We can evaluate the models from three different aspects:
- Learning/(re)training time
- Storage requirement
- Performance (accuracy)
The Mobilnet V1's retraining and validation process using the retrain.py module took less than an hour on a desktop (Intel Xenon CPU E5-1650 [email protected] and 32 GB RAM) with GPU support.
The storage/memory requirement of a model is an essential consideration for resource-constrained IoT devices. To evaluate the storage/memory footprint of the Mobilenet V1, we compared its storage requirement to another two similar networks' (the Incentive V3 and CIFAR-10 CNN) storage requirements. The following screenshot presents the storage requirements for the three models. As shown, Mobilenet V1 requires only 17.1 MB, less than one-fifth of the Incentive V3 (87.5 MB) and CIFAR-10 CNN (91.1 MB). In terms of storage requirements, Mobilenet...