Online traffic speed data of urban road networks serve as the foundation of modern intelligent transportation systems. Much research has been conducted on developing methods, mostly model-based or machine learning ones, to estimate the data with GPS record for one, few adjacent roads, or the entire vehicular transportation network. While the machine learning methods generally yield satisfactory estimation accuracy, their accomplishments are established on a plethora of historical GPS records which may not be readily available for many urban transportation systems. In this paper, we investigate a transfer learning approach to provide speed data estimations with few data. We ground this work on a graph convolutional generative autoencoder that can generate the estimations for an entire transportation network in one go, and modify its internal computation graph to reduce the size of network topology-dependent model parameters. Subsequently, pre-trained models from road networks with massive historical data can be re-used in other networks with few data, which are only employed to adjust a small number of parameters. To assess the effectiveness of the proposed approach, comprehensive case studies are conducted, in which outstanding speed estimations can be obtained with significantly shorter training time.