Electroencephalogram (EEG) based brain-computer interface (BCI) has emerged as a promising tool for communication and control. Temporal non-stationarity of the signal is one of the critical challenges faced by motor imagery (MI) classification for EEG based BCI. To address this challenge, this paper proposes a novel temporal attention convolutional network (TACNet) for MI classification. By combining two types of sub-networks through attention mechanisms, TACNet can selectively focus on valuable time slices of the signal to obtain task-related information. In TACNet architecture, a global sub-network is applied to the entire time horizon and guides the attention mechanism to select a few time slices to apply the local sub-networks. We compare TACNet with other deep learning models on two EEG datasets: BCI competition IV dataset 2a (BCIC IV 2a) and high gamma dataset (HGD). The results show that our approach achieves significantly better classification accuracies than other baseline models.