Generating Images with CNN: Why is the Size Not Increasing in Each Layer?

  • Context: Python 
  • Thread starter Thread starter BRN
  • Start date Start date
  • Tags Tags
    Images
Click For Summary

Discussion Overview

The discussion revolves around generating images using a Convolutional Neural Network (CNN) model, specifically focusing on the issue of output size not increasing as expected in each layer of a deconvolutional (transpose convolution) setup. The scope includes technical explanations and model adjustments related to neural network architecture.

Discussion Character

  • Technical explanation
  • Debate/contested

Main Points Raised

  • One participant describes their model setup and expresses confusion about consistently obtaining an 8x8x3 image instead of the desired 128x128x3.
  • Another participant provides a solution involving the correct equations for output size in transpose convolutions, suggesting adjustments to the model's parameters, including the stride and kernel size.
  • The initial poster acknowledges the proposed solution and expresses satisfaction with the resolution of their issue.
  • A later reply suggests that expanding acronyms like CNN could be beneficial for readers who may not be familiar with the terminology.
  • Another participant confirms the meaning of CNN as Convolutional Neural Network and agrees to provide clearer explanations in future posts.

Areas of Agreement / Disagreement

Participants generally agree on the solution provided for the output size issue, but there is a minor disagreement regarding the clarity of terminology used in the discussion.

Contextual Notes

The discussion includes assumptions about the understanding of CNNs and the implications of different parameters on output size, which may not be fully explored or resolved.

BRN
Messages
107
Reaction score
10
Hello everybody,
I have this problem:
starting from a vector of 100 random values, I have to generate an image of size 128x128x3 using a model consisting of a fully completely layer and 5 layer deconv.
This is my model

Python: 
def generator_model(noise_dim):
   
    n_layers = 5
    k_w, k_h = [8, 8]
    input_dim = (noise_dim,)
    i_w, i_h, i_d = [8, 8, 1024] # starting filters
    strides = (1, 1)
    weight_initializer = None

    model = tf.keras.Sequential()
   
    model.add(tf.keras.layers.Dense(i_w * i_h * i_d, input_shape = input_dim, kernel_initializer = weight_initializer))
    model.add(tf.keras.layers.BatchNormalization())
    model.add(tf.keras.layers.ReLU())
   
    model.add(tf.keras.layers.Reshape((i_w, i_h, i_d)))
    for i in range(n_layers - 1):
        print(k_w, k_h)
        model.add(tf.keras.layers.Conv2DTranspose(i_d, (k_w, k_h), strides, padding = 'same', use_bias = False))
        model.add(tf.keras.layers.BatchNormalization())
        model.add(tf.keras.layers.ReLU())
        i_d = int(i_d / 2)
        k_w = int(k_w * 2)
        k_h = int(k_h * 2)
       
    k_w = i_d
    k_h = i_d
    model.add(tf.keras.layers.Conv2DTranspose(3, (k_w, k_h), strides, padding = 'same', use_bias = False))

    return model

Why do I always get an 8x8x3 image without having an increase in size in each layer?

Thank's
 
Last edited by a moderator:
Technology news on Phys.org
Ok, I have the solution.

The size of the outputs of a CNN "conv" is given by the equation

$$o=\left ( \frac{i - k + 2p}{s} \right )+1$$

but, as in my case, for a transpose convolution "deconv" the size of the outputs is

$$o=s\left (i -1 \right )+ k - 2p$$

Then, with stride ##s=2##, the correct code is this

Python: 
def generator_model(noise_dim):
 
    n_layers = 4
    k_w, k_h = [16, 16] # starting kernel size
    input_dim = (noise_dim,)
    i_w, i_h, i_d = [8, 8, 1024] # starting filters
    strides = (2, 2)
    weight_initializer = None

    model = tf.keras.Sequential()
 
    model.add(tf.keras.layers.Dense(i_w * i_h * i_d, input_shape = input_dim, kernel_initializer = weight_initializer))
    model.add(tf.keras.layers.BatchNormalization())
    model.add(tf.keras.layers.ReLU())
 
    model.add(tf.keras.layers.Reshape((i_w, i_h, i_d)))
    i_d = int(i_d / 2)
    for i in range(n_layers - 1):
#        print(i_d, k_w, k_h)
        model.add(tf.keras.layers.Conv2DTranspose(i_d, (k_w, k_h), strides, padding = 'same', use_bias = False))
        model.add(tf.keras.layers.BatchNormalization())
        model.add(tf.keras.layers.ReLU())
        i_d = int(i_d / 2)
        k_w = int(k_w * 2)
        k_h = int(k_h * 2)

    model.add(tf.keras.layers.Conv2DTranspose(3, (k_w, k_h), strides, padding = 'same', use_bias = False))

    return model

And this solves my problem :smile:
 
Reply
  • Like
Likes   Reactions: Ibix
BRN said:
And this solves my problem :smile:
Great!
In the future, it would be helpful to readers to expand acronyms such as CNN, which might not be generally known. I'm assuming it has something to do with neural networks, but that's only a guess.
 
Reply
  • Like
Likes   Reactions: BRN
Hello and thanks!
You're right, CNN means Convolutional Neural Network.
Next time I will write the acronyms explicitly :smile: