view article
Figure 5
Secondary training of the previous algorithms (G50 or H50) on ten images of target G (H50 > G10, top panels A and B, and E and F) or target H (G50 > H10, bottom panels C and D, and G and H), using 10 epochs and learning rates of 1 × 10−3 (panels A and C, and E and G) or 1 × 10−4 (panels B and D, and F and H), and evaluation on target G (left, A–D) or H (right, E–H). Whereas G50 > H10 seems to be performed correctly on both targets (with low influence by the learning rate used), the H50 > G10 algorithm seems to require higher learning rates for adapting to the secondary target (losing in turn adaptation to its primary target). This shows that the learning rate is a parameter that must be tuned depending on each target's characteristics. Performance indicator values are shown in Table 1[link].

Journal logoJOURNAL OF
APPLIED
CRYSTALLOGRAPHY
ISSN: 1600-5767
Volume 57| Part 2| April 2024| Pages 266-275
https://doi.org/10.1107/S1600576724000682
Follow J. Appl. Cryst.
Sign up for e-alerts
E-alerts
Follow J. Appl. Cryst. on Twitter
Twitter
Follow us on facebook
Facebook
Sign up for RSS feeds
RSS