How to configure cfg.yaml file?

[DrWillway]

Hello! I am using bonnetal for semantic segmentation task. It works and I would like to ask you how to configure the parameters in cfg.yaml file to make the results better. I have only one class to segment and the second is background as in the example you provided.

If there is any documentation of the meaning of each of the parameters?

[tano297]

Hello @DrWillway, what specifically do you want to configure? :)

[DrWillway]

To start with I would like to know how to configure img_means, img_stds and img_prop

[tano297]

So:

• img_means and img_stds depends on your dataset. They represent the means and standard deviation of your R, G, and B channels. There is a script to generate these values from your data, which takes in a list of image paths.

• img_prop represents the size you will use to infer your neural network after training, during inference time. It usually also corresponds to the size of the images in your training data. If you check the cityscapes parser it provides a good example of a dataset that has fixed size image inputs, and the persons parser provides a good example of how to deal with datasets that contain images of different sizes. This parameter is, again, for INFERENCE. For training, crops of the images of the size ["train"]["crop_prop"] will be used. The best way to understand this is by looking at the parsers.

[DrWillway]

Thank you very much, I will update you later :)

[tano297]

any updates on this?

[tano297]

I'm closing this for inactivity. If you have any other doubts, let me know