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

[xiaowuhu]

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]

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]

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]

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

[xiaowuhu]

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]

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

[justinchuby]

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.