This paper describes two systems for the second subtask of CoNLL-SIGMORPHON 2018 shared task on universal morphological reinflection submitted by the University of Colorado Boulder team. Both systems are implementations of RNN encoder-decoder models with soft attention. The first system is similar to the baseline system with minor differences in architecture and parameters, and is implemented using PyTorch. It works for both track 1 and track 2 of the subtask and generally outperforms the baseline at low data settings in both tracks. The second system predicts the morphosyntactic description (MSD) of the lemma to be inflected using an MSD prediction model. The data for subtask 2 is processed and reformatted to subtask 1 data format to train an inflection model. Then the inflection model predicts the inflected form for the target lemma given the predicted MSD. This system achieves higher accuracies than the first system when the training data is the most limited, though it does not perform better when the training data is abundant.