Closed wujiaju closed 5 years ago
@wujiaju I also come across this issue, have you ever find some tricks to overcome this problem?
@zhouyuan888888 No. I did't check the code.
MobileNet-v2 requires different hyperparams + you need to load-in ImageNet-pretrained weights for it
@DrSleep Can you give me a link to ImageNet-pretrained weights~~0(_)0, thank you
@DrSleep Hi, thank you so much0(#_#)0
@DrSleep Hello,T_T, I have download the pretrained weights as you provided, but there are some misstakes: Missing key(s) in state_dict: "layer1.0.weight", "layer1.1.running_var", "layer1.1.bias", "layer1.1.weight", "layer1.1.running_mean", "layer2.0.output.0.0.weight", "layer2.0.output.0.1.running_var", "layer2.0.output.0.1.bias"........
can you give me some tips to fix this bugT_T.
There might have been some changes in that repo since I used it. You may need to convert the weights manually instead
On Sat, 29 Jun 2019 6:56 pm sky, notifications@github.com wrote:
@DrSleep https://github.com/DrSleep Hello,T_T, I have download the pretrained weights as you provided, but there are some misstakes: Missing key(s) in state_dict: "layer1.0.weight", "layer1.1.running_var", "layer1.1.bias", "layer1.1.weight", "layer1.1.running_mean", "layer2.0.output.0.0.weight", "layer2.0.output.0.1.running_var", "layer2.0.output.0.1.bias"........
can you give me some tips to fix this bugT_T.
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/DrSleep/light-weight-refinenet/issues/38?email_source=notifications&email_token=AB32NGAW4Y6YXPBAPWKF2ODP5AHF7A5CNFSM4HNTJPAKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODY4DLKQ#issuecomment-507000234, or mute the thread https://github.com/notifications/unsubscribe-auth/AB32NGA7AVB5FXKBUWR4UITP5AHF7ANCNFSM4HNTJPAA .
@DrSleep T_T
Can you provide the details of the hyperparams Settings @DrSleep
I adapted the function
create_segmenter
in train.py, by adding a branch as follow:I also adapted other necessary codes such as dataloader and then I ran the code using provided
mbv2
on NYU and cityscapes. But I got very low mIoUs on both these datasets. The results were similar to follow and almost invariant during training:It seemed that the model classify all pixels as background.
I also ran the provided resnet101 and its results were normal. Is there something wrong with the code in
mobilenet.py
?