Table 3
Numerical performance tests as a function of batch size.
| Batch size | Computer time per epoch [sec] | Number of batches | GPU usage [%] | CPU usage [%] |
|---|---|---|---|---|
| CPU only | ||||
| 32 | 120 | 28 125 | 0 | 90 |
| 64 | 82 | 14063 | 0 | 90 |
| 128 | 81 | 7032 | 0 | 90 |
| 256 | 66 | 3516 | 0 | 92 |
| 512 | 60 | 1758 | 0 | 93 |
| CPU + GPU | ||||
| 32 | 92 | 28 125 | 2 | 14 |
| 64 | 55 | 14063 | 5 | 17 |
| 128 | 28 | 7032 | 5 | 17 |
| 256 | 13 | 3516 | 6 | 10 |
| 512 | 8 | 1758 | 5 | 10 |
| 1024 | 6 | 879 | 5 | 10 |
| 2048 | 5 | 440 | 5 | 10 |
Notes. For this experiment in particular, we generated one million synthetic datasets following the time sampling of ɛ Eri(66 observations). For each experiment we trained a convolutional NN (three convolutional layers, followed by two fully connected ones). The input samples consisted of three spectroscopic indices (FWHM, BIS, and CON).
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.