Almost everything is seen as OK with default settings

Question

Almost everything is seen as OK with default settings

JulienMaille opened this issue 5 years ago · 17 comments

I trained segmentation and then decision with your default settings and the network barely detect anything. I get 99% of "OK" on the test dataset.
Did I missed something?

Answer 1 · 2019-07-30T13:31:06.000Z

Hi,
i have the same behaviour. Did you find an answer?

Answer 2 · 2019-07-30T19:43:05.000Z

Not really, I tried to setup a Google Colab notebook.
https://colab.research.google.com/drive/101HUaOo3PskWp7h6DucR50pIGxO_SX2m
The loss during training doesn't look good to me

Answer 3 · 2019-07-31T11:09:15.000Z

Hi,

i got it working. When you train the segmentation network please use only the defect inside Train_NG with lr 0.001. Train it for 100 epochs as default. When you train the decision network change the lr=0.1 with Adam or add some lr_scheduler like cycle learning. With these parameters i got both defect and good classification. Let me know if it is working also for you. Thanks

Answer 4 · 2019-07-31T19:10:31.000Z

Ok I will try that. Have you tried training it on your own dataset?

Answer 5 · 2019-08-02T10:17:15.000Z

Yes i have tried but with no good results

Answer 6 · 2019-09-04T08:21:57.000Z

when i using defualt settings in training, the output mask in "testResultSeg" is completely black, no defect was detected,same In all pictures of all epoch (100 epoch ).
How can i fix it.
thanks.

Answer 7 · 2019-09-09T06:32:15.000Z

when i using defualt settings in training, the output mask in "testResultSeg" is completely black, no defect was detected,same In all pictures of all epoch (100 epoch ).
How can i fix it.
thanks.

hi, I have the same problem(the output mask in "testResultDec" is completely black), have you sovled it?

Answer 8 · 2019-09-10T09:36:13.000Z

Hi @SampleW ,

have you tried with the images provided or with your own dataset?

Answer 9 · 2019-11-24T06:01:35.000Z

ValueError: num_samples should be a positive integer value, but got num_samples=0
how can i fix it?

Answer 10 · 2019-11-24T06:02:31.000Z

@eric-tc
@Nefast

Answer 11 · 2019-11-29T02:01:39.000Z

@LoveIsAGame
The same problem, how do you solve it？

Answer 12 · 2019-11-29T02:04:03.000Z

Tried to python train_segment.py
ValueError: num_samples should be a positive integer value, but got num_samples=0

THANKS

Answer 13 · 2019-11-29T08:38:44.000Z

Changed ".jpg" to ".bmp" at dataset.py .

ok

Answer 14 · 2020-06-19T06:15:56.000Z

try to use sigmoid but not rule in the last which is named layer 5

Answer 15 · 2020-06-29T02:33:34.000Z

oh,Tried to python train_segment.py is ok, but tried to python train_decision.py is ValueError: num_samples should be a positive integer value, but got num_samples=0, why? how do you solve it?
thanks.

Answer 16 · 2021-01-17T15:22:14.000Z

try to use sigmoid but not rule in the last which is named layer 5

Still not working. I doubt the implementation can work tbh

Answer 17 · 2024-04-16T05:31:53.000Z

try to use sigmoid but not rule in the last which is named layer 5

Still not working. I doubt the implementation can work tbh
just no act function，because loss function has the act.