This research presents a deep learning based approach to predict stress fields in the solid material elastic deformation using convolutional neural networks (CNN). Two different architectures are proposed to solve the problem. One is Feature Representation embedded Convolutional Neural Network (FR-CNN) with a single input channel, and the other is Squeeze-and-Excitation Residual network modules embedded Fully Convolutional Neural network (SE-Res-FCN) with multiple input channels. Both the tow architectures are stable and converged reliably in training and testing on GPUs. Accuracy analysis shows that SE-Res-FCN has a significantly smaller mean squared error (MSE) and mean absolute error (MAE) than FR-CNN. Mean relative error (MRE) of the SE-Res-FCN model is about 0.25% with respect to the average ground truth. The validation results indicate that the SE-Res-FCN model can accurately predict the stress field. For stress field prediction, the hierarchical architecture becomes deeper within certain limits, and then its prediction becomes more accurate. Fully trained deep learning models have higher computational efficiency over conventional FEM models, so they have great foreground and potential in structural design and topology optimization.