Retraining the model

[qubvel]

Retraining the model

[qubvel]

You have to pass for each image (e.g. shape is HxWx3) corresponding mask (labeled image, where each pixel assigned to specific class) with shape HxWxN, where N is a number of classes you have (+ background for multiclass problems). Read more about semantic segmentation problem.

If have more than one class pass kwargs: classes=<num_classes>, activation='softmax'

[qubvel]

You`re welcome)
P.S. Great repo with information about datasets and networks for semantic segmentation https://github.com/mrgloom/awesome-semantic-segmentation

[chm60]

Thanks :) - That saved me a lot of time. Another question if I may, do you have any resources that already contain labelled training images in that specific format, i'm struggling to find a reliable and accurate dataset.

[qubvel]

Try this one 😄

[chm60]

Thank you again :D