Trjtdtk

Colliding particles and Activation energy

Stark VS Thanos

What is the difference between `a[bc]d` (brackets) and `ab,cd` (braces)?

Pawn Sacrifice Justification

Past Perfect Tense

If Earth is tilted, why is Polaris always above the same spot?

How to figure out whether the data is sample data or population data apart from the client's information?

Find the coordinate of two line segments that are perpendicular

Why does Bran Stark feel that Jon Snow "needs to know" about his lineage?

Help, my Death Star suffers from Kessler syndrome!

A question regarding using the definite article

Where does the labelling of extrinsic semiconductors as "n" and "p" come from?

Is GOCE a satellite or aircraft?

Where did the extra Pym particles come from in Endgame?

Need help understanding harmonic series and intervals

Binary Numbers Magic Trick

Weird result in complex limit

Were there two appearances of Stan Lee?

Pressure to defend the relevance of one's area of mathematics

How to create an ad-hoc wireless network in Ubuntu

Is it cheaper to drop cargo drop than to land it?

Subtleties of choosing the sequence of tenses in Russian

Unexpected email from Yorkshire Bank

Why do computer-science majors learn calculus?

Dimension of feature vectors for classification task in the DCGAN paper

What are deconvolutional layers?How are 1x1 convolutions the same as a fully connected layer?My first machine learning experiment , model not converging , tips?Strange patterns from GANUnderstand the shape of this Convolutional Neural NetworkHow to change learning rate of MomentumOptimizer in tensorflowGANs and grayscale imagery colorizationExtracting Useful features from large convolutional layersHow to calculate $phi_i,j$ in VGG19 network?What is the motivation for row-wise convolution and folding in Kalchbrenner et al. (2014)?

0

$begingroup$

I am trying to implement one of the section in the DCGAN paper (https://arxiv.org/pdf/1511.06434.pdf) i.e. Using the Discriminator network trained on ImageNet-1k as a feature extractor to classify images in the CIFAR-10 dataset. However, I could not understand how to compute the dimension of 28672 of the feature vectors after concatenating outputs from all convolutional layers of the Discriminator.

I followed the paper and implement a maxpooling layer of size and stride (2, 2) (the paper did not specify this.) after each convolutional layer in the Discriminator network (there are 5 layers in total). However, after concatenating the dimension of the vector that I got is a dimension of $30,720$ as follows:

$$16 times 16 times 64 + 8 times 8 times 128 + 4 times 4 times 256 + 2 times 2 times 512 + 1 times 1 times 1= 16,384 + 8,192 + 4,096 + 2,048 + 1= 30,721.$$

If I just take the 3 first layers, then the dimension is $28,672$ as specified in the paper. However, the authors have said that:

To evaluate the quality of the representations learned by DCGANs for supervised tasks, we train on Imagenet-1k and then use the discriminator’s convolutional features from all layers, maxpooling each layers representation to produce a 4 × 4 spatial grid. These features are then flattened and concatenated to form a 28672 dimensional vector and a regularized linear L2-SVM classifier is trained on top of them.

So I think I have got something wrong. Would anyone please kindly explain for me how to get the dimension of the vectors as mentioned in the paper?

Thank you so much in advance for your help!

convnet feature-extraction convolution gan

share|improve this question

asked Apr 8 at 15:58

SophilSophil

1

$endgroup$

• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  Has anyone got any idea of this? I really appreciate your help!
  $endgroup$
  – Sophil
  Apr 10 at 8:47
  

  
  
  
  

add a comment | 

0

$begingroup$

I am trying to implement one of the section in the DCGAN paper (https://arxiv.org/pdf/1511.06434.pdf) i.e. Using the Discriminator network trained on ImageNet-1k as a feature extractor to classify images in the CIFAR-10 dataset. However, I could not understand how to compute the dimension of 28672 of the feature vectors after concatenating outputs from all convolutional layers of the Discriminator.

I followed the paper and implement a maxpooling layer of size and stride (2, 2) (the paper did not specify this.) after each convolutional layer in the Discriminator network (there are 5 layers in total). However, after concatenating the dimension of the vector that I got is a dimension of $30,720$ as follows:

$$16 times 16 times 64 + 8 times 8 times 128 + 4 times 4 times 256 + 2 times 2 times 512 + 1 times 1 times 1= 16,384 + 8,192 + 4,096 + 2,048 + 1= 30,721.$$

If I just take the 3 first layers, then the dimension is $28,672$ as specified in the paper. However, the authors have said that:

To evaluate the quality of the representations learned by DCGANs for supervised tasks, we train on Imagenet-1k and then use the discriminator’s convolutional features from all layers, maxpooling each layers representation to produce a 4 × 4 spatial grid. These features are then flattened and concatenated to form a 28672 dimensional vector and a regularized linear L2-SVM classifier is trained on top of them.

So I think I have got something wrong. Would anyone please kindly explain for me how to get the dimension of the vectors as mentioned in the paper?

Thank you so much in advance for your help!

convnet feature-extraction convolution gan

share|improve this question

asked Apr 8 at 15:58

SophilSophil

1

$endgroup$

• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  Has anyone got any idea of this? I really appreciate your help!
  $endgroup$
  – Sophil
  Apr 10 at 8:47
  

  
  
  
  

add a comment | 

$begingroup$

I am trying to implement one of the section in the DCGAN paper (https://arxiv.org/pdf/1511.06434.pdf) i.e. Using the Discriminator network trained on ImageNet-1k as a feature extractor to classify images in the CIFAR-10 dataset. However, I could not understand how to compute the dimension of 28672 of the feature vectors after concatenating outputs from all convolutional layers of the Discriminator.

I followed the paper and implement a maxpooling layer of size and stride (2, 2) (the paper did not specify this.) after each convolutional layer in the Discriminator network (there are 5 layers in total). However, after concatenating the dimension of the vector that I got is a dimension of $30,720$ as follows:

$$16 times 16 times 64 + 8 times 8 times 128 + 4 times 4 times 256 + 2 times 2 times 512 + 1 times 1 times 1= 16,384 + 8,192 + 4,096 + 2,048 + 1= 30,721.$$

If I just take the 3 first layers, then the dimension is $28,672$ as specified in the paper. However, the authors have said that:

To evaluate the quality of the representations learned by DCGANs for supervised tasks, we train on Imagenet-1k and then use the discriminator’s convolutional features from all layers, maxpooling each layers representation to produce a 4 × 4 spatial grid. These features are then flattened and concatenated to form a 28672 dimensional vector and a regularized linear L2-SVM classifier is trained on top of them.

So I think I have got something wrong. Would anyone please kindly explain for me how to get the dimension of the vectors as mentioned in the paper?

Thank you so much in advance for your help!

convnet feature-extraction convolution gan

share|improve this question

asked Apr 8 at 15:58

SophilSophil

1

$endgroup$

share|improve this question

asked Apr 8 at 15:58

SophilSophil

1

share|improve this question

asked Apr 8 at 15:58

SophilSophil

1

asked Apr 8 at 15:58

SophilSophil

1

asked Apr 8 at 15:58

SophilSophil

1