Predicted normals are all NaN

[eric11220]

As mentioned in #5,
The output of "layer_4" is NaN for Line 71 in modules/geo_predictors/omnidata/modules/midas/dpt_depth.py
layer_1, layer_2, layer_3, layer_4 = forward_vit(self.pretrained, x)

The pre-trained checkpoint seems to load successfully for the normal prediction model.
How to fix the issue? thanks!

[gard-n]

I describe a workaround here. It would be great if there was a solution where you didn't have to manipulate the backbone.