ConvTransposeWithDynamicPads msrc116345

[yuslepukhin]

This pull request strengthens input validation and error handling for the ConvTranspose operator (including dynamic pads) across both CPU and CUDA providers, and in shape inference logic. The main improvements ensure that invalid or inconsistent input shapes, attributes, and parameters are caught early with clear error messages, improving robustness and compliance with the ONNX specification.

Input shape and attribute validation:

• Added checks to ensure input tensor X and filter tensor W have at least 2 dimensions, and that the group attribute is positive in both the shape inference and kernel implementations. [1] [2] [3]
• Enhanced validation for the dynamic pads input: ensures it is present, is a 1D tensor, has the correct size, and contains only non-negative values. [1] [2]

Parameter validation (stride, dilation, output padding):

• Added checks to ensure that output_padding is less than either the stride or dilation for each spatial dimension, as required by the ONNX spec. [1] [2]
• Enforced that spatial dimension sizes, stride, kernel size, and dilation are all positive, and output padding is non-negative, in shape calculation routines.

Shape inference and calculation improvements:

• Improved error messages during shape inference when output shape values are inconsistent with input shape.
• Used SafeInt for safer arithmetic in output shape calculations to prevent overflow.

General code maintenance:

• Included necessary headers for SafeInt where used. [1] [2]

Testing:

• Added a basic functional test for 2D ConvTranspose with dynamic pads.

** References:
onnx/onnx#7997
#28524

[Copilot]

Pull request overview

This PR strengthens ConvTransposeWithDynamicPads validation across shape inference and CPU/CUDA execution paths, with added regression and functional tests for invalid shapes, dynamic pads, and output padding constraints.

Changes:

• Adds rank, dynamic pads, output padding, and arithmetic validation in CPU/CUDA ConvTranspose paths.
• Tightens ConvTransposeWithDynamicPads shape inference errors.
• Expands contrib op tests for dynamic pads and invalid inputs.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 6 comments.

File	Description
onnxruntime/core/graph/contrib_ops/contrib_defs.cc	Updates ConvTransposeWithDynamicPads shape inference validation and messages.
onnxruntime/core/providers/cpu/nn/conv_transpose_attributes.h	Adds runtime validation and SafeInt arithmetic for CPU/shared ConvTranspose attribute handling.
onnxruntime/core/providers/cuda/nn/conv_transpose.cc	Mirrors validation in CUDA ConvTranspose state setup.
onnxruntime/test/contrib_ops/conv_transpose_with_dynamic_pads_test.cc	Adds functional and failure tests for dynamic pads and ConvTranspose edge cases.

Comments suppressed due to low confidence (1)

onnxruntime/test/contrib_ops/conv_transpose_with_dynamic_pads_test.cc:139

• This test expects the shape-inference error text, but Pads is added as a runtime input because AddInput's initializer flag is not set. The kernel path reports "Pads input must be a 1D tensor...", so the expected substring with lowercase quoted 'pads' will not match. Make Pads an initializer here or update the expected error text.

  test.AddInput<int64_t>("Pads", {2, 2}, std::vector<int64_t>{0, 0, 0, 0});  // 2D instead of 1D
  test.AddOutput<float>("Y", {1, 1, 5, 5}, std::vector<float>(25, 0.0f));

  test.Run(OpTester::ExpectResult::kExpectFailure, "'pads' input must be a 1D");

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated 4 comments.

Comments suppressed due to low confidence (2)

onnxruntime/core/providers/cuda/nn/conv_transpose.cc:361

• This now treats Pads as required at runtime, but the ConvTransposeWithDynamicPads schema still declares input 2 as OpSchema::Optional (contrib_defs.cc:1730). Either the schema should be tightened to make Pads required, or the kernel should define behavior for an omitted Pads input; otherwise model validation can accept a graph that this kernel rejects as missing a required input.

      if (Pads == nullptr) {
        return ORT_MAKE_STATUS(ONNXRUNTIME, INVALID_ARGUMENT,
                               "Pads input is required for dynamic padding mode.");

onnxruntime/core/providers/cpu/nn/conv_transpose_attributes.h:276

• These new validations use ORT_ENFORCE inside a helper that is reached from PrepareForCompute; in ORT_NO_EXCEPTIONS builds ORT_ENFORCE aborts rather than returning a failure Status (see common.h:98-142). For user-controlled cases like negative output_padding, this can terminate the process instead of producing the INVALID_ARGUMENT status used by the other new checks. Please return/propagate Status from this helper or validate these conditions before calling it.

    ORT_ENFORCE(in_size > 0, "Input spatial dimension must be positive. Got: ", in_size);
    ORT_ENFORCE(stride > 0, "Stride must be positive. Got: ", stride);
    ORT_ENFORCE(kernel > 0, "Kernel size must be positive. Got: ", kernel);
    ORT_ENFORCE(dilation > 0, "Dilation must be positive. Got: ", dilation);
    ORT_ENFORCE(adj >= 0, "Output padding must be non-negative. Got: ", adj);

[yuslepukhin]

Merging to #28524 and closing.