The last two parts of this tutorial explained a lot about the selection and acquisition of suitable data and its preparation for training deep learning models for object recognition. Using Python and the freely available Google Colab, developers have seen how the data can be used within a Jupyter notebook environment and how the training time can be reduced through hardware acceleration.
Notebooks and Colab represent the tools that are currently being used to develop and train machine learning models. The cloud is ideally suited as a training ground, which is not only due to the fact that all major cloud providers provide notebook environments, but that these can be scaled on demand. If it is foreseeable that a model training session on the booked hardware will take several days, developers in the cloud can reduce the training time with larger hardware.
The question remains how to get trained models for the actual target systems. This step is necessary in many application scenarios, since powerful graphics cards from server hardware are not available everywhere. Today, most of the deep learning models can be found in mobile devices, industrial robots or cars. These applications run on limited or at least specialized hardware. Power consumption is particularly important in mobile systems. In contrast to GPU modules in servers, which quickly need several 100 watts for operation, mobile devices have to optimize the coefficient of performance and energy consumption, i.e. the number of GPU cores for power consumption in watts.
- Access to all heise + content
- exclusive tests, advice & background: independent, critically well-founded
- Read c’t, iX, Technology Review, Mac & i, Make, c’t photography directly in your browser
- register once – read on all devices – can be canceled monthly
- first month free, then € 9.95 per month
- Weekly newsletter with personal reading recommendations from the editor-in-chief
Start FREE month
Start your FREE month now
Already subscribed to heise +?
Sign up and read
Register now and read articles immediately
More information about heise +