Deep learning based device fingerprinting has emerged as a key method of identifying and authenticating devices solely via their captured RF transmissions. Conventional approaches are not portable to different domains in that if a model is trained on data from one domain, it will not perform well on data from a different but related domain. Examples of such domains include the receiver hardware used for collecting the data, the day/time on which data was captured, and the protocol configuration of devices. This work proposes Tweak, a technique that, using metric learning and a calibration process, enables a model trained with data from one domain to perform well on data from another domain. This process is accomplished with only a small amount of training data from the target domain and without changing the weights of the model, which makes the technique computationally lightweight and thus suitable for resource-limited IoT networks. This work evaluates the effectiveness of Tweak vis-a-vis its ability to identify IoT devices using a testbed of real LoRa-enabled devices under various scenarios. The results of this evaluation show that Tweak is viable and especially useful for networks with limited computational resources and applications with time-sensitive missions.