ADE20k Learning Rate

[cjrd]

Section A.4 of the paper has "We search for the optimal lr for each entry in Table 5 " when referring to segmentation on ADE20k. Could you share these learning rates? My reproduced baselines are about 2 points too low.

Thanks!

[implus]

see our reproduction here: #2 (comment)

[endernewton]

Closing with the reproduced results above. Please check lr, layer-wise rate decay, and drop path rate similar to what we have done in detection.

[youngwanLEE]

@endernewton Hi,

Are LR, layer-wise rate decay, and drop-path hyper-parameters of UperNet the same as those of detection?

Do you have a plan to release the segmentation code and weights?

[hellojialee]

Dose layer-wise rate decay really make differences obviously?