Add Op (_upsample_bilinear2d_aa, _upsample_bicubic2d_aa) | feat(torchlib) #1259

xiaowuhu · 2024-01-26T11:07:21Z

It seems that the antialias method is different between ONNX and PyTorch, so we can just compare the shape, instead of the value.

Below is the difference between ONXN and PyTorch:

# ONNX
import numpy as np
self = np.array([[[[2,1,1,1],
                   [1,1,1,1],
                   [1,1,1,1],
                   [1,1,1,1]]]]).astype(np.float32)
print(self.shape)
output_size = np.array([1,1]).astype(np.int64)
align_corners = True
r = aten__upsample_bicubic2d_aa(self, output_size, align_corners)
print(r)

ONXN output = [[[[1.390625]]]]

# PyTorch
import torch as t
r = t.ops.aten._upsample_bicubic2d_aa(t.tensor(self), t.tensor(output_size), align_corners)
print(r)

Torch output = tensor([[[[2.2656]]]])

I also tried some other parameters combination but none of them can match with torch.

codecov · 2024-01-26T11:13:31Z

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (ce3eb4a) 78.68% compared to head (778b799) 78.85%.
Report is 8 commits behind head on main.

❗ Current head 778b799 differs from pull request most recent head cf4f4af. Consider uploading reports for the commit cf4f4af to get more accurate results

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1259      +/-   ##
==========================================
+ Coverage   78.68%   78.85%   +0.17%     
==========================================
  Files         119      119              
  Lines       15762    15700      -62     
  Branches     2486     2481       -5     
==========================================
- Hits        12403    12381      -22     
+ Misses       2950     2911      -39     
+ Partials      409      408       -1

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

github-actions · 2024-01-26T11:21:05Z

Test Results

24 files ± 0 24 suites ±0 1h 41m 53s ⏱️ + 11m 9s
11 405 tests + 6 8 439 ✅ + 4 2 952 💤 ± 0 14 ❌ +2
274 768 runs +16 882 63 102 ✅ +4 302 211 460 💤 +12 578 206 ❌ +2

For more details on these failures, see this check.

Results for commit cf4f4af. ± Comparison against base commit 457e52e.

This pull request removes 29 and adds 35 tests. Note that renamed tests count towards both.

onnxscript.tests.function_libs.torch_lib.ops_test.TestFunctionValidity ‑ test_function_has_op_schema_327_aten_upsample_bilinear2d_vec
onnxscript.tests.function_libs.torch_lib.ops_test.TestFunctionValidity ‑ test_function_has_op_schema_328_aten_upsample_bicubic2d
onnxscript.tests.function_libs.torch_lib.ops_test.TestFunctionValidity ‑ test_function_has_op_schema_329_aten_upsample_bicubic2d_vec
onnxscript.tests.function_libs.torch_lib.ops_test.TestFunctionValidity ‑ test_function_has_op_schema_330_aten_upsample_linear1d
onnxscript.tests.function_libs.torch_lib.ops_test.TestFunctionValidity ‑ test_function_has_op_schema_331_aten_upsample_nearest1d
onnxscript.tests.function_libs.torch_lib.ops_test.TestFunctionValidity ‑ test_function_has_op_schema_332_aten_upsample_nearest2d
onnxscript.tests.function_libs.torch_lib.ops_test.TestFunctionValidity ‑ test_function_has_op_schema_333_aten_upsample_nearest3d
onnxscript.tests.function_libs.torch_lib.ops_test.TestFunctionValidity ‑ test_function_has_op_schema_334_aten_upsample_trilinear3d
onnxscript.tests.function_libs.torch_lib.ops_test.TestFunctionValidity ‑ test_function_has_op_schema_335_aten_ones_like
onnxscript.tests.function_libs.torch_lib.ops_test.TestFunctionValidity ‑ test_function_has_op_schema_336_aten_roll
…

onnxscript.tests.function_libs.torch_lib.ops_test.TestFunctionValidity ‑ test_function_has_op_schema_327_aten__upsample_bilinear2d_aa
onnxscript.tests.function_libs.torch_lib.ops_test.TestFunctionValidity ‑ test_function_has_op_schema_328_aten_upsample_bilinear2d_vec
onnxscript.tests.function_libs.torch_lib.ops_test.TestFunctionValidity ‑ test_function_has_op_schema_329_aten_upsample_bicubic2d
onnxscript.tests.function_libs.torch_lib.ops_test.TestFunctionValidity ‑ test_function_has_op_schema_330_aten_upsample_bicubic2d_vec
onnxscript.tests.function_libs.torch_lib.ops_test.TestFunctionValidity ‑ test_function_has_op_schema_331_aten__upsample_bicubic2d_aa
onnxscript.tests.function_libs.torch_lib.ops_test.TestFunctionValidity ‑ test_function_has_op_schema_332_aten_upsample_linear1d
onnxscript.tests.function_libs.torch_lib.ops_test.TestFunctionValidity ‑ test_function_has_op_schema_333_aten_upsample_nearest1d
onnxscript.tests.function_libs.torch_lib.ops_test.TestFunctionValidity ‑ test_function_has_op_schema_334_aten_upsample_nearest2d
onnxscript.tests.function_libs.torch_lib.ops_test.TestFunctionValidity ‑ test_function_has_op_schema_335_aten_upsample_nearest3d
onnxscript.tests.function_libs.torch_lib.ops_test.TestFunctionValidity ‑ test_function_has_op_schema_336_aten_upsample_trilinear3d
…

♻️ This comment has been updated with latest results.

justinchuby · 2024-01-27T14:19:30Z

It would be better if we match values because the values should be deterministic. Do we know how PyTorch does it?

xiaowuhu · 2024-02-19T08:37:11Z

It would be better if we match values because the values should be deterministic. Do we know how PyTorch does it?

please see the description for this PR. I add comparison between onnx and torch.

justinchuby · 2024-02-20T10:43:20Z

Would it be helpful to consult the PyTorch implementation? I suspect we need additional processing to implement antialiasing.

justinchuby · 2024-03-05T01:59:00Z

From our discussion: understanding the PyTorch implementation proved to be harder than anticipated (https://github.com/pytorch/pytorch/blob/bcf35c6ae62bb6560befa3550e37a8283944e5f4/aten/src/ATen/native/cpu/UpSampleKernel.cpp#L2009). We will seek additional help for this.

…ilinear2d_aa functions (#2383) This PR implements the missing anti-aliasing (AA) variants of upsample functions that were requested in issue #1159: - `aten__upsample_bicubic2d_aa` - bicubic 2D upsampling with anti-aliasing - `aten__upsample_bilinear2d_aa` - bilinear 2D upsampling with anti-aliasing ## Changes Made ### Core Implementation - **Modified helper functions** to support anti-aliasing: - Added `antialias` parameter (default=0) to `_aten_upsample_output_size()` - Added `antialias` parameter (default=0) to `_aten_upsample_scales()` - Maintains backward compatibility with existing code - **Implemented AA functions** with same signatures as regular variants: ```python def aten__upsample_bicubic2d_aa(self, output_size, align_corners, scales_h=None, scales_w=None) def aten__upsample_bilinear2d_aa(self, output_size, align_corners, scales_h=None, scales_w=None) ``` Both functions pass `antialias=1` to enable ONNX Resize anti-aliasing. ### Test Configuration - **Added OpInfo entries** in `extra_opinfo.py` for both AA functions - **Added TorchLibOpInfo entries** in `ops_test_data.py` with `compare_shape_only_for_output=(0,)` since ONNX and PyTorch use different anti-aliasing algorithms ## Technical Details The AA variants use the same underlying logic as regular upsample functions but enable anti-aliasing in the ONNX Resize operation. As noted in the original issue discussion, ONNX and PyTorch implement different anti-aliasing methods, so tests compare shapes rather than exact values. Example usage: ```python import numpy as np from onnxscript.function_libs.torch_lib.ops.nn import aten__upsample_bicubic2d_aa # Create test input input_tensor = np.array([[[[2,1,1,1], [1,1,1,1], [1,1,1,1], [1,1,1,1]]]]).astype(np.float32) output_size = np.array([1,1]).astype(np.int64) # Use AA upsampling result = aten__upsample_bicubic2d_aa(input_tensor, output_size, align_corners=True) print(result) # Output: [[[[1.390625]]]] ``` ## Testing Results - ✅ All new AA function tests pass (2 passed, 1 skipped as expected for trace-only functions) - ✅ All existing upsample function tests continue to pass - no regressions - ✅ Functions produce expected different output when AA is enabled vs disabled - ✅ Helper functions work correctly with both `antialias=0` and `antialias=1` This implementation matches the approach from the previous PR #1259 and completes the upsample function suite requested in the issue. Fixes #1159. Fixes pytorch/pytorch#128818  --- 💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more [Copilot coding agent tips](https://gh.io/copilot-coding-agent-tips) in the docs. --------- Co-authored-by: copilot-swe-agent[bot] <[email protected]> Co-authored-by: justinchuby <[email protected]> Co-authored-by: titaiwangms <[email protected]> Co-authored-by: Justin Chu <[email protected]>

update

f814821

xiaowuhu changed the title ~~AddOp(_upsample_bilinear2d_aa, _upsample_bicubic2d_aa) | feat(TorchLib)~~ Add Op (_upsample_bilinear2d_aa, _upsample_bicubic2d_aa) | feat(torchlib) Jan 26, 2024

Update nn.py

778b799

xiaowuhu mentioned this pull request Jan 29, 2024

[torchlib] Implement various upsample functions #1159

Open

8 tasks

xiaowuhu added 3 commits February 19, 2024 15:26

Merge branch 'main' into xiaowu/AddOp(upsample_bicubic2d_aa)

f1cd040

Update nn.py

3855b1c

Merge branch 'main' into xiaowu/AddOp(upsample_bicubic2d_aa)

cf4f4af

Copilot AI mentioned this pull request Jun 14, 2025

[torchlib] Implement aten__upsample_bicubic2d_aa and aten__upsample_bilinear2d_aa functions #2383

Merged

justinchuby closed this Jun 14, 2025

github-project-automation bot moved this from Todo to Done in ONNX Script Review Board Jun 14, 2025

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Add Op (_upsample_bilinear2d_aa, _upsample_bicubic2d_aa) | feat(torchlib) #1259

Add Op (_upsample_bilinear2d_aa, _upsample_bicubic2d_aa) | feat(torchlib) #1259

Uh oh!

xiaowuhu commented Jan 26, 2024 •

edited

Loading

Uh oh!

codecov bot commented Jan 26, 2024 •

edited

Loading

Uh oh!

github-actions bot commented Jan 26, 2024 •

edited

Loading

Uh oh!

justinchuby commented Jan 27, 2024

Uh oh!

xiaowuhu commented Feb 19, 2024

Uh oh!

justinchuby commented Feb 20, 2024

Uh oh!

justinchuby commented Mar 5, 2024

Uh oh!

Uh oh!

Add Op (_upsample_bilinear2d_aa, _upsample_bicubic2d_aa) | feat(torchlib) #1259

Add Op (_upsample_bilinear2d_aa, _upsample_bicubic2d_aa) | feat(torchlib) #1259

Uh oh!

Conversation

xiaowuhu commented Jan 26, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

codecov bot commented Jan 26, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

github-actions bot commented Jan 26, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Test Results

Uh oh!

justinchuby commented Jan 27, 2024

Uh oh!

xiaowuhu commented Feb 19, 2024

Uh oh!

justinchuby commented Feb 20, 2024

Uh oh!

justinchuby commented Mar 5, 2024

Uh oh!

Uh oh!

xiaowuhu commented Jan 26, 2024 •

edited

Loading

codecov bot commented Jan 26, 2024 •

edited

Loading

github-actions bot commented Jan 26, 2024 •

edited

Loading