camlaedtke / segmentation_pytorch Goto Github PK

Hello, I would like to ask you if its possible to use your implementation for a binary segmentation task on a grayscale image data. And if yes, what the simplest way would look like, taking into consideration the loss function and the model parameters, such as num_classes or in_chans?

Thank you

Hi, Thanks for contributing to this repo, which makes me easier to understand SegFormer's architecture.
Did you validate the performance achieved by this implementation? Could it reach the results reported in the paper?
I would appreciate it a lot for your answer.
Best regards

missing keys in source state_dict: linear_c4.proj.weight, linear_c4.proj.bias, linear_c3.proj.weight, linear_c3.proj.bias, linear_c2.proj.weight, linear_c2.proj.bias, linear_c1.proj.weight, linear_c1.proj.bias, linear_fuse.weight, linear_fuse.bias, linear_pred.weight, linear_pred.bias

Hello author, could you please tell me which pre-trained models are used in the code, such as mit_b0.pth?
Thank you very much for your help.

First of all, thank you very much for the great implementation of SegFormer!

While working with your SegFormer version and the pre-trained weights of the original paper, I noticed some minor inconsistencies.

1. The defaults in the constructor are mixed from different SegFormer versions (or from no version at all).
   • The combination of embed_dims=[64, 128, 256, 512], is not mentioned in the paper
   • depths=[3, 6, 40, 3] and mlp_ratios=[4, 4, 4, 4] are from version b5
   • decoder_dim = 256 is from version b0
2. There is a sneaky batch norm layer in the original code that your implementation is missing.

Further I managed to convert the pre-trained weights from the original repo to your implementation. Would you like me to add a small conversion util to your repo?

I already opened a PR, but would of course like to discuss the changes here first 🥳.

The Segformer model only accepts a single int for img_size. Does this mean this model can only be used with square images? If not, how can I use rectangular images?

Thank you.

Hello
How are you?
Thanks for contributing to this project.
I found that the sigmoid/softmax activation is NOT applied for the model's output when it is evaluated.
Please check the function "validate" in your notebook file "segmentation_segformer.ipynb".
Could u explain the reason?
Thanks