Designing the Model for Emojify_V2

Course Q&A Deep Learning Specialization Sequence Models
1

Hello,
In Emojify_V2 of the EMojify assignment, we were presented with a Neural Network design to use.

emojifier-v2
emojifier-v21666×1094 71 KB 

How was the network designed?
How did the designers know we know we need 2 layers of LSTM?
How did the designers know we know we need a dropout?
What is the general approach when you start trying to design a network?

Thanks in Advance!

2

There is no fixed algorithm for NN architecture. Packages like keras tuner are used to compare different architectures before finalizing a model architecture.

Here are some points (based on Andrews advice):

  1. See if an existing model satisfies your requirements.
  2. If it doesn’t, see how parts of an existing model architecture can be reused or customized for your problem.
  3. If that doesn’t work, invent your architecture from scratch.