keras.Sequential（）的“ input_shape”格式是什么？

Question

Following this tensorflow tutorial, under the Build the model section, the first layer of keras.Sequential() is given parameter vocab_size=10000 . What does it mean?

After looking at the documentation I figured out that, the Sequential model needs to know the shape of the input tensor(tensor to the input layer), and input_shape attribute defines it. But input_shape should be like (30, 50, 50, 3) (for 30, 50x50 RGB images). The value in the tutorial is 10000.

Answer 1

For the first part of your question:

The example uses an embedding layer. Think of embedding layer as a lookup matrix. Each row will represent a word vector. The vocab_size is identifying the size of this matrix (or in other words number of words represented in in this matrix - which is essentially the number of rows of the matrix).

Now to represent an input string, each word will be represented by an integer - which corresponds to the row of the embedding matrix.

For second part, I am not sure which example you are referring or what you are trying to do. However, you might want to know that the number of samples in the batch will not be a part of input_shape. So the input_shape to pass a batch of 30 samples of 50*50 RGB images will be (50,50,3)

Answer 2

Sequential is just a way to define the model, a sort of wrapper, if you want.

It wraps a sequence of layers. The first layer needs an input size.

the input shape will be different for each element you are studying: you are correct that the input shape for an image could be (50, 50, 3).

In the example you provided, the first layer handles word embeddings: the embedding shape is defined by the function that creates the data (lower in your example page there is imdb.load_data() )

the vocab_size is not the full shape, but how many "rows" of embeddings you are going to treat as a single input.

You can see a clearer example here: https://keras.io/layers/embeddings/