Layers-Sequential, input_shape, Flatten, Dense

• Sequential

- To construct with a sequential structure, a stack from the input layer to the output layer.

- It lets stack the layers one by one.

- It connects several functions such as Wx+b, Sigmoid.

 

• model.add()

- Adding the layer-an input layer, a hidden layer, and an output layer

- You can construct the model using the various layers(Dense, Conv2D, MaxPooling2D, LSTM, etc.).

 

• input_shape

- Only specified in the first layer, and the subsequent layers automatically estimate the shape of the input data.

- The tuple must be entered.

---

• 2d data

input_shape=(num_features,)

 

For example,

28*28 black and white by each images can be represented

model = Sequential()
model.add(Dense(256, activation='relu', input_shape=(28*28,)))

---

• 3d data

input_shape=(num_time_steps, num_features)

 

num_time_steps : the number of time steps in sequential data

 

For example,

390 : time stages per day

3 : each time stages measure 3 characteristics such as stock price, trading volume, index.

 

 

model = Sequential()
model.add(LSTM(256, input_shape=(390, 3)))

---

• 4d data

Typically, color image data.

input_shape=(height, width, channel)

 

model = Sequential()
model.add(Conv2D(32, (3, 3), activation='relu', input_shape=(32, 32, 3)))

 

• Flatten

Flatten the input data to one dimension, converting a 2D or 3D feature map to a 1D vectorwithout changing the format. Therefore, it does not have weights.

model = tf.keras.Sequential()
model.add(tf.keras.layers.Conv1D(32,3, input_shape=(100,1)))
model.output_shape#(None, 98, 32)

model.add(MaxPooling1D(2))
model.output_shape#(None, 49, 32)

model.add(Flatten())
model.output_shape#(None, 1568) <- 49*32 = 1568

 

• Dense

The structure that both input and output neurons are connected.

In image classification, it converts 1d vector by Flatten layer, and then add Dense layer for class probability. 

 

128 : num of output neurons.

tf.keras.Sequential([
        tf.keras.layers.Flatten(input_shape=(32, 32, 3)),
        tf.keras.layers.Dense(128, activation='relu')