I am running a GAN, DCGAN (https://github.com/carpedm20/DCGAN-tensorflow) using TensorFlow. I notice that about 73% of the execution time is spent in a module called dgrad_engine. Can anyone tell me what this dgrad_engine is ?
moved to tensorflow board for more coverage.
A dgrad operation computes the gradient of a convolution layer with respect to the input “data”. The gradient with respect to the kernels weights (wgrad) is computed separately. Depending on various parameters (e.g., dilation, batch size, etc.) gradient kernels sometimes require challenging data access patterns that make them run more slowly than, for example, the forward convolution.