Added support for aten::randperm and aten::polar #29585

vijaykr338 · 2025-03-19T17:17:46Z

Details:

added support for aten::randperm and tests

currently working on aten::polar and others

Tickets:

[Good First Issue]: support aten::polar and aten::randperm #29547

vijaykr338

@mvafin
I found out that pytorch.polar returns complex tensors meanwhile OpenVINO does not support them right now

to make it work, I separately computed real and imag parts using cos and sine then wrapped it with a ComplexTypeMark.

I will now work on str and delete

src/frontends/pytorch/src/op/polar.cpp

tests/layer_tests/pytorch_tests/test_randperm.py

tests/layer_tests/pytorch_tests/test_polar.py

mvafin · 2025-03-24T12:36:45Z

Also please fix code style

src/frontends/pytorch/src/op/polar.cpp:1:-#include "openvino/op/cos.hpp"
src/frontends/pytorch/src/op/polar.cpp:2:-#include "openvino/op/sin.hpp"
src/frontends/pytorch/src/op/polar.cpp:3:-#include "openvino/op/multiply.hpp"
src/frontends/pytorch/src/op/polar.cpp:4:-#include "openvino/op/concat.hpp"
src/frontends/pytorch/src/op/polar.cpp:2:+#include "openvino/op/concat.hpp"
src/frontends/pytorch/src/op/polar.cpp:4:+#include "openvino/op/cos.hpp"
src/frontends/pytorch/src/op/polar.cpp:5:+#include "openvino/op/multiply.hpp"
src/frontends/pytorch/src/op/polar.cpp:6:+#include "openvino/op/sin.hpp"
src/frontends/pytorch/src/op/polar.cpp:20:-    auto real = context.mark_node(std::make_shared<v1::Multiply>(abs,context.mark_node(std::make_shared<v0::Cos>(angle))));
src/frontends/pytorch/src/op/polar.cpp:21:-    auto imag = context.mark_node(std::make_shared<v1::Multiply>(abs,context.mark_node(std::make_shared<v0::Sin>(angle))));
src/frontends/pytorch/src/op/polar.cpp:20:+    auto real =
src/frontends/pytorch/src/op/polar.cpp:21:+        context.mark_node(std::make_shared<v1::Multiply>(abs, context.mark_node(std::make_shared<v0::Cos>(angle))));
src/frontends/pytorch/src/op/polar.cpp:22:+    auto imag =
src/frontends/pytorch/src/op/polar.cpp:23:+        context.mark_node(std::make_shared<v1::Multiply>(abs, context.mark_node(std::make_shared<v0::Sin>(angle))));
src/frontends/pytorch/src/op/randperm.cpp:1:-#include "openvino/op/topk.hpp"
src/frontends/pytorch/src/op/randperm.cpp:2:-#include "openvino/op/random_uniform.hpp"
src/frontends/pytorch/src/op/randperm.cpp:5:-#include "utils.hpp"
src/frontends/pytorch/src/op/randperm.cpp:3:+#include "openvino/op/random_uniform.hpp"
src/frontends/pytorch/src/op/randperm.cpp:5:+#include "openvino/op/topk.hpp"
src/frontends/pytorch/src/op/randperm.cpp:6:+#include "utils.hpp"
src/frontends/pytorch/src/op/randperm.cpp:23:-            OPENVINO_ASSERT(dtype_value == 4, "Only dtype value 4 (int64) is supported for aten::randperm, got: ", dtype_value);
src/frontends/pytorch/src/op/randperm.cpp:23:+            OPENVINO_ASSERT(dtype_value == 4,
src/frontends/pytorch/src/op/randperm.cpp:24:+                            "Only dtype value 4 (int64) is supported for aten::randperm, got: ",
src/frontends/pytorch/src/op/randperm.cpp:25:+                            dtype_value);
src/frontends/pytorch/src/op/randperm.cpp:28:-            OPENVINO_ASSERT(dtype_value == 4, "Only dtype value 4 (int64) is supported for aten::randperm, got: ", dtype_value);
src/frontends/pytorch/src/op/randperm.cpp:30:+            OPENVINO_ASSERT(dtype_value == 4,
src/frontends/pytorch/src/op/randperm.cpp:31:+                            "Only dtype value 4 (int64) is supported for aten::randperm, got: ",
src/frontends/pytorch/src/op/randperm.cpp:32:+                            dtype_value);
src/frontends/pytorch/src/op/randperm.cpp:31:-            PYTORCH_OP_CONVERSION_CHECK(false, "Unexpected number of inputs for aten::randperm: ", num_inputs);
src/frontends/pytorch/src/op/randperm.cpp:35:+        PYTORCH_OP_CONVERSION_CHECK(false, "Unexpected number of inputs for aten::randperm: ", num_inputs);
src/frontends/pytorch/src/op/randperm.cpp:34:-    return {context.mark_node(v0::Constant::create(element::i64, Shape{0},std::vector<int64_t>{}))};}
src/frontends/pytorch/src/op/randperm.cpp:38:+        return {context.mark_node(v0::Constant::create(element::i64, Shape{0}, std::vector<int64_t>{}))};
src/frontends/pytorch/src/op/randperm.cpp:39:+    }
src/frontends/pytorch/src/op/randperm.cpp:41:-    auto topk = context.mark_node(std::make_shared<v11::TopK>(random_tensor, k, axis, ov::op::TopKMode::MIN, ov::op::TopKSortType::SORT_VALUES, element::i64, false));
src/frontends/pytorch/src/op/randperm.cpp:46:+    auto topk = context.mark_node(std::make_shared<v11::TopK>(random_tensor,
src/frontends/pytorch/src/op/randperm.cpp:47:+                                                              k,
src/frontends/pytorch/src/op/randperm.cpp:48:+                                                              axis,
src/frontends/pytorch/src/op/randperm.cpp:49:+                                                              ov::op::TopKMode::MIN,
src/frontends/pytorch/src/op/randperm.cpp:50:+                                                              ov::op::TopKSortType::SORT_VALUES,
src/frontends/pytorch/src/op/randperm.cpp:51:+                                                              element::i64,
src/frontends/pytorch/src/op/randperm.cpp:52:+                                                              false));

vijaykr338 · 2025-03-26T12:22:11Z

additionally, for randperm produces random permutation, we can't expect the output of OpenVINO to match Pytorch's output, so instead we just verify if the output has correct shape and when sorted gives 0 to n-1

That approach is fine. A better approach would be to use randperm to randomize the tensor as this might be the useful usage of this function and since the output will not be the same between openvino and torch, sort the tensor and compare. That actually can be done inside a torch model, so regular _test function can be used. Could you implement such approach?

thank you for the insight, I updated the model to use randperm and sort it within the model itself to use the test function

I also separated the sin and cos for easier reading

tests/layer_tests/pytorch_tests/test_randperm.py

src/frontends/pytorch/src/op/polar.cpp

src/frontends/pytorch/src/op/randperm.cpp

mvafin · 2025-03-26T13:58:55Z

tests/layer_tests/pytorch_tests/test_randperm.py

+                                         device=x.device, pin_memory=False)
+                else:
+                    raise ValueError("Invalid num_inputs")
+                sorted_p, _ = torch.sort(p)


My idea was to use randperm output to permute the actual tensor. BTW, x can be such tensor, otherwise it is unused variable here, and sort the tensor after permutation to verify that permutation works

alright, I updated it so that randperm's o/p shuffles the actual tensor, and then sorts it

please check

src/frontends/pytorch/src/op/randperm.cpp

Co-authored-by: Maxim Vafin <maxim.vafin@intel.com>

now it can handle value of n dynamically

vijaykr338 · 2025-03-27T08:35:38Z

@mvafin
should I make any further changes?

mvafin · 2025-03-27T10:46:03Z

tests/layer_tests/pytorch_tests/test_polar.py

+        atol = 1e-4 if precision == "FP32" else 1e-3
+        rtol = 1e-4


Suggested change

atol = 1e-4 if precision == "FP32" else 1e-3

rtol = 1e-4

you are not using atol and rtol here

ah yes, I have removed them now, I was initially using atol and rtol as I was worried about the floating point differences that might be produced, but the default test layer handles it fine!

tests/layer_tests/pytorch_tests/test_randperm.py

added a comment to explain why we sort

mvafin · 2025-03-28T12:12:10Z

build_jenkins

mvafin

Thanks for your contribution. Everything seems to be okay now.

vijaykr338 · 2025-03-28T12:25:22Z

@mvafin Thank you very much for your patience and feedback throughout the process

I appreciate the valuable insights gained and will apply them to my future PRs

vijaykr338 · 2025-03-31T09:31:26Z

Hi @mvafin, are there any issues with the merging?

mvafin · 2025-03-31T09:47:22Z

Hi @mvafin, are there any issues with the merging?

Don't worry about this. That is cause by the issues in our CI, I will look into it.

vijaykr338 · 2025-03-31T10:15:41Z

Hi @mvafin, are there any issues with the merging?

Don't worry about this. That is cause by the issues in our CI, I will look into it.

ah, alright, thanks for the update.

mvafin · 2025-03-31T11:07:09Z

build_jenkins

vijaykr338 and others added 4 commits March 15, 2025 10:24

added aten::take function in pytorch, but tests are not running

0656890

implemented aten::str, aten::delete but unable to write their tests

98c1c61

Merge branch 'master' into master

725c474

added support for aten::randperm with tests

4176919

github-actions bot added the category: PyTorch FE OpenVINO PyTorch Frontend label Mar 19, 2025

vijaykr338 added 5 commits March 19, 2025 22:50

Delete src/frontends/pytorch/src/op/delete.cpp

e32628f

cleaned

1483255

Delete tests/layer_tests/pytorch_tests/test_take.py

8a16331

Delete src/frontends/pytorch/src/op/take.cpp

e8b35e8

Delete src/frontends/pytorch/src/op/str.cpp

16f11fa

sys-openvino-ci added ExternalPR External contributor labels Mar 19, 2025

This was referenced Mar 22, 2025

[Good First Issue]: Support request for aten::pop and aten::reverse #29630

Open

[Good First Issue]: Support aten::smooth_l1_loss #29632

Open

vijaykr338 added 2 commits March 23, 2025 12:38

added support for aten::polar

cd18e01

fixed code style for aten::polar and aten::randperm

113d07e

vijaykr338 commented Mar 23, 2025

View reviewed changes

Merge branch 'master' into work2

c6ec86e

vijaykr338 marked this pull request as ready for review March 23, 2025 12:52

vijaykr338 requested a review from a team as a code owner March 23, 2025 12:52

vijaykr338 requested review from PiotrKrzem and cavusmustafa March 23, 2025 12:52

wang-zhuoran mentioned this pull request Mar 24, 2025

[Good First Issue]: Support aten::cross_entropy_loss #29647

Open

Merge branch 'master' into work2

bf75e7b

vijaykr338 changed the title ~~Added support for aten::randperm with tests [WIP - aten::polar]~~ Added support for aten::randperm and aten::polar Mar 24, 2025

vijaykr338 mentioned this pull request Mar 24, 2025

[Good First Issue]: Operation list to be supported #28584

Open

mvafin requested changes Mar 24, 2025

View reviewed changes

src/frontends/pytorch/src/op/polar.cpp Outdated Show resolved Hide resolved

tests/layer_tests/pytorch_tests/test_randperm.py Outdated Show resolved Hide resolved

mvafin reviewed Mar 24, 2025

View reviewed changes

tests/layer_tests/pytorch_tests/test_polar.py Outdated Show resolved Hide resolved

Merge branch 'master' into work2

c8dac15

vijaykr338 added 2 commits March 26, 2025 17:45

Update polar.cpp

db261fc

Update test_randperm.py

bbb6952

mvafin requested changes Mar 26, 2025

View reviewed changes

mvafin reviewed Mar 26, 2025

View reviewed changes

src/frontends/pytorch/src/op/randperm.cpp Outdated Show resolved Hide resolved

mvafin reviewed Mar 26, 2025

View reviewed changes

src/frontends/pytorch/src/op/randperm.cpp Outdated Show resolved Hide resolved

vijaykr338 and others added 6 commits March 26, 2025 19:52

Update tests/layer_tests/pytorch_tests/test_randperm.py

9cc31cd

Co-authored-by: Maxim Vafin <maxim.vafin@intel.com>

Update polar.cpp

7c9e745

Update randperm.cpp

7b4b3b7

now it can handle value of n dynamically

Update test_randperm.py

e8f42fa

Update test_randperm.py, added the mark

b27a3f2

Update test_randperm.py

f041050

mvafin reviewed Mar 27, 2025

View reviewed changes

tests/layer_tests/pytorch_tests/test_randperm.py Show resolved Hide resolved

vijaykr338 added 2 commits March 28, 2025 12:35

Update test_randperm.py

e7e3999

added a comment to explain why we sort

Update test_polar.py

e4d204d

mvafin approved these changes Mar 28, 2025

View reviewed changes

Merge branch 'master' into work2

0102220

mvafin added this pull request to the merge queue Apr 1, 2025

github-merge-queue bot removed this pull request from the merge queue due to failed status checks Apr 1, 2025

mvafin added this pull request to the merge queue Apr 1, 2025

Merged via the queue into openvinotoolkit:master with commit f2182b3 Apr 1, 2025
185 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Added support for aten::randperm and aten::polar #29585

Added support for aten::randperm and aten::polar #29585

vijaykr338 commented Mar 19, 2025

vijaykr338 left a comment •

edited

Loading

mvafin commented Mar 24, 2025

vijaykr338 commented Mar 26, 2025

mvafin Mar 26, 2025

vijaykr338 Mar 26, 2025

vijaykr338 commented Mar 27, 2025

mvafin Mar 27, 2025

vijaykr338 Mar 28, 2025 •

edited

Loading

mvafin commented Mar 28, 2025

mvafin left a comment •

edited

Loading

vijaykr338 commented Mar 28, 2025

vijaykr338 commented Mar 31, 2025

mvafin commented Mar 31, 2025

vijaykr338 commented Mar 31, 2025

mvafin commented Mar 31, 2025

Added support for aten::randperm and aten::polar #29585

Added support for aten::randperm and aten::polar #29585

Conversation

vijaykr338 commented Mar 19, 2025

Details:

Tickets:

vijaykr338 left a comment • edited Loading

Choose a reason for hiding this comment

mvafin commented Mar 24, 2025

vijaykr338 commented Mar 26, 2025

mvafin Mar 26, 2025

Choose a reason for hiding this comment

vijaykr338 Mar 26, 2025

Choose a reason for hiding this comment

vijaykr338 commented Mar 27, 2025

mvafin Mar 27, 2025

Choose a reason for hiding this comment

vijaykr338 Mar 28, 2025 • edited Loading

Choose a reason for hiding this comment

mvafin commented Mar 28, 2025

mvafin left a comment • edited Loading

Choose a reason for hiding this comment

vijaykr338 commented Mar 28, 2025

vijaykr338 commented Mar 31, 2025

mvafin commented Mar 31, 2025

vijaykr338 commented Mar 31, 2025

mvafin commented Mar 31, 2025

vijaykr338 left a comment •

edited

Loading

vijaykr338 Mar 28, 2025 •

edited

Loading

mvafin left a comment •

edited

Loading