Model Zoo ResNet implementation fails with TensorFlow 2.5 custom ops

Description

I'm encountering compatibility issues with the ResNet implementation in Model Zoo when using TensorFlow 2.5.

Environment

• TensorFlow version: 2.5
• Models repository version: (latest as of report date)
• Operating system:
• Python version:

Error Details

When attempting to use the ResNet implementation from the Model Zoo with TensorFlow 2.5, the model fails to load due to incompatibility with custom operations.

Failure Mode

The model fails specifically when loading custom operations that were compiled against previous TensorFlow versions. The error occurs during model initialization before any training or inference can take place.

Affected Components

• Model Zoo ResNet implementation
• Custom operations in TensorFlow 2.5
• Model loading pipeline

Steps to Reproduce

1. Install TensorFlow 2.5
2. Clone the tensorflow/models repository
3. Attempt to load/use the ResNet implementation from Model Zoo
4. Observe the failure with custom ops

Expected Behavior

The ResNet implementation should load and function correctly with TensorFlow 2.5.

Potential Workarounds

• Downgrading to TensorFlow 2.4.x temporarily resolves the issue
• Modifying the custom op loading mechanism to be compatible with TF 2.5's API changes
• Using the Keras built-in ResNet implementation as an alternative

Any guidance or fixes to resolve this compatibility issue would be greatly appreciated.

Additional Context

This issue appears to be related to the changes in TensorFlow's C++ API for custom ops between version 2.4 and 2.5.

Start by reproducing the failure with TensorFlow 2.5 in the Model Zoo ResNet implementation, then compare custom-operation loading under TensorFlow 2.4. The issue does not name a source file or test; done means identifying and fixing the compatibility problem so the ResNet model loads and functions with TensorFlow 2.5.