Add torch.unique_consecutive #19060

zasdfgbnm · 2019-04-09T06:31:12Z

Fixes: #19045

This is independent on the #18649 series. This will cause merge conflicts in #18649 series, but please merge this first, and I will resolve the merge conflicts there.

The new feature is exposed in _unique2_temporary_will_remove_soon and _unique_dim2_temporary_will_remove_soon. But not at torch.unique yet. I will take care of the API after #18649 series get merged completely.

Benchmark on a tensor of shape torch.Size([15320, 2]):

print(torch.__version__)
print()
a = tensor.sort().values.to('cpu')
print('cpu, sorted_input=False:')
%timeit torch._unique2_temporary_will_remove_soon(a)
%timeit torch._unique2_temporary_will_remove_soon(a, return_inverse=True)
%timeit torch._unique2_temporary_will_remove_soon(a, return_counts=True)
%timeit torch._unique2_temporary_will_remove_soon(a, return_inverse=True, return_counts=True)
print()
print('cpu, sorted_input=True:')
%timeit torch._unique2_temporary_will_remove_soon(a, sorted_input=True)
%timeit torch._unique2_temporary_will_remove_soon(a, sorted_input=True, return_inverse=True)
%timeit torch._unique2_temporary_will_remove_soon(a, sorted_input=True, return_counts=True)
%timeit torch._unique2_temporary_will_remove_soon(a, sorted_input=True, return_inverse=True, return_counts=True)
print()
a = a.to('cuda')
print('cuda, sorted_input=False:')
%timeit torch._unique2_temporary_will_remove_soon(a); torch.cuda.synchronize()
%timeit torch._unique2_temporary_will_remove_soon(a, return_inverse=True); torch.cuda.synchronize()
%timeit torch._unique2_temporary_will_remove_soon(a, return_counts=True); torch.cuda.synchronize()
%timeit torch._unique2_temporary_will_remove_soon(a, return_inverse=True, return_counts=True); torch.cuda.synchronize()
print()
print('cuda, sorted_input=True:')
%timeit torch._unique2_temporary_will_remove_soon(a, sorted_input=True); torch.cuda.synchronize()
%timeit torch._unique2_temporary_will_remove_soon(a, sorted_input=True, return_inverse=True); torch.cuda.synchronize()
%timeit torch._unique2_temporary_will_remove_soon(a, sorted_input=True, return_counts=True); torch.cuda.synchronize()
%timeit torch._unique2_temporary_will_remove_soon(a, sorted_input=True, return_inverse=True, return_counts=True); torch.cuda.synchronize()

1.1.0a0+2addccc

cpu, sorted_input=False:
340 µs ± 5.88 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
717 µs ± 14.9 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
52.3 ms ± 2.75 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
52.3 ms ± 1.79 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

cpu, sorted_input=True:
32.8 µs ± 285 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
49.9 µs ± 557 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
51.6 µs ± 1.08 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)
78 µs ± 782 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)

cuda, sorted_input=False:
213 µs ± 1.52 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
291 µs ± 3.81 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
250 µs ± 1.05 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
321 µs ± 1.59 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

cuda, sorted_input=True:
45.6 µs ± 2.13 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)
110 µs ± 2.47 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)
82 µs ± 857 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
143 µs ± 409 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)

print(torch.__version__)
print()
a1, a2 = tensor.unbind(1)
indices = (a1 * tensor.max() + a2).sort().indices 
a = tensor.index_select(0, indices).to('cpu')
print('cpu, sorted_input=False:')
%timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0)
%timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, return_inverse=True)
%timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, return_counts=True)
%timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, return_inverse=True, return_counts=True)
print()
print('cpu, sorted_input=True:')
%timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted_input=True)
%timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted_input=True, return_inverse=True)
%timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted_input=True, return_counts=True)
%timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted_input=True, return_inverse=True, return_counts=True)
print()
a = a.to('cuda')
print('cuda, sorted_input=False:')
%timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0); torch.cuda.synchronize()
%timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, return_inverse=True); torch.cuda.synchronize()
%timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, return_counts=True); torch.cuda.synchronize()
%timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, return_inverse=True, return_counts=True); torch.cuda.synchronize()
print()
print('cuda, sorted_input=True:')
%timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted_input=True); torch.cuda.synchronize()
%timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted_input=True, return_inverse=True); torch.cuda.synchronize()
%timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted_input=True, return_counts=True); torch.cuda.synchronize()
%timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted_input=True, return_inverse=True, return_counts=True); torch.cuda.synchronize()

cpu, sorted_input=False:
55.4 ms ± 1.12 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
55.8 ms ± 616 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
55.2 ms ± 402 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
55.1 ms ± 725 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

cpu, sorted_input=True:
54.7 ms ± 585 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
55.2 ms ± 1.23 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
54.5 ms ± 865 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
54.9 ms ± 577 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

cuda, sorted_input=False:
171 µs ± 783 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
220 µs ± 1.65 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
203 µs ± 2.95 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
251 µs ± 2.83 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

cuda, sorted_input=True:
59.6 µs ± 757 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
113 µs ± 431 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
93.2 µs ± 2.13 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)
147 µs ± 2.81 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)

The CPU implementation of unique_dim is super slow, see #18987, but this PR will not worry about this issue.

zasdfgbnm · 2019-04-09T06:39:54Z

CI is broken...

zasdfgbnm · 2019-04-09T06:40:03Z

@pytorchbot retest this please

zasdfgbnm · 2019-04-09T13:32:56Z

@pytorchbot retest this please

ngimel · 2019-04-09T16:26:55Z

Looks good, but when it's finally merged it will require a documentation for sorted_input argument. Generally though I'm unsure if we need this option - it complicates API, it is error prone (user can pass sorted=True when input is in fact unsorted), and with a slight perf impact the implementation itself can verify if input is sorted or not, and not call sort if input is sorted.
cc @jcjohnson @ezyang for the general API question.
Also, while you are working with unique, can you please look into implementing #16330?

zasdfgbnm · 2019-04-09T16:41:35Z

@ngimel

Looks good, but when it's finally merged it will require a documentation for sorted_input argument.

Yes, I will document this when it's ready to modify API.

Generally though I'm unsure if we need this option - it complicates API, it is error prone (user can pass sorted=True when input is in fact unsorted), and with a slight perf impact the implementation itself can verify if input is sorted or not, and not call sort if input is sorted.

Not exactly, sorted_input actually has another benefit: if the user want the same behavior as thrust::unique or std::unique, then the user could set sorted_input=True to easily achieve this.

[1, 1, 2, 2, 1, 1, 2, 2, 3] --> [1, 2, 1, 2, 3]

The sorted argument is even worse: It basically tells ATen that it might use hashtable to optimize performance, but unfortunately only CPU implementation of unique cares about it. GPU implementation of unique_dim, unique and CPU implementation of unique_dim doesn't care about it

cc @jcjohnson @ezyang for the general API question.

Yes, please comment.

Also, while you are working with unique, can you please look into implementing #16330?

Sure, no problem.

ngimel · 2019-04-09T16:44:51Z

I see. I'd say that passing sorted_input=True for an unsorted input to achieve this result is abuse of the API, but it might have its uses.

zasdfgbnm · 2019-04-09T16:50:35Z

@ngimel Yes, kind of abuse in its current design. But I prefer to design the API in a way that, users clearly knows that this is not an abuse, and encourage them to use this way when needed. Not sure how to do this though. Open to suggestions :)

Maybe adding a new operator called unique_consecutive? But if we do agree this use case worth being merged into PyTorch, we could first just make it the way as in this PR for now (this PR does not introduce any public API, just a modification of a temporary operator), and discuss the API later.

VitalyFedyunin

What will happen if I pass sorted_input and randomly shuffled tensor? Garbage in - garbage out or sygfault?

aten/src/ATen/native/Unique.cpp

zasdfgbnm · 2019-04-09T19:07:45Z

@VitalyFedyunin Actually it is not garbage in, if sorted_input is enabled, then it will have the same behavior as std::unique and thrust::unique

[1, 1, 2, 2, 1, 1, 2, 2, 3] --> [1, 2, 1, 2, 3]

ezyang · 2019-04-09T19:15:45Z

I think I am generally in favor of API renaming so that it's clear that we are doing std/thrust style unique-semantics; this is much better than sorted_input. If you want to merge this patch as is and do the rename in a follow up, that's acceptable me as well.

I'm kind of sympathetic to @ngimel's suggestion that we automatically detect if we can do the faster thing, but that's an enhancement we can always add at some later point in time.

zasdfgbnm · 2019-04-09T19:50:57Z

@ezyang Yes I agree. Do you have suggestions on how to name that new operator? Does unique_consecutive sounds good? Since this is a new operator, I think we could simply name it as what it should be, instead of modifying _unique2_temporary_will_remove_soon and _unique_dim2_temporary_will_remove_soon.

zasdfgbnm · 2019-04-09T19:52:28Z

BTW, I will be very busy in the following couple days, so I might not be able to reply to this.

ezyang · 2019-04-09T20:44:04Z

I like unique_consecutive. Let's do it!

zasdfgbnm · 2019-04-09T22:33:15Z

@VitalyFedyunin @ngimel @ezyang I have refactored this PR to create a new operator torch.unique_consecutive. It has documentation. The modification to the two temporary operators has been reverted. Review comments by @VitalyFedyunin has been resolved.

Please take another look: @VitalyFedyunin :)

Also, could we land this first and then work on #18649? This way the merge conflicts should be a bit easier to resolve...

ezyang · 2019-04-09T22:37:36Z

please 'merge this please' when tests are passing, thx

zasdfgbnm · 2019-04-10T02:59:01Z

@pytorchbot merge this please

facebook-github-bot

@ezyang is landing this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

Summary: Fixes: pytorch/pytorch#19045 Please review: VitalyFedyunin ngimel This is independent on the #18649 series. This will cause merge conflicts in #18649 series, but please merge this first, and I will resolve the merge conflicts there. The new feature is exposed in `_unique2_temporary_will_remove_soon` and `_unique_dim2_temporary_will_remove_soon`. But not at `torch.unique` yet. I will take care of the API after #18649 series get merged completely. Benchmark on a tensor of shape `torch.Size([15320, 2])`: ```python print(torch.__version__) print() a = tensor.sort().values.to('cpu') print('cpu, sorted_input=False:') %timeit torch._unique2_temporary_will_remove_soon(a) %timeit torch._unique2_temporary_will_remove_soon(a, return_inverse=True) %timeit torch._unique2_temporary_will_remove_soon(a, return_counts=True) %timeit torch._unique2_temporary_will_remove_soon(a, return_inverse=True, return_counts=True) print() print('cpu, sorted_input=True:') %timeit torch._unique2_temporary_will_remove_soon(a, sorted_input=True) %timeit torch._unique2_temporary_will_remove_soon(a, sorted_input=True, return_inverse=True) %timeit torch._unique2_temporary_will_remove_soon(a, sorted_input=True, return_counts=True) %timeit torch._unique2_temporary_will_remove_soon(a, sorted_input=True, return_inverse=True, return_counts=True) print() a = a.to('cuda') print('cuda, sorted_input=False:') %timeit torch._unique2_temporary_will_remove_soon(a); torch.cuda.synchronize() %timeit torch._unique2_temporary_will_remove_soon(a, return_inverse=True); torch.cuda.synchronize() %timeit torch._unique2_temporary_will_remove_soon(a, return_counts=True); torch.cuda.synchronize() %timeit torch._unique2_temporary_will_remove_soon(a, return_inverse=True, return_counts=True); torch.cuda.synchronize() print() print('cuda, sorted_input=True:') %timeit torch._unique2_temporary_will_remove_soon(a, sorted_input=True); torch.cuda.synchronize() %timeit torch._unique2_temporary_will_remove_soon(a, sorted_input=True, return_inverse=True); torch.cuda.synchronize() %timeit torch._unique2_temporary_will_remove_soon(a, sorted_input=True, return_counts=True); torch.cuda.synchronize() %timeit torch._unique2_temporary_will_remove_soon(a, sorted_input=True, return_inverse=True, return_counts=True); torch.cuda.synchronize() ``` ``` 1.1.0a0+2addccc cpu, sorted_input=False: 340 µs ± 5.88 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each) 717 µs ± 14.9 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each) 52.3 ms ± 2.75 ms per loop (mean ± std. dev. of 7 runs, 10 loops each) 52.3 ms ± 1.79 ms per loop (mean ± std. dev. of 7 runs, 10 loops each) cpu, sorted_input=True: 32.8 µs ± 285 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each) 49.9 µs ± 557 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each) 51.6 µs ± 1.08 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each) 78 µs ± 782 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each) cuda, sorted_input=False: 213 µs ± 1.52 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each) 291 µs ± 3.81 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each) 250 µs ± 1.05 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each) 321 µs ± 1.59 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each) cuda, sorted_input=True: 45.6 µs ± 2.13 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each) 110 µs ± 2.47 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each) 82 µs ± 857 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each) 143 µs ± 409 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each) ``` ```python print(torch.__version__) print() a1, a2 = tensor.unbind(1) indices = (a1 * tensor.max() + a2).sort().indices a = tensor.index_select(0, indices).to('cpu') print('cpu, sorted_input=False:') %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0) %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, return_inverse=True) %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, return_counts=True) %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, return_inverse=True, return_counts=True) print() print('cpu, sorted_input=True:') %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted_input=True) %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted_input=True, return_inverse=True) %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted_input=True, return_counts=True) %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted_input=True, return_inverse=True, return_counts=True) print() a = a.to('cuda') print('cuda, sorted_input=False:') %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0); torch.cuda.synchronize() %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, return_inverse=True); torch.cuda.synchronize() %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, return_counts=True); torch.cuda.synchronize() %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, return_inverse=True, return_counts=True); torch.cuda.synchronize() print() print('cuda, sorted_input=True:') %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted_input=True); torch.cuda.synchronize() %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted_input=True, return_inverse=True); torch.cuda.synchronize() %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted_input=True, return_counts=True); torch.cuda.synchronize() %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted_input=True, return_inverse=True, return_counts=True); torch.cuda.synchronize() ``` ``` cpu, sorted_input=False: 55.4 ms ± 1.12 ms per loop (mean ± std. dev. of 7 runs, 10 loops each) 55.8 ms ± 616 µs per loop (mean ± std. dev. of 7 runs, 10 loops each) 55.2 ms ± 402 µs per loop (mean ± std. dev. of 7 runs, 10 loops each) 55.1 ms ± 725 µs per loop (mean ± std. dev. of 7 runs, 10 loops each) cpu, sorted_input=True: 54.7 ms ± 585 µs per loop (mean ± std. dev. of 7 runs, 10 loops each) 55.2 ms ± 1.23 ms per loop (mean ± std. dev. of 7 runs, 10 loops each) 54.5 ms ± 865 µs per loop (mean ± std. dev. of 7 runs, 10 loops each) 54.9 ms ± 577 µs per loop (mean ± std. dev. of 7 runs, 10 loops each) cuda, sorted_input=False: 171 µs ± 783 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each) 220 µs ± 1.65 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each) 203 µs ± 2.95 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each) 251 µs ± 2.83 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each) cuda, sorted_input=True: 59.6 µs ± 757 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each) 113 µs ± 431 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each) 93.2 µs ± 2.13 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each) 147 µs ± 2.81 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each) ``` The CPU implementation of `unique_dim` is super slow, see pytorch/pytorch#18987, but this PR will not worry about this issue. Pull Request resolved: pytorch/pytorch#19060 Differential Revision: D14866909 Pulled By: ezyang fbshipit-source-id: d20012cec68c37b05cf770a6f4d6524f910b950f

facebook-github-bot · 2019-04-10T15:07:40Z

@ezyang merged this pull request in ea2405c.

Summary: Fixes: pytorch#19045 Please review: VitalyFedyunin ngimel This is independent on the pytorch#18649 series. This will cause merge conflicts in pytorch#18649 series, but please merge this first, and I will resolve the merge conflicts there. The new feature is exposed in `_unique2_temporary_will_remove_soon` and `_unique_dim2_temporary_will_remove_soon`. But not at `torch.unique` yet. I will take care of the API after pytorch#18649 series get merged completely. Benchmark on a tensor of shape `torch.Size([15320, 2])`: ```python print(torch.__version__) print() a = tensor.sort().values.to('cpu') print('cpu, sorted_input=False:') %timeit torch._unique2_temporary_will_remove_soon(a) %timeit torch._unique2_temporary_will_remove_soon(a, return_inverse=True) %timeit torch._unique2_temporary_will_remove_soon(a, return_counts=True) %timeit torch._unique2_temporary_will_remove_soon(a, return_inverse=True, return_counts=True) print() print('cpu, sorted_input=True:') %timeit torch._unique2_temporary_will_remove_soon(a, sorted_input=True) %timeit torch._unique2_temporary_will_remove_soon(a, sorted_input=True, return_inverse=True) %timeit torch._unique2_temporary_will_remove_soon(a, sorted_input=True, return_counts=True) %timeit torch._unique2_temporary_will_remove_soon(a, sorted_input=True, return_inverse=True, return_counts=True) print() a = a.to('cuda') print('cuda, sorted_input=False:') %timeit torch._unique2_temporary_will_remove_soon(a); torch.cuda.synchronize() %timeit torch._unique2_temporary_will_remove_soon(a, return_inverse=True); torch.cuda.synchronize() %timeit torch._unique2_temporary_will_remove_soon(a, return_counts=True); torch.cuda.synchronize() %timeit torch._unique2_temporary_will_remove_soon(a, return_inverse=True, return_counts=True); torch.cuda.synchronize() print() print('cuda, sorted_input=True:') %timeit torch._unique2_temporary_will_remove_soon(a, sorted_input=True); torch.cuda.synchronize() %timeit torch._unique2_temporary_will_remove_soon(a, sorted_input=True, return_inverse=True); torch.cuda.synchronize() %timeit torch._unique2_temporary_will_remove_soon(a, sorted_input=True, return_counts=True); torch.cuda.synchronize() %timeit torch._unique2_temporary_will_remove_soon(a, sorted_input=True, return_inverse=True, return_counts=True); torch.cuda.synchronize() ``` ``` 1.1.0a0+2addccc cpu, sorted_input=False: 340 µs ± 5.88 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each) 717 µs ± 14.9 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each) 52.3 ms ± 2.75 ms per loop (mean ± std. dev. of 7 runs, 10 loops each) 52.3 ms ± 1.79 ms per loop (mean ± std. dev. of 7 runs, 10 loops each) cpu, sorted_input=True: 32.8 µs ± 285 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each) 49.9 µs ± 557 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each) 51.6 µs ± 1.08 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each) 78 µs ± 782 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each) cuda, sorted_input=False: 213 µs ± 1.52 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each) 291 µs ± 3.81 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each) 250 µs ± 1.05 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each) 321 µs ± 1.59 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each) cuda, sorted_input=True: 45.6 µs ± 2.13 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each) 110 µs ± 2.47 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each) 82 µs ± 857 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each) 143 µs ± 409 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each) ``` ```python print(torch.__version__) print() a1, a2 = tensor.unbind(1) indices = (a1 * tensor.max() + a2).sort().indices a = tensor.index_select(0, indices).to('cpu') print('cpu, sorted_input=False:') %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0) %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, return_inverse=True) %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, return_counts=True) %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, return_inverse=True, return_counts=True) print() print('cpu, sorted_input=True:') %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted_input=True) %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted_input=True, return_inverse=True) %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted_input=True, return_counts=True) %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted_input=True, return_inverse=True, return_counts=True) print() a = a.to('cuda') print('cuda, sorted_input=False:') %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0); torch.cuda.synchronize() %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, return_inverse=True); torch.cuda.synchronize() %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, return_counts=True); torch.cuda.synchronize() %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, return_inverse=True, return_counts=True); torch.cuda.synchronize() print() print('cuda, sorted_input=True:') %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted_input=True); torch.cuda.synchronize() %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted_input=True, return_inverse=True); torch.cuda.synchronize() %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted_input=True, return_counts=True); torch.cuda.synchronize() %timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted_input=True, return_inverse=True, return_counts=True); torch.cuda.synchronize() ``` ``` cpu, sorted_input=False: 55.4 ms ± 1.12 ms per loop (mean ± std. dev. of 7 runs, 10 loops each) 55.8 ms ± 616 µs per loop (mean ± std. dev. of 7 runs, 10 loops each) 55.2 ms ± 402 µs per loop (mean ± std. dev. of 7 runs, 10 loops each) 55.1 ms ± 725 µs per loop (mean ± std. dev. of 7 runs, 10 loops each) cpu, sorted_input=True: 54.7 ms ± 585 µs per loop (mean ± std. dev. of 7 runs, 10 loops each) 55.2 ms ± 1.23 ms per loop (mean ± std. dev. of 7 runs, 10 loops each) 54.5 ms ± 865 µs per loop (mean ± std. dev. of 7 runs, 10 loops each) 54.9 ms ± 577 µs per loop (mean ± std. dev. of 7 runs, 10 loops each) cuda, sorted_input=False: 171 µs ± 783 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each) 220 µs ± 1.65 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each) 203 µs ± 2.95 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each) 251 µs ± 2.83 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each) cuda, sorted_input=True: 59.6 µs ± 757 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each) 113 µs ± 431 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each) 93.2 µs ± 2.13 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each) 147 µs ± 2.81 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each) ``` The CPU implementation of `unique_dim` is super slow, see pytorch#18987, but this PR will not worry about this issue. Pull Request resolved: pytorch#19060 Differential Revision: D14866909 Pulled By: ezyang fbshipit-source-id: d20012cec68c37b05cf770a6f4d6524f910b950f

Add sorted_input to unique

6eb25f6

VitalyFedyunin reviewed Apr 9, 2019

View reviewed changes

aten/src/ATen/native/Unique.cpp Show resolved Hide resolved

aten/src/ATen/native/Unique.cpp Outdated Show resolved Hide resolved

aten/src/ATen/native/Unique.cpp Outdated Show resolved Hide resolved

ezyang added the module: operators label Apr 9, 2019

zasdfgbnm changed the title ~~Add sorted_input to unique~~ Add torch.unique_consecutive Apr 9, 2019

zasdfgbnm added 5 commits April 9, 2019 17:55

refactor to name it as unique_consecutive

f512a10

fix docs

3a742eb

optim

91d14f1

typo

0af8541

method and docs

4c60537

ezyang approved these changes Apr 9, 2019

View reviewed changes

fix type hint

eca4368

pytorchbot added the merge-this-please Was marked for merge with @pytorchbot merge this please label Apr 10, 2019

facebook-github-bot reviewed Apr 10, 2019

View reviewed changes

facebook-github-bot closed this in ea2405c Apr 10, 2019

facebook-github-bot added the merged label Apr 10, 2019

ezyang added the open source label Jun 24, 2019

zasdfgbnm deleted the unique-sorted_input branch November 5, 2019 01:00

1112zakaria mentioned this pull request Nov 5, 2025

[PT FE] Added support for aten::unique_consecutive openvinotoolkit/openvino#32688

Draft

Add torch.unique_consecutive #19060

Add torch.unique_consecutive #19060

Uh oh!

Conversation

zasdfgbnm commented Apr 9, 2019

Uh oh!

zasdfgbnm commented Apr 9, 2019

Uh oh!

zasdfgbnm commented Apr 9, 2019

Uh oh!

zasdfgbnm commented Apr 9, 2019

Uh oh!

ngimel commented Apr 9, 2019

Uh oh!

zasdfgbnm commented Apr 9, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ngimel commented Apr 9, 2019

Uh oh!

zasdfgbnm commented Apr 9, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

VitalyFedyunin left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

zasdfgbnm commented Apr 9, 2019

Uh oh!

ezyang commented Apr 9, 2019

Uh oh!

zasdfgbnm commented Apr 9, 2019

Uh oh!

zasdfgbnm commented Apr 9, 2019

Uh oh!

ezyang commented Apr 9, 2019

Uh oh!

zasdfgbnm commented Apr 9, 2019

Uh oh!

ezyang commented Apr 9, 2019

Uh oh!

zasdfgbnm commented Apr 10, 2019

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

facebook-github-bot commented Apr 10, 2019

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

zasdfgbnm commented Apr 9, 2019 •

edited

Loading

zasdfgbnm commented Apr 9, 2019 •

edited

Loading