Squeeze making wonders? Bias is -ve?

[MuhammadBilal848]

I wrote the Linear Regression Class a bit differently using nn.Linear method. The thing I noticed is that, in the video (Pytorch workflow), we were assigning the nn.Parameters(torch.rand()) so rand wasn't assigning the w and b as negative, so the model was taking off from positive w and b and was able to successfully converge.

But in nn.Linear(), the method assigns the w and b as negative, and from that point, the model makes efforts to converge and takes multiple attempts to finally make good predictions. I don't get it why?

Also, I tried removing the .squeeze(1) from the forward() method and it was also affecting the model's accuracy. could anyone explain this?