We can evaluate the models in three different aspects:
- Learning/(re)training time
- Storage requirement
- Performance (accuracy)
In terms of training time, in a desktop (Intel Xenon CPU E5-1650 v3@3.5 GHz and 32 GB RAM) with GPU support, LSTM on the HAR dataset, CNN on FER2013, and Mobilenet V1 on the FER2019 dataset, it took less than an hour to train/retrain the model.
The storage requirement of a model is an essential consideration in resource-constrained IoT devices. The following diagram presents the storage requirements for the three models we tested for the two use cases. As shown in the diagram, the simple CNN takes up only 2.6 MB, smaller than one sixth of the Mobilenet V1 (17.1 MB). Also, the LSTM for the HAR took up 1.6 MB (not in the diagram) of storage. In terms of storage requirements, all the models are fine to be deployed in many resource-constrained...
