READ ME File For 'Dataset for "Incremental Training and Group Convolution Pruning for Runtime DNN Performance Scaling on Heterogeneous Embedded Platforms"'

Dataset DOI: 10.5258/SOTON/D1245

ReadMe Author: Lei Xun, University of Southampton [0000-0003-0765-6726]

This dataset supports the publication:
AUTHORS: Lei Xun, Long Tran-Thanh, Bashir M Al-Hashimi, Geoff V. Merrett
TITLE: Incremental Training and Group Convolution Pruning for Runtime DNN Performance Scaling on Heterogeneous Embedded Platforms
JOURNAL: ACM/IEEE Workshop on Machine Learning for CAD 2019 (MLCAD'19)
PAPER DOI IF KNOWN:


This dataset contains:
Data for Fig 2,3,4,5(a-b)

The figures are as follows:

Fig 2. Top-1 image classification accuracy on 10,000 CIFAR10 validation images. Our Dynamic DNN has four different model configurations which have four different accuracies, at runtime, dynamic DNN can switch to smaller configuration for time/energy reduction with accuracy loss, or switch back to larger models for the accuracy recovery once more computing resources become available.

Fig 3. Total confidence score of the correct DNN output over 10,000 CIFAR10 validation images. The value is normalized to the 100\% model. The confidence is improved when more groups are added to dynamic DNN, this indicate different feature filters are learnt in later groups.

Fig 4. Inference time on four heterogeneous cores of two hardware platforms. Our group convolution pruning is fully compatible with both CPU and GPU.

Fig 5. Dynamic DNN (different symbols) is combined with task mapping (different colours) and DVFS (different points). Different configurations are used for different runtime energy (a), power (b) and time budget. A standalone dynamic DNN  with four increments can only provide four trade-off points over a limited dynamic range. With the combination of task mapping and DVFS, the trade-offs are finer and the dynamic ranges are wider.

Date of data collection: 2019

Information about geographic location of data collection: Southampton, UK

Licence: CC BY

Related projects: International Centre for Spatial Computation

Date that the file was created: FEB 2020