`tf.raw_ops.AvgPool`: negative kernel size is not checked at shape inference step

Issue type

Bug

Have you reproduced the bug with TensorFlow Nightly?

Yes

Source

source

TensorFlow version

tf 2.17

Custom code

Yes

OS platform and distribution

Linux Ubuntu 22.04 LTS

Mobile device

No response

Python version

3.11.7

Bazel version

6.5.0

GCC/compiler version

clang 16

CUDA/cuDNN version

No response

GPU model and memory

No response

Current behavior?

Currently shape inference step of tf.raw_ops.AvgPool allows negative kernel size.
Note that debug build rejects it here:

  TF_RETURN_IF_ERROR(GetWindowedOutputSizeFromDims(
      c, in_rows_dim, kernel_rows, stride_rows, padding, &output_rows));

, where kernel_rows is converted to DimensionOrConstant and ends up with assertion failure here:

inline DimensionOrConstant::DimensionOrConstant(int64_t val) : val(val) {
  DCHECK(val >= 0 || val == InferenceContext::kUnknownDim)
      << "Dimension must be non-negative or equal to "
         "InferenceContext::kUnknownDim but got "
      << val;
}

Standalone code to reproduce the issue

import tensorflow as tf

tf.compat.v1.disable_eager_execution()

x = tf.raw_ops.AvgPool(
    value=tf.random.normal([1,1,1,1]),
    ksize=[1,-2,1,1],
    strides=[1,1,1,1],
    padding="SAME",
    data_format='NHWC',
    name=None
)

print(x)

Relevant log output

Release Build:

Tensor("AvgPool:0", shape=(1, 1, 1, 1), dtype=float32)

Debug Build:

2024-02-23 22:18:43.783609: F ./tensorflow/core/framework/shape_inference.h:891] Check failed: val >= 0 || val == InferenceContext::kUnknownDim Dimension must be non-negative or equal to InferenceContext::kUnknownDim but got -2
Aborted (core dumped)