CODEWITHSUNDEEP
pythontensorflowkerasdeep-learninglstm
KyuMirthu
KyuMirthu

Reputation: 427

Tensoflow2 LSTM - Unused parameter input_shape?

So I have build Neural Network with the following code:

import tensorflow as tf

tf_model = tf.keras.Sequential()
tf_model.add(tf.keras.layers.LSTM(50, activation='relu'))
tf_model.add(tf.keras.layers.Dense(20, activation='relu'))
tf_model.add(tf.keras.layers.Dense(10, activation='relu'))
tf_model.add(tf.keras.layers.Dense(1, activation='linear'))
tf_model.compile(optimizer='Adam', loss='mse')

My training set is shaped as follows:

>> ts_train_X.shape
(16469, 3, 21)

I have read numerous articles and questions here on stackoverflow in order to bring the data frame in the right shape for the LSTM. Almost every of the pages I found specified the input_shape parameter and passed it either to LSTM(..) or Sequential(..).

When I look at the LSTM API I cannot find a reference to this parameter. I also had a glimpse on the source code and to me it seems that the shape is somehow automatically inferred, but I am not sure about this.

This leads me to my question: Why does my code work? How can the LSTM layer as the first layer know the shape of my inputs, if I don't specify the input_shape parameter?

edit: change title as per suggestion in comments.

Upvotes: 2

Views: 467

Answers (1)

Dr. Snoopy
Dr. Snoopy

Reputation: 56417

The parameter input_shape can be given to the constructor of any keras Layer subclass, as this is how the API is defined.

The code works because input_shape is passed as a keyword argument (the **kwargs), then these keyword arguments are passed by the LSTM constructor to the Layer constructor, which then proceeds to store the information for later use. This effectively means that the input_shape parameter does not have to be defined in each layer, and it is passed as a keyword argument instead.

I think the issue is that since keras has been moved to tensorflow, the documentation might not be complete. You can find more information about the input_shape parameter in the Guide to the Sequential API.

Upvotes: 3

Related Questions