The DeepLab implementation seems to have a long-lasting unobserved issue in Classifier_Module:
in line 116 the out variable is returned in the first iteration of the loop. So it only sums up the output of the first and second atrous convolutions in ASPP module (not all 4 of them).
The source of this problem probably is a bug in the original implementation which was fixed later.
Unfortunately many other researchers have been used this implementation (or the original implementation before its bug being fixed). So whoever reads this issue better be cautious about the DeepLabv2 pytorch implementation used in other repositories too.
The DeepLab implementation seems to have a long-lasting unobserved issue in Classifier_Module:
in line 116 the out variable is returned in the first iteration of the loop. So it only sums up the output of the first and second atrous convolutions in ASPP module (not all 4 of them).
The source of this problem probably is a bug in the original implementation which was fixed later.
Unfortunately many other researchers have been used this implementation (or the original implementation before its bug being fixed). So whoever reads this issue better be cautious about the DeepLabv2 pytorch implementation used in other repositories too.