Loss Functions and Segmentation metrics

[ansariyusuf]

trafficstars

Hi,

I trained the model as shown in the tutorial and noticed that the default loss function is cross-entropy. Is there a way I can train the network with other loss functions, like dice loss, IoU loss, combo loss, etc? Also, I would appreciate it if there is a way to evaluate the performance of the model on different segmentation metrics, like, precision, accuracy, IoU, dice coefficient, etc.

Kindly assist and provide potential directions

Yusuf

[h-sami-ullah]

yes, you can do that, you will need to play with train.py file, maybe add the loss you are interested in and call it under Line 114 https://github.com/divamgupta/image-segmentation-keras/blob/dc830bbd76371aaedbf8cb997bdedca388c544c4/keras_segmentation/train.py#L114

[mavaylon1]

Changing to custom loss and metric still reports accuracy and cross-entropy. It doesn't have errors, it simply just reports accuracy even if you change it.