Research Article
Local Epochs Inefficiency Caused by Device Heterogeneity in Federated Learning
Table 10
The mean ranking of convolutional feature weights.
| Features | P100_Conv | V100_Conv | K40_Conv | MeanRank |
| Batchsize | 8.692 | 9.574 | 5.702 | 7.989333333 | Elements_matrix | 5.8396 | 8.6824 | 7.5608 | 7.360933333 | Elements_kernel | 11.2952 | 11.7014 | 10.1 | 11.0322 | Channels_in | 1.09 | 1.0334 | 1.0112 | 1.044866667 | Channels_out | 7.0608 | 5.6072 | 12.0992 | 8.255733333 | Padding | 14.1344 | 13.7096 | 13.9144 | 13.91946667 | Strides | 8.9108 | 9.8826 | 8.4852 | 9.092866667 | Use_bias | 14.5956 | 14.595 | 15.7176 | 14.9694 | Opt_SGD | 6.3264 | 6.853 | 4.2328 | 5.804066667 | Opt_Adadelta | 3.764 | 7.3048 | 7.0828 | 6.050533333 | Opt_Adagrad | 7.732 | 5.3538 | 6.3752 | 6.487 | Opt_Momentum | 5.2476 | 5.3584 | 11.0724 | 7.226133333 | Opt_Adam | 10.1944 | 3.6814 | 5.1448 | 6.3402 | Opt_RMSProp | 2.9736 | 4.107 | 5.9512 | 4.343933333 | Act_relu | 15.8232 | 15.6576 | 7.3332 | 12.938 | Act_tanh | 14.0576 | 15.8164 | 15.514 | 15.12933333 | Act_sigmoid | 15.2628 | 14.082 | 15.7032 | 15.016 |
|
|