Convolutional neural network errors

Citation Author(s):
Fernando Fernandes dos Santos
Submitted by:
Fernando dos Santos
Last updated:
Tue, 05/17/2022 - 22:17
DOI:
10.21227/H2WT0P
Data Format:
Research Article Link:
License:
0
0 ratings - Please login to submit your rating.

Abstract 

This file contains all data used on paper "Analyzing and Increasing the Reliability of Convolutional Neural Networks on GPUs"

Instructions: 

Data is organized in the following files:

-- sassifi_inst: contains all errors obtained on fault injections using INST mode

-- sassifi_rf: contains all errors obtained on fault injections using RF mode

-- ecc_on: contains all observed errors under the beam for K40 with ECC enabled

-- ecc_off: contains all observed errors under the beam for K40 (ECC OFF), Titan X, and X

Each file is organized as follows:

BENCHMARK_AND_MACHINE_NAME: which could be cudaDarknet_carol-k402 for YOLO(Darknet) on K40, PyFasterRcnn_carol-k402 Faster RCNN on K40, cudaDarknet_carolx1 Darknet on Tegra X1, PyFasterRcnn_carol-tx Faster RCNN on Titan X or cudaDarknet_carol-tx Darknet on Titan X
log_name:\

Each log output file which contains:
-date and time of the test
sdc_iteration
-iteration of SDC, once not all executions produced SDC
-
it_errors: how many errors in this SDC
-
ERROR_LIST
All errors listed, to compare with golden value the errors are printed always using the value read (x_r, y_r, prob_r for object probability, h_r for height, w_r for width) and expected value (x_e, y_e, prob_e for object probability, h_e for height, w_e for width)