Training problems

[Zhou-pf]

Hello:
I used the default configuration file (--hparams ./configs/coco_search18_dense_SSL_TP.json) for TP training, but I noticed that some of the metrics in the final results didn’t reach even half of the values reported in the paper (my experiment results: cIG=1.067, cNSS=2.6845; paper results: cIG=2.399, cNSS=5.086). Additionally, when I checked TensorBoard, I found that the metrics showed a downward trend towards the later stages of training. So, I changed the max_iters in the default configuration file to 6000, but I found that these metrics dropped significantly. Could you please explain why this happened? Is 3000 iterations the optimal training number for HAT?

[Zhou-pf]

[RuoyuX-2018]

The image link is unavailable, could you please reupload it? We checked our history TensorBoard and the performance keeps increasing at 6000 iterations.

[Zhou-pf]

Sorry, the image upload failed. I will upload it again now