Problem generalising 01. PyTorch Workflow Fundamentals model_0 to different data

[amadanmath]

I have tried to play with the model_0 from https://www.learnpytorch.io/01_pytorch_workflow/
When I execute it as written (except for increasing epochs), I can see both training and evaluation shrink, and the parameters end up very close to the gold around 170th epoch:

Epoch: 170 | MAE Train Loss: 0.008932482451200485 | MAE Test Loss: 0.005023092031478882
OrderedDict([('weights', tensor([0.6951])), ('bias', tensor([0.2993]))])

For the record, here are the dataset and gold parameters used:

# Create *known* parameters
weight = 0.7
bias = 0.3

# Create data
start = 0
end = 1
step = 0.02

However, when I just change the dataset parameters to this:

start = 1
end = 101
step = 1

the loss and the state dictionary stabilise to this, around 360th epoch:

Epoch: 360 | MAE Train Loss: 12.168511390686035 | MAE Test Loss: 9.61581039428711
OrderedDict([('weights', tensor([0.8074])), ('bias', tensor([0.1928]))])

which is not really close to the gold parameter values. Why was the training unsuccessful here, and what else should be adjusted in order for this to work as expected? Does the training process assume the input data is between 0 and 1? If so, which part of the training loop code is sensitive to this? (It may be somewhere later in the lessons, but I have not yet gotten there.)

It seems to only happen when X is large; when I made bias and weight larger, so that y was large, the training converged to the correct answer (though I had to increase epochs and learning rate in order to reach it). The thing I don't understand is why making X larger made the result to stabilise on a wrong value (as opposed to e.g. not converge at all).

[PastorJordi]

I had exactly the same problem when playing around with the exercises.
Try standardizing your data to see whether it helps :)

Here's the SO solution I found that helped me