Question about Sequence Models Week 2 Emojify

Course Q&A Deep Learning Specialization Sequence Models
1

In the assignment last step( Exercise 5 - constructing the Emojify_V2 function), we were supposed to construct multiple(2) layers of LSTM as shown in the figure

image
image739×447 38.4 KB

In have coded correctly and passed all the tests but just have some unclear points about the structure.

  1. to confirm the the input shape
    The model takes as input an array of sentences of shape (m , max_len , ) defined by input_shape —> m should be the batch size, max_len = Tx, and last dim should be the embedding dim, if we are using 50 embedding dim, input will be m,Tx,50?
  2. The figure shows that we have Tx LSTM units in the first and second layer, but in the code function, only one LSTM with 128 units are coded. So my confusion is that does the keras layer automatically unroll the single LSTM unit Tx times (since they are sharing the same weights in the same layer)? Because I didnt see any for loop like for i in range(Tx) to loop all the Tx inputs

Any clarification appreciated, thanks in advantage!

2
  1. Yes
  2. The number of units is not the same as Tx (max_len). If we have 128 units and Tx = 10 and if we input only one word at each time step, then all 128 LSTM units simultaneously contribute to processing the input word.

Moreover, it is mentioned in the instructions that you have to specify the number of units in the 2nd LSTM layer as well.

3

That clears out all my confusion! :grinning: Thanks for your reply Saif.