Transfer Learning and meta-learning have been effective in improving performance across multiple domains. It has also been applied successfully to EEG decoding where there is a lack of data. However, there are unique challenges for transfer learning with EEG data across datasets due to differences in experimental setup, like different numbers of electrodes, different positions of the electrodes, and different task definitions. To tackle the issue of cross-dataset training across heterogeneous electrode configuration EEG datasets we introduce a novel method, CoordinateAttention, that uses 3-D coordinates of the electrode sensors to learn the spatial relationship between the electrode's positions to dynamically generate spatial convolution kernels for feature extraction. We show that our model has good performance in EEG decoding across settings and is robust to data corruption. CoordinateAttention is a general-purpose method for feature extraction and data fusion using geometric positional information.