[Bug-Report|缺陷反馈]: TensorFlow Mean output shape loses dimensions after a zero-sized dimension

Problem

TensorFlow Mean output descriptor translation drops dimensions after a zero-sized dimension. For a serialized output shape [0, 3], TensorFlowUtil::TransTensorDescriptor returns [0]. A normal [2, 3] shape is preserved.

The corruption is caused by a special case for Mean that replaces the descriptor shape with a one-dimensional shape when it encounters any zero dimension, then stops scanning dimensions. The source comment says the Mean and FusedBatchNormGrad workarounds address inaccurate TensorFlow inference for ResNet50; this case does not verify a full ResNet50 model.

Reproduction

The parser unit test UtestTensorflowParser.tensorflow_MeanPreservesEmptyOutputShape_test passes serialized TensorFlow output descriptors for [0, 3] and [2, 3] through TransTensorDescriptor and compares the resulting dimensions.

At source commit abaf3791a9ed60b5f871baf5eaade7b58fbbecd9, the empty-shape case fails with actual dimensions [0]; the ordinary shape control passes. After removing the Mean rewrite, both cases pass.

Fix

Remove the Mean zero-dimension shape rewrite. Keep the FusedBatchNormGrad adjustment unchanged. The change updates descriptor translation, adds a parser regression test, and documents the serialized-shape behavior.

Validation

The parser regression test reproduces the shape collapse before the change and passes after it. The [0, 3] empty shape and [2, 3] control are both preserved. The ut_parser target builds successfully. The target repository's applicable pre-commit hooks, clang-format, and git diff --check pass.

These test and build results were recorded against the original source commit and have not been rerun locally against the current develop base. GitCode Actions reports ci-pipeline-passed for the current head of PR !5216. The historical ResNet50 workaround was not validated with a full model, so compatibility with that scenario remains unverified.

关联 PR: https://gitcode.com/cann/ge/merge_requests/5216