pytorch-fcn icon indicating copy to clipboard operation
pytorch-fcn copied to clipboard

why all my output results of the fcn32 is ZERO

Open daiyp3360xxxxxisxxxxxxmine opened this issue 5 years ago • 16 comments

after the fcn32 was trained for 9 epochs, the output value score of it is still zero. however, the weight of each layers are not zero. i am wondering if somebody has suggestions to fix this problem.

I think you should try using training data as validation data, and check if it works. If it doesn't work, that means you have a serious bug in your code.

wkentaro avatar Apr 09 '19 14:04 wkentaro

I have the sane problem with fcn8s, help anyone?

jdavidd avatar Apr 14 '19 11:04 jdavidd

after the fcn32 was trained for 9 epochs, the output value score of it is still zero. however, the weight of each layers are not zero. i am wondering if somebody has suggestions to fix this problem.

Hi,Did you solved this problem?I trained 50 epoch,but when I use the trained model to test,the output is zero.

BCWang93 avatar May 02 '19 11:05 BCWang93

I have the sane problem with fcn8s, help anyone?

Hi,Did you solved this problem?I trained 50 epoch,but when I use the trained model to test,the output is zero.

BCWang93 avatar May 02 '19 11:05 BCWang93

I have the sane problem with fcn8s, help anyone?

For training fcn8s, you first start training fcn32s and fcn16s.

wkentaro avatar May 02 '19 11:05 wkentaro

I have the sane problem with fcn8s, help anyone?

For training fcn8s, you first start training fcn32s and fcn16s.

Hi,I trained the fcn32 network,but I also has this problem,the output of the network is always zero.

BCWang93 avatar May 02 '19 12:05 BCWang93

I found that in the initialization function, weights are initialized to zero, which may cause the problem. Does anyone knows why the fcn8s need to initialize all the weights to zero?

ShawnNew avatar May 28 '19 17:05 ShawnNew

I found that in the initialization function, weights are initialized to zero, which may cause the problem. Does anyone knows why the fcn8s need to initialize all the weights to zero?

Because of this https://github.com/wkentaro/pytorch-fcn/issues/40

wkentaro avatar May 28 '19 18:05 wkentaro

I didn't solve this problem yet, for these days I was tracked by some trifles, I am just going to resume this work. I do think this problem might caused by the initial strategy. i will response here as soon as i solve this problem.

daiyp3360 avatar May 29 '19 02:05 daiyp3360

I didn't solve this problem yet, for these days I was tracked by some trifles, I am just going to resume this work. I do think this problem might caused by the initial strategy. i will response here as soon as i solve this problem.

Hi,Did you solved this problem?Thanks!

BCWang93 avatar Jun 03 '19 02:06 BCWang93

I have the same problem with fcn32s, all output results are zeros. Anyone can help ?

phucvu460 avatar Sep 24 '19 10:09 phucvu460

I have the same problem, all output results are zeros. Anyone can help ?

prfans avatar Nov 20 '19 02:11 prfans

Hi, @daiyp3360xxxxxisxxxxxxmine , have figure out the problem yet?

songyuc avatar Jan 17 '20 06:01 songyuc

Hi @wkentaro , I think #40 may have no relation to the initiation strategy, so it cannot account for the problem here.

songyuc avatar Jan 17 '20 06:01 songyuc

I have same problem, did you solve it? @daiyp3360xxxxxisxxxxxxmine

hylovezhang avatar Mar 25 '20 08:03 hylovezhang

can you run the code?i have lots of problems about this code.Could you please share me your code?My email:[email protected]。Thank you very much!!!

liul-ing avatar Nov 06 '21 07:11 liul-ing