Hyperspectral imaging can provide more spectral information than an ordinary RGB camera. The spectral information has been beneficial to numerous applications, such as monitoring natural environment changes and classifying plants and soils in agriculture. The hyperspectral images reconstruction from a single RGB image is severely unconstrained problem. Previous methods need additional components or the spectral response by commercial camera. An end-to-end conditional generative adversarial network was proposed with modified residual network as backbone. The feature pyramid was used inside the network and a scale attention module was designed to fuse local and global information. In order to provide more accurate solution, another distinct architecture was proposed, named WNet. Experiments manifested the superiority of the proposed method over other representative methods in terms of quality and quantity. Experiments used both synthesized RGB images using public hyperspectral data and real-world image by ordinary camera demonstrate that proposed method outperforms the state-of-the-art. The WNet drops 45% and 50% in terms of RMSE and relative RMSE on the ICVL dataset than sparse coding.