l2_reg_ortho returns different values for frozen model weights

[souravdalai18]

I tried running

model = model.cpu().eval()
with torch.no_grad():
    print(l2_reg_ortho(model.visual.trunk))

it returns different values everytime