A new approach to auto-calibrating, coil-by-coil parallel imaging reconstruction is presented. It is a generalized reconstruction framework based on deep learning. A neural network consisting of three Dense layer (Fully connected layer) units, an RNN layer and an output Dense unit is designed and trained to identify the mapping relationship between the zero-filled and fully-sampled k-space data. The training process could be separated into two steps: pre-training and fine-tuning. Results show our proposed model could be robust to arbitrary undersampling patterns in k-space and shows a higher structural similarity index compared with traiditional k-space based methods.