Design and Development of an Optimized Fast Transformation Module (OpFTM) for GAN Accelerator with Computation Efficiency

Generative Adversarial Network (GAN) is a type of deep learning-based neural network which is gaining popularity, in image processing domain, computer gaming, etc. A general GAN architecture consists of two important neural network elements namely, generator and discriminator. The generator part of the GAN predominantly uses deconvolutional operations to produce the output tiles. But the deconvolution layers presented in various research did not address the performance imbalance. This proposed work focuses on the optimization of the deconvolution layer to address the performance issues through the best Fast Transformation Module (FTM). The proposed FTM functions are based on the Fast Transformation Algorithm (FTA) which reduces the number of multiplications needed to compute the deconvolution operation. To optimize the performance of the FTM, Baugh Wooley multiplier and ripple carry look-ahead adder (RCLA) circuits have been used.