add tensor_hash

[lessw2020]

Summary:
This adds a tensor hash function for multi-dim tensors.
It's purpose is to allow easy full tensor verification on forward tensors for other unit tests
by storing the hash of the answer tensor.
This strikes a middle ground between simple verification (i.e. first row, mean of an axis) and every item in the tensor verification which would require storing the entire tensor in a unit test.

Usage =
hash the answer tensor, save in your unit test.
Run a forward in unit test, hash the output tensor.
Compare hashed tensors to confirm unit test result.

Test plan:
Verified a + .01 change to single item in a multi-dim tensor is detected, verified identical tensors generate identical hashes.
Since this is a unit test function, not sure that unit test for a unit test function is needed as will not change unless function itself is modified.

Fixes #{issue number}

[lessw2020]

unit test failure is not related to this PR

[pbontrager]

Could you provide a little more context for this function? It seems to compare sums for the final dimension, does this provide more precision than mean? It's just one extra dimension of data. I guess this would be more likely to catch tensors that had the right data but the wrong shape, is that the primary use case? To add clarity to my question, I am wondering what the potential mistakes are that are not caught by mean or mean(-1) and if this implementation of a hash catches all of them.

[lessw2020]

Hi @pbontrager,
My concern with the mean test is that it's an average, and multiple combinations of numbers could potentially end up at the same mean, so I felt the mean test alone is not sufficient to fully check a tensor.

I can't guarantee that a hash will catch all issues and probably using both would result in the best coverage, but I show a simple example below where using the mean fails to detect the difference of .01 to an entry, vs the hash does using the default all_close settings that are in the multi-modal tests.

Here adding .01 to a single entry is not detected by mean (see line 18 allclose = True), but is detected with the hashing.

[rohan-varma]

just testing

[rohan-varma]

test

[facebook-github-bot]

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

[facebook-github-bot]

@ebsmothers merged this pull request in 1034ed7.