[Segnet] TypeError: new() received an invalid combination of arguments

[tailongnguyen]

Hi,
I am using pytorch version 0.4.0, python 3.6 and this code appears to need a minor modification to make Segnet works.
Here is the bug:

I found out that the line 12 should change from
middle_channels = in_channels / 2
to
middle_channels = in_channels // 2

[oujieww]

@tailongnguyen is your training ok? i also want to train segnet on voc but with training go on ,my mean iu is always =0.03469,can you give me your setting about trainning code,i use the setting same as fcn,lr=2.5e-4

[tailongnguyen]

@oujieww
I used this code but did not focused on this SegNet architecture, thus only tried on training with one set of hyper parameters (lr = 1e-4, negative log likelihood loss function with weighted scheme following E-Net) . The performance was unexpectedly bad, too. I am not sure if there was something wrong with the code or my dataset is just not suitable with this model, so I skipped it.
However, if you ask about the training, my mean IoU DID change over the training phase.

[oujieww]

@tailongnguyen yeah, is your training mIoU is nromal? such as bigger than 0.5? and i have tried this code on fcn32s even is better than official code mIoU is 0.6447 i got. but for other model of this code i can not work, ,XD , do u know where can find a good code of semantic segmentation , i want to do some test on segmentation XD

[tailongnguyen]

Awesome-semantic-segmentation github repo is what you need.

…

On Thu, Jul 26, 2018, 09:49 oujieww ***@***.***> wrote: @tailongnguyen <https://github.com/tailongnguyen> yeah, is your training mIoU is nromal? such as bigger than 0.5? and i have tried this code on fcn32s even is better than official code mIoU is 0.6447 i got. but for other model of this code i can not work, ,XD , do u know where can find a good code of semantic segmentation , i want to do some test on segmentation XD — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#42 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AWZ-P_rLMYZcYfB-8iNVvYWuCYCaCoDMks5uKS4igaJpZM4VSia4> .

[oujieww]

@tailongnguyen thank you !!!

[robinreni96]

Its simple in the Decoderblock inside init function instead of
middle_channels = in_channels / 2
use
middle_channels = int(in_channels / 2)