Size mismatch for bn running_mean and running_var, us_mobilenet_v1

I trained the us_mobilenet_v1 on a different dataset and when loading the best_model_bn_calibrated.pt, I get lots of size mismatch errors for bn running_mean and running_var.
Any idea on how to overcome these errors?
Also, in your opinion is is possible to train the universally slimmable model to lower widths than 0.25 (e.g. 0.1, 0.15) ?