Hey Daniel, thanks for covering that topic. I also implemented that approach some time ago. However, From my point of view it will never be a complete solution to accelerate Deep Learning as a whole on your local machine.
The first problem is, that to my knwoledge, Keras will no longer support multi backend support. Thus, from the next generation ongoing, plaidml will not longer work. Hence, new models, layers etc. that might come to Keras will not be available.
Secondly, it is advised to use Tf.keras anyway from now on, because they have merged into Tensorflow, thus support from the community will be low in the future. This somehow covers the same problem like my first point.
Last but not least, community based plattforms wirh pre trained models etc. Rely on Tensorflow and not Keras. So, using the Object Detection Api for example will receive no benefit at all with that setup
That is why I would recommend to everyone who has a deeper interest in Deep Learning should use something like Google Colab, if he wants to stick to a Mac or should leave MacOs und join Linux (or Windows) with a Nvidia GPU.