.layers.Layer which contains the trainable parameter.

I think it could make sense to generate embedding data outside your model if you are, for example, using pretrained context-sensitive vectors and you do not want to update their values during training. But again, it’s all up to you.