Compressive sensing (CS) is widely used to reduce the image acquisition time of magnetic resonance imaging (MRI). Though CS based undersampling has numerous benefits, like high quality images with less motion artefacts, low storage requirement, etc., the reconstruction of the image from the CS-undersampled data is an ill-posed inverse problem which requires extensive computation and resources. In this paper, we propose a novel deep network that can process complex-valued input to perform high-quality reconstruction. Our model is based on generative adversarial network (GAN) that uses residual-in-residual dense blocks in a modified U-net generator with patch based discriminator. We introduce a wavelet based loss in the complex GAN model for better reconstruction quality. Extensive analyses on different datasets demonstrate that the proposed model significantly outperforms the existing CS reconstruction techniques in terms of peak signal-to-noise ratio and structural similarity index.