Evaluating code generation with pretrained models of code.