READ ME File For 'Dataset supporting the thesis "Realising the Benefits of Dynamic DNNs on Reconfigurable Hardware"'

Dataset DOI: 10.5258/SOTON/D3483

ReadMe Author: Anastasios Dimitriou, University of Southampton, 0009-0005-0925-8459

This dataset supports the thesis entitled "Realising the Benefits of Dynamic DNNs on Reconfigurable Hardware"
AWARDED BY: University of Southampton
DATE OF AWARD: 2025

DESCRIPTION OF THE DATA
The data contain all the necessary information that are portrayed in the figures and tables of the thesis "Realising the Benefits of Dynamic DNNs on Reconfigurable Hardware". No specialist software needed to view the data.

This dataset contains:
- Chapter 3 folder: Data supporting figures and tables of the third chapter of the thesis.
	- 'Table3_1.csv': Data supporting TABLE 3.1: FPGA Resource Utilisation.  
	- 'Fig3_1.csv': Data supporting FIGURE 3.10: Experimental results comparing average Execution Time per sample. Average values are calculated based on each exit point trigger rate. 
	- 'Table3_2.csv': Data supporting TABLE 3.2: Execution Time (ms) of a static (Normal) and an early-exit dynamic LeNet-5 DNN
	- 'Table3_3.csv': Data supporting TABLE 3.3: Energy Consumption (mJ), with (w/ EE) and without (w/o EE) early exits of a dynamic LeNet-5 DNN.
	- 'Fig3_2.csv': Data supporting FIGURE 3.11: Experimental results comparing average Energy Consumption per sample. Average values are calculated based on each exit point trigger rate.
  
- Chapter 4 folder: Data supporting figures and tables of the fourth chapter of the thesis. 
	- 'Table4_1.csv': Data supporting TABLE 4.1: Resource Utilisation.  
	- 'Fig4_1.csv': Data supporting FIGURE 4.2: Experimental results comparing average Execution Time per sample. Average values are calculated based on each exit point trigger rate (ai). 
	- 'Fig4_2.csv': Data supporting FIGURE 4.3: Experimental results comparing average Energy Consumption per sample. Average values are calculated based on each exit point trigger rate (ai).
	- 'Table4_2.csv': Data supporting TABLE 4.2: Execution Time (ms).
	- 'Table4_3.csv': Data supporting TABLE 4.3: Energy Consumption (mJ), with (w/ EE) and without (w/oEE) early exits.
	- 'Fig4_3.csv': Data supporting FIGURE 4.4: Comparison of the pipeline and parallel designs over average execution time across early-exit LeNet-5, AlexNet, VGG19 and ResNet32.  
	- 'Fig4_4.csv': Data supporting FIGURE 4.5: Comparison of the pipeline and parallel designs over average energy consumption across early-exit LeNet-5, AlexNet, VGG19 and ResNet32.  
	- 'Table4_4.csv': Data supporting TABLE 4.4: Performance Comparison With Existing Implementations
	- 'Fig4_5.csv': Data supporting FIGURE 4.6: Comparison of the pipeline and parallel designs over average data movement across early-exit LeNet-5, AlexNet, VGG19 and ResNet32.  
	- 'Fig4_6.csv': Data supporting FIGURE 4.7: On a 3 point early-exit Resnet-32 a) shows each exit’s trigger rate and b) the percentage difference of parallel over pipeline approaches over Energy and Time for different Decision Thresholds. 
	- 'Fig4_7.csv': Data supporting FIGURE 4.8: On an 3 point early-exit Resnet-32 a) shows each exit’s trigger rate and b) the percentage difference of parallel over pipeline approaches over Energy and Time for different Decision Thresholds.  

- Chapter 5 folder: Data supporting figures and tables of the fifth chapter of the thesis. 
	- 'Fig5_1.csv': Data supporting FIGURE 5.1: The exit trigger rates (ai) of an early-exit ResNet-32 Dynamic DNN with an intermediate classifier after every layer.
	- 'Fig5_2.csv': Data supporting FIGURE 5.2: Average δj for the input samples that failed to trigger exit points 2 and 14 but achieved it at points 3 to 6 and 15 to 19 accordingly. 
	- 'Table5_1.csv': Data supporting TABLE 5.1: Exit Trigger Rates (%)  
	- 'Table5_2.csv': Data supporting TABLE 5.1: Execution Time (ms)  
	- 'Fig5_3.csv': Data supporting FIGURE 5.3: Early-Exit Branches Memory Needs over different quantisation levels.
	- 'Fig5_4.csv': Data supporting FIGURE 5.4: Percentage of samples to trigger each exit point over different quantisation levels of the exit branch.
	- 'Table5_3.csv': Data supporting TABLE 5.3: Per Sample Execution Time (ms) of Early-Exit Dynamic ResNet-32 over Different Quantisation Levels of Intermediate Classifiers
	- 'Table5_4.csv': Data supporting TABLE 5.4: Per Sample Energy Consumption (mJ) of Early-Exit Dynamic ResNet-32 over Different Quantisation Levels of Intermediate Classifiers
	- 'Fig5_5.csv': Data supporting FIGURE 5.5: Early-exit ResNet-32 memory needs over different quantisation levels.
	- 'Fig5_6.csv': Data supporting FIGURE 5.6: Percentage of samples to trigger each exit point over different quantisation levels of the early-exit network.
	- 'Table5_5.csv': Data supporting TABLE 5.5: Per Sample Execution Time (ms) of Early-Exit Dynamic ResNet-32 over Different Quantisation Levels
	- 'Table5_6.csv': Data supporting TABLE 5.6: Per Sample Energy Consumption (mJ) of Early-Exit Dynamic ResNet-32 over Different Quantisation Levels
	- 'Fig5_7.csv': Data supporting FIGURE 5.8: The number of inputs for which δ ≤ 0.1, δ ≤ 0.2 and δ ≤ 0.3 at Exit 1 and Exit 2.
	- 'Fig5_8.csv': Data supporting FIGURE 5.9: Percentage of samples to trigger each exit point over different dynamic quantisation levels when δ ≤ 0.1.
	- 'Fig5_9.csv': Data supporting FIGURE 5.10: Percentage of samples to trigger each exit point over different dynamic quantisation levels when δ ≤ 0.2.
	- 'Fig5_10.csv': Data supporting FIGURE 5.11: Percentage of samples to trigger each exit point over different dynamic quantisation levels when δ ≤ 0.3.
	- 'Table5_7.csv': Data supporting TABLE 5.7: Per Sample Execution Time (ms) of Early-Exit Dynamic ResNet-32 with and without applying dynamic quantisation.


Date of data collection: 09/2021 to 02/2024

Information about geographic location of data collection: University of Southampton, SO17 1BJ, Southampton, United Kingdom 

Licence:
CC BY 4.0

Related projects/Funders: Engineering and Physical Sciences Research Council (EPSRC) under EP/S030069/1

Date that the file was created: April, 2025