Fixing user inputs for low, high in make_tensor

[krshrimali]

TODOs:

• Do not clamp inputs for low and high when given and valid.
• Devise rules for modifying low and high when extremals/invalid values passed.
• Testing with test_references_numerics_hard with the revised changes. (I've tested locally, the changes will take place in a separate PR though after offline discussion with @mruberry)
• Revise comments/documentation for make_tensor

See #61758 for tracker issue.

cc: @mruberry @pmeier

[facebook-github-bot]

🔗 Helpful links

• 🧪 See artifacts and rendered test results at hud.pytorch.org/pr/61108
• 📄 Preview docs built from this PR

💊 CI failures summary and remediations

As of commit cd166fa (more details on the Dr. CI page):

---

💚 💚 Looks good so far! There are no failures yet. 💚 💚

---

This comment was automatically generated by Dr. CI (expand for details).

Follow this link to opt-out of these comments for your Pull Requests.

Please report bugs/suggestions to the (internal) Dr. CI Users group.

Click here to manually regenerate this comment.

[pmeier]

Thanks @krshrimali for the PR. Apart from my inline comments, I remember that discussed using torch.distributions.Uniform as the base sampler rather than torch.rand. What became of that plan?

[krshrimali]

Thanks @krshrimali for the PR. Apart from my inline comments, I remember that discussed using torch.distributions.Uniform as the base sampler rather than torch.rand. What became of that plan?

Thanks for the review, @pmeier - I have addressed a couple of your comments, will take a better look and update this PR with changes suggested. Regarding using: torch.distriutions.uniform.Uniform rather than torch.rand, I don't exactly remember why I reverted the commit. Will have to take a look again and see if it broke any relevant test. I'll get back to you on this after some testing. Thanks!

[krshrimali]

This is ready for another round of review, unless and until there are unexpected test failures. A couple of comments still require attention, I'm looking at them. Thanks!

Leaving some sample outputs from modified make_tensor:

Click to expand!

>>> make_tensor((1, 2), device='cpu', dtype=torch.float32, low=-9, high=9)
tensor([[-8.4784, -1.7658]])
>>> make_tensor((1, 2), device='cpu', dtype=torch.uint8, low=-9, high=9)
tensor([[3, 2]], dtype=torch.uint8)
>>> make_tensor((1, 2), device='cpu', dtype=torch.uint8, low=-9, high=None)
tensor([[6, 4]], dtype=torch.uint8)
>>> make_tensor((1, 2), device='cpu', dtype=torch.uint8, low=None, high=None)
tensor([[2, 5]], dtype=torch.uint8)
>>> make_tensor((1, 2), device='cpu', dtype=torch.float32, low=None, high=None)
tensor([[ 3.1285, -3.0337]])
>>> make_tensor((1, 10), device='cpu', dtype=torch.float32, low=None, high=None)
tensor([[ 5.1067,  1.1351,  4.9473,  5.7744, -3.9730,  3.2708, -3.8939,  2.8211,
         -4.7024,  4.1631]])
>>> make_tensor((1, 10), device='cpu', dtype=torch.float32, low=None, high=float('inf'))
tensor([[2.0459e+38, 1.0356e+38, 8.6688e+37, 2.1416e+38, 3.2889e+38, 2.5179e+38,
         1.5371e+38, 1.6188e+38, 2.6685e+38, 5.1892e+37]])
>>> make_tensor((1, 2), device='cpu', dtype=torch.float32, low=float('inf'), high=float('inf'))
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/krshrimali/pytorch/torch/testing/_internal/common_utils.py", line 2047, in make_tensor
    low, high = _modify_low_high(ranges_floats, low, high, default_values=(-9, 9))
  File "/home/krshrimali/pytorch/torch/testing/_internal/common_utils.py", line 2028, in _modify_low_high
    low = _for_val('low')
  File "/home/krshrimali/pytorch/torch/testing/_internal/common_utils.py", line 2008, in _for_val
    raise ValueError(f"Found invalid value {val} for {inp}")
ValueError: Found invalid value inf for low
>>> make_tensor((1,2), 'cpu', torch.float32, low=float('nan'), high=float('nan'))
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/krshrimali/pytorch/torch/testing/_internal/common_utils.py", line 2048, in make_tensor
    low, high = _modify_low_high(ranges_floats, low, high, default_values=(-9, 9))
  File "/home/krshrimali/pytorch/torch/testing/_internal/common_utils.py", line 2029, in _modify_low_high
    low = _for_val('low')
  File "/home/krshrimali/pytorch/torch/testing/_internal/common_utils.py", line 2009, in _for_val
    raise ValueError(f"Found invalid value {val} for {inp}")
ValueError: Found invalid value nan for low

cc: @pmeier @mruberry

[mruberry]

Overall this looks good, @krshrimali. I just made a couple suggestions and then we should wait for the tests

[krshrimali]

This is ready for a final review @mruberry - thanks for your patience and help. PTAL, whenever you find the time. :)

[mruberry]

Cool; finally a make_tensor() that respects the user's low and high!

[facebook-github-bot]

@mruberry has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[facebook-github-bot]

@mruberry merged this pull request in fcc1f87.