如何在 keras 中将密集层连接到 Conv2D

Question

I want to map an input of 61 intensity values to an image of size 64×64. The code I use is given below.

Network Input=61×1 (intensity values)

OUTPUT=64×64 (image)

input_img = Input(shape=(61,))

x = Dense(250, activation='relu')(input_img)
x = Dense(500, activation='relu')(x)
x = Dense(1000, activation='relu')(x)
x = Dense(4096, activation='relu')(x)

x=Conv2D(16,(3,3),padding='same',kernel_regularizer=regularizers.l2(0.001),kernel_initializer='glorot_uniform')(x)


x=Conv2D(1,(3,3),padding='same',kernel_regularizer=regularizers.l2(0.001),kernel_initializer='glorot_uniform')(x)

The dimensions are giving me problems. How can I shape the dimensions in the code so that I can get correct mapping as 64×64 size at output.

The code error is ValueError: Input 0 is incompatible with layer conv2d_14: expected ndim=4, found ndim=2

Thank you

Answer 1

The probable problem is input_img shape.

It should actually contain 3 dimensions. And internally keras will add the batch dimension making it 4.

Since you used input_img with 1 dimensions (vector), keras is adding the 2nd.

You should correct the shape of your input_img