Using images with mutiple channels in Ganomaly

[marchgeo]

Hello,
While using GANomaly to process 3-channel (RGB) images, I came across functionality that appears to support multi-channel input, such as RGB+NIR. According to the Anomalib documentation, this can be configured using the num_input_channels parameter. However, I encountered difficulties setting this parameter via the CLI.
It seems that when invoking GANomaly through the CLI, the num_input_channels parameter is defined within the GanomalyModel, not directly exposed in the configuration file.
Has anyone experienced this issue or found a workaround? I'd appreciate any guidance or suggestions.

[abc-125]

Hello,
The fastest way probably will be to install Anomalib from source and change num_input_channels in the Ganomaly class:

anomalib/src/anomalib/models/image/ganomaly/lightning_model.py

Lines 158 to 165 in f6ec1c5

	self.model = GanomalyModel(
	input_size=self.input_size,
	num_input_channels=3,
	n_features=self.n_features,
	latent_vec_size=self.latent_vec_size,
	extra_layers=self.extra_layers,
	add_final_conv_layer=self.add_final_conv_layer,
	)

You will also need to change or turn off validation here to allow usage of an image with four channels:

anomalib/src/anomalib/data/validators/torch/image.py

Lines 516 to 518 in f6ec1c5

	if image.shape[1] != 3:
	msg = f"Image must have 3 channels, got {image.shape[1]}."
	raise ValueError(msg)

[marchgeo]

Thank you, I will try it.