I'm currently building a GAN with Tensorflow 2 and Keras and noticed a lot of the existing Neural Networks for the generator and discriminator use Conv2D and Conv2DTranspose in Keras.
I'm struggling to find something that functionally explains the difference between the two. Can anyone explain what these two different options for making a NN in Keras mean?
Conv2D applies Convolutional operation on the input. On the contrary, Conv2DTranspose applies a Deconvolutional operation on the input.
For example:
x = tf.random.uniform((1,3,3,1))
conv2d = tf.keras.layers.Conv2D(1,2)(x)
print(conv2d.shape)
# (1, 2, 2, 1)
conv2dTranspose = tf.keras.layers.Conv2DTranspose(1,2)(x)
print(conv2dTranspose.shape)
# (1, 4, 4, 1)
Conv2D is mainly used when you want to detect features, e.g., in the encoder part of an autoencoder model, and it may shrink your input shape.
Conversely, Conv2DTranspose is used for creating features, for example, in the decoder part of an autoencoder model for constructing an image. As you can see in the above code, it makes the input shape larger.
For example:
kernel = tf.constant_initializer(1.)
x = tf.ones((1,3,3,1))
conv = tf.keras.layers.Conv2D(1,2, kernel_initializer=kernel)
y = tf.ones((1,2,2,1))
de_conv = tf.keras.layers.Conv2DTranspose(1,2, kernel_initializer=kernel)
conv_output = conv(x)
print("Convolution\n---------")
print("input shape:",x.shape)
print("output shape:",conv_output.shape)
print("input tensor:",np.squeeze(x.numpy()).tolist())
print("output tensor:",np.around(np.squeeze(conv_output.numpy())).tolist())
'''
Convolution
---------
input shape: (1, 3, 3, 1)
output shape: (1, 2, 2, 1)
input tensor: [[1.0, 1.0, 1.0], [1.0, 1.0, 1.0], [1.0, 1.0, 1.0]]
output tensor: [[4.0, 4.0], [4.0, 4.0]]
'''
de_conv_output = de_conv(y)
print("De-Convolution\n------------")
print("input shape:",y.shape)
print("output shape:",de_conv_output.shape)
print("input tensor:",np.squeeze(y.numpy()).tolist())
print("output tensor:",np.around(np.squeeze(de_conv_output.numpy())).tolist())
'''
De-Convolution
------------
input shape: (1, 2, 2, 1)
output shape: (1, 3, 3, 1)
input tensor: [[1.0, 1.0], [1.0, 1.0]]
output tensor: [[1.0, 2.0, 1.0], [2.0, 4.0, 2.0], [1.0, 2.0, 1.0]]
'''
To sum up:
Conv2D:
Conv2DTranspose:
And if you want to know how Conv2DTranspose enlarges input, here you go:
