Now that I can predict, I can adapt: Towards CNN adaptation of Edge GPUs through modelling of the training process

Add to your list(s) Download to your calendar using vCal

• Aditya Rajagopal
• Monday 19 October 2020, 16:00-17:00
• Teams.

If you have a question about this talk, please contact George A Constantinides.

With an increase in the memory and processing capabilities of today’s edge devices, there is a push towardsincreased edge intelligence. For Convolutional Neural Networks (CNNs), pruning and retraining networks on theedge, adapting their structure to the data being processed and resulting in systems with lower memory, latencyand power consumption, is a desirable capability towards achieving edge intelligence. However, the limitedcompute resources and memory budget of edge devices restrict the model size that can be trained under a certaintime budget. This work describes an automated methodology for developing accurate models that predict CNNtraining memory consumption and latency given a target device, network and pruning level. With PyTorch asthe framework and NVIDIA Jetson TX2 as the target device, the developed models predict training memoryconsumption and latency with 91.0% and 88.8% accuracy respectively for a wide range of networks and pruninglevels. Additionally the work develops a classification model that predicts with 93.4% accuracy whether a givennetwork and pruning level will perform training on the target device given the available memory.

This talk is part of the CAS Talks series.

This talk is included in these lists:

• Andrea Picciau's list
• CAS Talks
• Circuits and Systems Group: Internal Seminars
• Teams

Note that ex-directory lists are not shown.