Fix some checks not ideal to onnx-ml #781
Conversation
1. Need to consider cases where no default ONNX operator set exists 2. Cannot assume that value_info's are all tensors (CoreML's neural network classifiers contain ZipMap, the they may produce Maps)
| enforce_has_field(type, elem_type); | ||
| enforce_has_field(type, shape); | ||
| } break; | ||
| default: |
There was a problem hiding this comment.
we should add #ifdef ONNX_ML and add cases of map/sequence in-place.
onnx/checker.cc
Outdated
| } | ||
| auto dit = opset_imports.find(ONNX_DOMAIN); | ||
| if (dit == opset_imports.end()) { | ||
| if (opset_imports.size() < 1) { |
|
✅ Build onnx 0.3.2387 completed (commit 4fec2d4b7b by @wschin) |
onnx/checker.cc
Outdated
| } | ||
| auto dit = opset_imports.find(ONNX_DOMAIN); | ||
| if (dit == opset_imports.end()) { | ||
| if (opset_imports.empty()) { |
There was a problem hiding this comment.
This doesn't seem right to me. Do you want to avoid the check when you have other domain?
There was a problem hiding this comment.
The following checks are not necessary if at least one domain exists because (1) opset_import is good with IR>=3 (2) default opset_imports should be set only if opset_import is empty.
There was a problem hiding this comment.
I think the right check should be:
if (model.ir_version() >= 3) {
// enforce opset_imports is not empty
} else {
// enforce opset_imports is empty,
// and set the default value for ONNX_DOMAIN
}
| } else { | ||
| if (opset_imports.empty()) | ||
| opset_imports[ONNX_DOMAIN] = 1; | ||
| } |
There was a problem hiding this comment.
else {
fail_check(
"model with IR version < 3 should have no opset_import specified for ONNX");
}
There was a problem hiding this comment.
It breaks a lot of torch tests.
|
@houseroad , looks like some of the tests are not following the fact that opset_import should not be there if ir_version <= 2. Could you help on it please? |
|
❌ Build onnx 0.3.2402 failed (commit bfc1669fef by @wschin) |
|
❌ Build onnx 0.3.2405 failed (commit a101afba21 by @wschin) |
|
#788 should solve the problem |
|
❌ Build onnx 0.3.2418 failed (commit 897f159ca2 by @linkerzhang) |
|
❌ Build onnx 0.3.2427 failed (commit f95c64fec7 by @houseroad) |
|
❌ Build onnx 0.3.2429 failed (commit 1f2d3ac5f6 by @bddppq) |
|
❌ Build onnx 0.3.2433 failed (commit af7fef5b47 by @bddppq) |
|
@houseroad looks like the tests still fail :(. |
|
@linkerzhang Some test cases were not updated. Checking the reason |
|
❌ Build onnx 0.3.2437 failed (commit 9ad9ab73ba by @bddppq) |
|
❌ Build onnx 0.3.2441 failed (commit c826f80c41 by @linkerzhang) |
|
❌ Build onnx 0.3.2444 failed (commit ba96ba9bab by @linkerzhang) |
* Fix some checks not ideal to onnx-ml 1. Need to consider cases where no default ONNX operator set exists 2. Cannot assume that value_info's are all tensors (CoreML's neural network classifiers contain ZipMap, the they may produce Maps) * Address comments and improve some more error messages * Make SSA not abbreviated * Address a comment * a fix (cherry picked from commit e3da0f9)
* Cherry picked fix on python Schema API * Merge Rel 1.1.0 branch into master (#657) * Increase Version number for release 1.1.0 * Increase version number to 1.1.1 (cherry picked from commit ecac1c1) * Fix some checks not ideal to onnx-ml (#781) * Fix some checks not ideal to onnx-ml 1. Need to consider cases where no default ONNX operator set exists 2. Cannot assume that value_info's are all tensors (CoreML's neural network classifiers contain ZipMap, the they may produce Maps) * Address comments and improve some more error messages * Make SSA not abbreviated * Address a comment * a fix (cherry picked from commit e3da0f9) * Cherry picked updated testcases * Cherry pick updates on testcases (1) * Remove incorrect cases (#791) * Update the remainig cases (#794) * Fix incorrect conflict merge * Remove unpassed testcases * Bump up version number
* Cherry picked fix on python Schema API * Merge Rel 1.1.0 branch into master (#657) * Increase Version number for release 1.1.0 * Increase version number to 1.1.1 (cherry picked from commit ecac1c1) * Fix some checks not ideal to onnx-ml (#781) * Fix some checks not ideal to onnx-ml 1. Need to consider cases where no default ONNX operator set exists 2. Cannot assume that value_info's are all tensors (CoreML's neural network classifiers contain ZipMap, the they may produce Maps) * Address comments and improve some more error messages * Make SSA not abbreviated * Address a comment * a fix (cherry picked from commit e3da0f9) * Cherry picked updated testcases * Cherry pick updates on testcases (1) * Remove incorrect cases (#791) * Update the remainig cases (#794) * Fix incorrect conflict merge * Remove unpassed testcases * Bump up version number * Kezhan/update ml op spec (#799) * update normalizer. * update ml op spec. * update traditional ops. * update ml op doc (cherry picked from commit bf00ae6)
* Fix some checks not ideal to onnx-ml 1. Need to consider cases where no default ONNX operator set exists 2. Cannot assume that value_info's are all tensors (CoreML's neural network classifiers contain ZipMap, the they may produce Maps) * Address comments and improve some more error messages * Make SSA not abbreviated * Address a comment * a fix
* Fix some checks not ideal to onnx-ml 1. Need to consider cases where no default ONNX operator set exists 2. Cannot assume that value_info's are all tensors (CoreML's neural network classifiers contain ZipMap, the they may produce Maps) * Address comments and improve some more error messages * Make SSA not abbreviated * Address a comment * a fix
network classifiers contain ZipMap, so they may produce Maps)