oniroai / semantic-segmentation-with-mobilenetv3 Goto Github PK

Thanks for the great work!
Your model can do inference with very high fps!

I am trying to train model with your train code. And I found the process is very slow and does not use GPU.
I read your code and I found the line where GPU is enabled is commented out.

    def fit(self, best_on_val=True):
        # with tf.device(self.device):

1. Why did you comment out the GPU line?
2. I have modified the line to enable the GPU, but the process doesn't become fast. (GPU usage is about 10% or so with RTX 2080Ti). Do you have any idea about it?

Wanted to test this out but couldn't get access to the models for testing.

Thanks for your nice work !
How to allow programs in Ubuntu environment, not docker, cannot allow make build

Hi there! Amazing repo you have here. I already tried both training and inferencing but i have few question to ask.

Do you have the image prediction output for this repository that you already tried before?
Cuz i already tried both training and inferencing using this repo, but the result was really bad.
I trained with 356 Image-Mask data.

For _dataset.py, why there are a lot of complicated process like ImageTargetDataset, RandomConncatDataset etc.?
Why can't just use normal image loader practice like in keras or maybe just load images-mask using opencv and append them on list instead of appending their path?

in vis_dataset(), why i can only visualize the Image data and not the Target(Mask)? I tried but errors occurred. How can i know the images have correct corresponding mask?

in Testing Image section at train-mobilenet.ipynb, what exactly would be the output? When i tried, it don't even display/show the prediction mask. It outputs back the original test image. When i try to output "out_img" variable, noisy images were displayed.

Hi, thanks for the project. It's really interesting to play and learn from it. However I got ValueError: Shapes (1, 1, 24, 2) and (1, 1, 24, 1) are incompatible when trying to load the pretrained model. Seems like it has some problem while loading weights from the model. I also noticed that you mentioned the pretrained model was trained with TF2.0. Could that be the cause of this issue? I am with TF2.3 btw.

I took a look at a bunch of segmentation based mv3 repos on github. Most including yours did not implement the atrous convolution on the final convolution block unlike in the paper. Any particular reason for this?

Could you please re-upload the pretrained model to google drive? I think they broke.

Would it be possible to provide a pre-trained model for quick evaluation?

Many thanks in advance!