Statistical Analysis Stack Exchange
Statistical Analysis Stack Exchange

Questions tagged [transposed-convolution]

Transposed convolution (a.k.a. deconvolution) is an upsampling operation in a neural network that works by swapping the forward and backward passes of a convolution.

21 questions
17
votes
4 answers

In CNN, are upsampling and transpose convolution the same?

Both the terms "upsampling" and "transpose convolution" are used when you are doing "deconvolution" (<-- not a good term, but let me use it here). Originally, I thought that they mean the same thing, but it seems to me that they are different after…
RockTheStar
  • 11,277
  • 31
  • 63
  • 89
3
votes
1 answer

Why do DeconvNet use ReLU in the backward pass?

Why does DeconvNet (Zeiler, 2014) use ReLU in the backward pass (after unpooling)? Are not the feature maps values already positive due to the ReLU in the forward pass? So, why do the authors apply the ReLU again coming back to the input? ref:…
volperossa
  • 625
  • 5
  • 9
2
votes
0 answers

Should convolutions or transposed convolutions be used in the decoder part of a Conv-based autoencoder?

I am implementing a convolutional autoencoder. For the decoder part of the model, some examples (such as this one from Francois Chollet) use standard convolutional layers (Conv2D in keras) in the decoder part of the model (in combination with an…
user3825755
  • 183
  • 4
2
votes
1 answer

Mismatching dimensions of input/output in the WaveNet model for text-to-speech generation?

I have been trying to understand the model of how speech generation works, particularly in WaveNet model by Google. I was referring to the original WaveNet paper and this implementation: I find the model very confusing in the input it takes and the…
Joe Black
  • 299
  • 1
  • 10
2
votes
0 answers

How do upsampling layers work for coarse-to-fine output in semantic segmentation?

Here is a figure illustrating the Fully Convolutionnal Network (FCN) of the Fully Convolutionnal Paper for Semantic Segmentation : The upsampling layer at the end confuses me. I cannot understand how it learns. I know how it works technically, I…
Soltius
  • 808
  • 6
  • 12
2
votes
0 answers

How to define a loss function for discrete fourier series?

In each batch there are 8000 sample points, and I apply discrete Fourier transform on them. The original samples are real valued, so only the half of the result is needed. The end result is 4000 complex numbers. I would like to build an autoencoder…
Iter Ator
  • 123
  • 4
1
vote
1 answer

Why does the number of filters does not determine the output size in the CNN layer?

The formula to determine the output size is [(W−K+2P)/S]+1. Here the number of filters used is not significant for determining the ouput size, I was wondering why? Is it because Each filter is separately applied to the input to obtain a feature map…
user9683713
  • 13
  • 3
1
vote
0 answers

Explanation of effect of bias in deconvolution

I've been reading the deconvolution article on distill. I am not able to figure out the meaning of the text These artifacts tend to be most prominent when outputting unusual colors. Since neural network layers typically have a bias (a learned value…
pramesh
  • 123
  • 5
1
vote
0 answers

Reusing Weights in Transposed Convolution

As far as I know it's possible to reuse the weights of a convolution in a transposed convolution to upsample an image. However when reusing the weights, the resulting restored images aren't even close to the original…
Tom Dörr
  • 331
  • 1
  • 5
1
vote
0 answers

Transposed convolution layer in a CNN

So I've seen a convolutional neural network that uses the traditional convolutional layers, but then after the down-sampling of the input and some pooling layer, they've added a transposed convolution layer which up-sample the resulting…
MercyDude
  • 111
  • 5
1
vote
0 answers

How to pad skip connections when using transposed / deconvolutional layers

if you have a standard CNN architecture with convolutional layers there are 2 reasons why the identity of the skip connection can't be added with the current output. 1) There was pooling between the identity and the current output 2) The amount of…
Peter111
  • 158
  • 4
1
vote
0 answers

calculate Inception Score and FID in GAN

Do we calculate Inception Score and Fréchet Inception Distance (FID) on the images generated by generator in parallel during training ? Or do we save the images generated by the generator and later calculate the score ?
Bishwa Karki
  • 111
  • 3
1
vote
0 answers

How to add bias in convolution transpose?

My question is regarding the transposed convolution operation (also commonly called deconvolution or upconvolution). In TensorFlow, for instance, I refer to this layer. My question is, how / when do we add the bias (intercept) term when applying…
James Bond
  • 163
  • 7
1
vote
1 answer

Concatentation of feature maps in U-net

I am looking at the following snippet of code: c4 = Conv2D(64, (3, 3), activation='relu', padding='same') (p3) c4 = Conv2D(64, (3, 3), activation='relu', padding='same') (c4) p4 = MaxPooling2D(pool_size=(2, 2)) (c4) c5 = Conv2D(128, (3, 3),…
Christian
  • 1,382
  • 3
  • 16
  • 27
1
vote
0 answers

Short Sentence Generation using CNNs

I am investigating whether building a classifier for sentence classification using CNN can be used for sentence generation. Say, we are classifying news articles' titles (classes such as sports, business, etc.). The question is whether the same…
Amir
  • 379
  • 2
  • 8
1
2