Working of DCGAN


  1. I did not quite understand the architecture of the generator mentioned in the notebook. (image in notebook)

  2. I also did not quite understand the statement for the generator
    “You may notice that instead of passing in the image dimension, you will pass the number of image channels to the generator. This is because with DCGAN, you use convolutions which don’t depend on the number of pixels on an image. However, the number of channels is important to determine the size of the filters.”
    I did understand to get a image out of a noise vector we use deconvolution ( for increasing the size of the image from the low level noise)
    But , I did not quite understand how we get it just by passing the number of channels.

Hi @Prasanna_Hegde,

Please consider that the size of the filters is not determined by the image dimensions (pixels), but by the number of channels. In summary, this helps to get the proper size of the filters. Where each channel, corresponds to a particular aspect or feature in the input data.

@Prasanna_Hegde, a little about your question #1 about the image in the notebook:

The main thing to keep in mind is that this is a diagram from a paper that was using different values in its model, so it’s best to think of it as just giving the general idea of what’s going on, rather than any specifics. For example:

1). The diagram shows the generator starting with a noise vector and going through several transpose convolution blocks to finally create an image-sized output with the appropriate number of channels. (In the case of the diagram, the final output is a 64x64 image, with 3 channels. In the assignment, the final output is a 28x28 image with one channel (since our assignment is for a black&white image)).
2). The diagram shows the general pattern of the “image” size increasing for each block while the number of channels decreases, with the image going from 4x4->8x8->16x16->32x32->64x64 and channels going from 1024->512->…->3. Similarly, with each block in our assignment, the “image” size increases, while the channels decrease until we get to the size of a 28x28 image with one channel.

Thank you Nydia and Wendy