Are you sure about the value of "batch_norm_decay" is 0.1 ?

[dreambear1234]

According the official example of tensorflow, https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/slim/python/slim/nets/resnet_utils.py in line 226, the decay value is 0.997; I'm using a revised ENet in tensorflow, but I can't get the iou value as well as Torch, about 4 less than Torch, Appreciate for your reply!

[kwotsin]

The value is different as it was based upon the original torch implementation of the ENet authors (see references in README).

[dreambear1234]

@kwotsin ,Are you explain me that the hyper parameter "momentum"=0.9? maybe it is a optimizer for training process(accelerate convergence) in back propagation; But the decay of BN is used to get a statistics value of mean and var in forward propagation(moving mean, moving var). they are two separate value(momentum, bn_decay). Is it correct? Maybe I'm confused. Thank you for your reply

[kwotsin]

@dreambear1234 Unfortunately I've not yet have the time to test this out - have you experimented with the value you chosen and see if it results in a better performance? In your revised ENet, what did you change?