seanxyz/segdecnet

Pytorch Implement of the paper "Segmentation-Based Deep-Learning Approach for Surface Defect Detection"

Pytorch Implement of the paper "Segmentation Based Deep-Learning Approach for Surface DefectDetection" https://arxiv.org/abs/1903.08536

Modified little things without changing the accuracy

step1: train segment net

python train_segment.py

step2: train decision net

python train_decision.py

step3: test

python test.py

segdecnet's Issues

Accuracy is not up to par

I ran the code, but the accuracy was not as good as in the paper.

about initialize_weight

I think the function of initialize_weight can be:
def initialize_weights(*models):
for model in models:
for m in model.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight.data, nonlinearity='relu')
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1.)
m.bias.data.fill_(1e-4)
elif isinstance(m, nn.Linear):
m.weight.data.normal_(0.0, 0.0001)
m.bias.data.zero_()

While using ReLU, Kaiming initializing is better.

From raw KolektorSDD dataset to NG and OK

Hi,

Thanks for sharing! Would you mind share how you convert raw KolektorSDD to your NG/OK split dataset? I am working to reproduce all result from the paper(eg different dilation), however the shared GPU do not have cuda 9.0, I would like to reproduce using pytorch and python.

Thanks

Question regarding compared models

My bad, I thought you were part of the team who wrote the paper.

about valid result

I trained the segmentation network and got the best result in 50 epochs. However, after 60 epochs the valid result is vanishing. Do you know why it is? Thank you!

error

raise ValueError("num_samples should be a positive integer "
ValueError: num_samples should be a positive integer value, but got num_samples=0

Tried to python train_segment.py

about Target size value error

I'm not able to train the model, since the loss function gives me back a Value error on the target size... do you know ho to fix it?

weird error

File "C:\Users\Administrator\anaconda3\envs\pytorch\lib\multiprocessing\spawn.py", line 138, in _check_not_importing_main
raise RuntimeError('''
RuntimeError:
An attempt has been made to start a new process before the
current process has finished its bootstrapping phase.

    This probably means that you are not using fork to start your
    child processes and you have forgotten to use the proper idiom
    in the main module:

        if __name__ == '__main__':
            freeze_support()
            ...

    The "freeze_support()" line can be omitted if the program
    is not going to be frozen to produce an executable.

Question about custom dataset

Hi, if i want to train my own datasets, how to create the mask images e.g. *_label.bmp?

Wierd Error

When running train_segment.py and train_decision.py, the error occurs:

usage: ipykernel_launcher.py [-h] [--cuda CUDA] [--gpu_num GPU_NUM]
[--worker_num WORKER_NUM]
[--batch_size BATCH_SIZE] [--lr LR] [--b1 B1]
[--b2 B2] [--begin_epoch BEGIN_EPOCH]
[--end_epoch END_EPOCH] [--need_test NEED_TEST]
[--test_interval TEST_INTERVAL]
[--need_save NEED_SAVE]
[--save_interval SAVE_INTERVAL]
[--img_height IMG_HEIGHT] [--img_width IMG_WIDTH]
ipykernel_launcher.py: error: unrecognized arguments: -f /root/.local/share/jupyter/runtime/kernel-b7c2d070-caa1-4edd-9d2d-2fcf3fd83b3e.json

An exception has occurred, use %tb to see the full traceback.

SystemExit: 2

/opt/conda/lib/python3.6/site-packages/IPython/core/interactiveshell.py:2971: UserWarning: To exit: use 'exit', 'quit', or Ctrl-D.
warn("To exit: use 'exit', 'quit', or Ctrl-D.", stacklevel=1)

dataset

Hello, thanks for your share. I am confuse about why you divide the data set into trainOK and trainNG. What do NG and OK mean respectively

Tried to python train_segment.py

About Licensing...

Can you add a licence to this repository @seanXYZ ? Would be very nice to use some of your awesome Code! :)

In segmentation training process, the model can't converge

I try to train the segmentation network in the koltektorSDD Dataset, either the MSEloss, or the BCEloss(), the training process cannot converge~~no matter the lr = 0.005, 0.01,0.1 ...

I use the dilation = 5, and scale(704, 256), and SGD optimizer

Can anyone give me any suggestion?

about loss function

In this paper, the author said the loss function is perfectly crossentropy, but when I modify the MSELoss to CrossEntropy Loss. The code do not work. How to modify the code better? Thank you!

Almost everything is seen as OK with default settings

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?