It’ll be very helpful if the TFMOT team can share the pruning & fine-tuning configuration for MobileNets to reproduce the above accuracies. The tutorials and examples are based on toy problems, and don’t offer enough guidance on how to use TFMOT to optimize production-level models. Thanks.
Thanks for the comment. The result shown in the screenshot is from the published paper in 2017 - https://arxiv.org/pdf/1712.05877.pdf (This is also linked in the same page), so you can find the training parameters in Appendix of the paper.
We also provide some proudction-level model examples in collaboration with Model Garden. You may be able to find the information here - Adding Quantization-aware Training and Pruning to the TensorFlow Model Garden — The TensorFlow Blog
The model garden example contains the real configuration & script to train MobilenetV2 and ResNet50 with pruning, so please check it out!