Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Where's DeepLabv3+ model Decoder? #26

Open
Albert-learner opened this issue Apr 6, 2022 · 3 comments
Open

Where's DeepLabv3+ model Decoder? #26

Albert-learner opened this issue Apr 6, 2022 · 3 comments

Comments

@Albert-learner
Copy link

Hello, I'm trying to study about your code. But when I analysis your code, I can't find DeepLabV3+ Decoder in your code. Could you please tell me where the DeepLabV3+ model's Decoder?

@Albert-learner
Copy link
Author

And I have another question. You're using one of Semantic Segmentation model, DeepLab V3+. But I know that the result of Semantic Segmentation result is [height, width, channels], and apply argmax so that I could get one channel array. When I analyze the result of Semantic Segmentation ath NYU Depth V2 dataset, the result array(inference DeepLab V3+) consists of different min cost and max cost at each channel array.
I'd like to know that why result array of each category's range is different.

@charlesCXK
Copy link
Owner

Hello, I'm trying to study about your code. But when I analysis your code, I can't find DeepLabV3+ Decoder in your code. Could you please tell me where the DeepLabV3+ model's Decoder?

Hi, DeepLabV3+ decoder is defined here.

@charlesCXK
Copy link
Owner

And I have another question. You're using one of Semantic Segmentation model, DeepLab V3+. But I know that the result of Semantic Segmentation result is [height, width, channels], and apply argmax so that I could get one channel array. When I analyze the result of Semantic Segmentation ath NYU Depth V2 dataset, the result array(inference DeepLab V3+) consists of different min cost and max cost at each channel array. I'd like to know that why result array of each category's range is different.

The absolute values of the result array do not make much sense.

  • During training, the result array is supervised by the CrossEntropy loss, which pushes the confidence of the correct category to be higher than other categories. In the meanwhile, the SoftMax operation in CrossEntropy loss hopes that the confidence of the correct category is significantly different from the other categories. For other categories, we don't care what their confidence numbers are, just make them as low as possible.
  • During inference, we simply select the category with the maximum confidence as the prediction, which is consistent with the training.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants