Table 2
Performances of the individual and ensemble network models evaluated on the test dataset are reported here.
Architecture | Parameters | CPU hours | TPR | FPR | AUROC | Candidates | Lens | Model reference |
---|---|---|---|---|---|---|---|---|
BaseNet | 5 81 828 | 104 | 0.906 | 0.097 | 0.973 | 62326 | 13 | Andika et al. (2023) |
RegNetX002 | 2 338 692 | 1202 | 0.924 | 0.072 | 0.983 | 160852 | 17 | Radosavovic et al. (2020) |
RegNetY002 | 2 816896 | 1317 | 0.899 | 0.078 | 0.977 | 192797 | 13 | Radosavovic et al. (2020) |
MobileNetV3Large | 3 000484 | 148 | 0.938 | 0.046 | 0.989 | 190808 | 16 | Howard et al. (2019) |
EfficientNetB0 | 4 055 275 | 314 | 0.945 | 0.042 | 0.991 | 147705 | 19 | Tan & Le (2019) |
NASNetMobile | 4 274 520 | 434 | 0.948 | 0.048 | 0.991 | 169309 | 17 | Zoph et al. (2018) |
EfficientNetV2B0 | 5 925 012 | 243 | 0.958 | 0.030 | 0.994 | 151 319 | 15 | Tan & Le (2021) |
ViT-Vanilla | 9 208 772 | 1185 | 0.949 | 0.050 | 0.990 | 24642 | 16 | Dosovitskiy et al. (2020) |
ViT-Lite | 9 230060 | 3499 | 0.960 | 0.029 | 0.994 | 12 872 | 16 | Lee et al. (2021) |
Xception | 20 870 252 | 1041 | 0.970 | 0.017 | 0.997 | 216620 | 17 | Szegedy et al. (2016) |
InceptionV3 | 21 811 556 | 549 | 0.967 | 0.020 | 0.996 | 186190 | 17 | Szegedy et al. (2015) |
ResNet50V2 | 23 579 268 | 1094 | 0.967 | 0.021 | 0.996 | 164722 | 18 | He et al. (2016) |
ResNetRS50 | 33 705 060 | 1245 | 0.966 | 0.021 | 0.996 | 155 262 | 16 | Bello et al. (2021) |
InceptionResNetV2 | 54 343 460 | 1630 | 0.966 | 0.014 | 0.996 | 150 435 | 18 | Chollet (2016) |
Ensemble | 195 741 135 | … | 0.963 | 0.016 | 0.996 | 3080 | 16 | This work |
Notes. We train the models using computing nodes that contain Intel Xeon E5-2680 2.4 GHz, 28 cores, and 252 GiB RAM each. Columns, from left to right, correspond to the architecture name, the number of parameters in the model, the total CPU hours required to train the network, the resulting TPR, the FPR, the AUROC, the number of selected lens candidates, the number of recovered known lensed quasars, and the literature explaining the architecture design.
Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.