heatmap_size parameter vs adjusting heatmap size

[smidm]

The top down models feature the data_cfg['heatmap_size'] parameter. It seems that the parameter is used only during the training in the target generation.

Is the parameter only a workaround to get the actual heatmap size determined by the number of deconvolution layers in the top down network head?

What is the proper way to change the size of the heatmap while keeping the input size constant? Adjust the number of deconvolution layers (num_deconv_layers) in the head and set the heatmap_size to the head ouput dimensions?

Thanks!