Much Lower performance on ResNet101, scripts for reproducing needed on ResNet101

I followed your instructions to train a ResNet101-DeeplabV3+-decouple just like your provided script to train the ResNet50-DeeplabV3+-decouple. Despite I got a decent results for the first stage to train the base modle, I only got 81.8 after I trained the ResNet101-DeeplabV3+-decouple for the second stage，which is 1.7 miou lower than the claimed performance in your ECCV passage. I've tried many different settings like changing the learning rate of the second stage, changing the weight of the joint_edgeseg_loss, but 81.8 (ms+flip inference) is the best results I can get using the provided code.

So could you please, or anyone who have reproduced the claimed results on ResNet101, provide the scripts for reproducing your claimed ResNet101 results in the passage?
Although you have provided the trained modle, I am pretty sure that there gotta be someone like me who are desperate to reproducing this amazing work on ResNet101.

Start with the provided ResNet50-DeeplabV3+-decouple training script and the supplied ResNet101 model, then compare the second-stage settings, including learning rate and joint_edgeseg_loss weight. Re-run the ResNet101 second stage and verify whether the claimed performance can be reproduced instead of the reported 81.8 mIoU.