Week4 Question to the padding of Conv2Dtranspose

Course Q&A Deep Learning Specialization Convolutional Neural Networks
1

I read the question in the following link.

In the Answer of the link,the Mentor said that same padding is used in order not to shrink the volume during the upsampling. I understand that this is true to the conv2d’s padding in the upsampling block,but is it also true to the conv2Dtranspose’s padding? I think in the transpose convolution,we can’t get the smaller output than the input tensor,So why the “same” padding is used in the conv2dtranspose?

2

Your understanding is correct. Conv2DTranspose output can’t be smaller than input. As far as padding parameter goes, this link should help understand how output size is a scaled version of the input dimension.

3

I was running into the same question too, and from a quick search, I was surprised by the lack of thorough explanation of what this is. Here I’m trying my best:

On tensorflow doc, “same” results in padding evenly to the left/right or up/down of the input. When padding=“same” and strides=1, the output has the same size as the input.

  • Normally for transpose convolution, we have output_size = stride * (input_size - 1) + Kernel - 2 * padding. As a side note, padding controls how much space is added around the output. There is some further explanation here.
  • The doc says

“The output has the same size as the input”

This is inconsistent with what I observed. I ran an experiment below. The behavior I saw was it’s trying to keep output_size = stride * input_size

Experiment:

import tensorflow as tf

t = tf.constant([
    [
        [0, 1],
        [5, 6],
    ]
], dtype=tf.float32)

t = tf.expand_dims(t, axis=-1)  # Now t has shape (1, 2, 2, 1)

# Needs 4d input, channels_last is chosen. That's why we need an input of (1,2,2,1)
conv = tf.keras.layers.Conv2DTranspose(
    filters=1,
    kernel_size=3,
    strides=(2, 2),
    padding='same',
    data_format='channels_last',
    dilation_rate=(1, 1),
    activation=None,
    use_bias=False,
    kernel_initializer='he_normal',
    kernel_regularizer=None,
    bias_regularizer=None,
    activity_regularizer=None,
    kernel_constraint=None,
    bias_constraint=None,
)(t)
conv.shape

I see (1,4,4,1) as output. I tried inputs with other shapes, such as (2,3), (3,3), the behavior output_size = stride * input_size is consistent. I was inspired by ChatGPT o1-preview, so please do point out if there’s any incorrectness here (I couldn’t find more detailed explanation on this though)

@ai_curious I saw a nice response from you in 2021. As of 2024, the output padding part has been removed from the TensorFlow API

4

Agree that post is obsolete due to evolution of the API in the last 3 years. I have edited it to reflect the change.