Running on a data without test labels

[absagargupta]

Hello there

Suppose I have trained my model on a dataset having the images as well as masks in the training as well as test set (Cityscape dataset). Is it possible that after training it on the (cityscape) dataset, I run it for creating masks of MSMT17_V2 image dataset? If yes what all changes should I do? Thank you well in advance for the reply :-)

[qubvel]

Hi, @absagargupta
You can make prediction on any images/any dataset (make sure to preprocess them same as in training stage), but model will produce label it have been trained on.

[absagargupta]

Thank you so much for the prompt reply. In case I need just the labels of people and the background what changes should I take care of?

[qubvel]

Suppose you have trained you model for C=10 classes, where value 3 is people class.
Output of network will be a tensor output with shape (N, H, W, 10).
To get mask of people class:
mask = (np.argmax(output, axis=-1) == 3)

[absagargupta]

Thank you. :-)