Fix AttributeError in Eagle image processor for transformers 4.53+

[Rayen023]

What this does

Fixes a (🐛 Bug) that prevents GR00T model training with transformers == 4.53.x

The Eagle image processor in image_processing_eagle2_5_vl_fast.py calls _prepare_image_like_inputs() which doesn't exist in the BaseImageProcessorFast parent class in transformers 4.53.x , causing training to fail immediately with:

AttributeError: 'Eagle25VLImageProcessorFast' object has no attribute '_prepare_image_like_inputs'

Changes:

• Add runtime detection for _prepare_input_images vs _prepare_image_like_inputs
• Use hasattr() to check which method is available in the parent class
• Maintains backward/forward compatibility with different transformers versions

How it was tested

• ✅ Tested with transformers 4.53.3
• ✅ GR00T-N1.5-3B training completes training successfully (previously failed immediately)

How to checkout & try? (for the reviewer)

lerobot-train \
  --policy.type="groot" \
  --policy.base_model_path="nvidia/GR00T-N1.5-3B" \
  --dataset.repo_id="repo_id" \

[Copilot]

Pull Request Overview

This PR adds backward compatibility handling for the transformers library API changes in the Eagle25VLImageProcessorFast class. The code now dynamically checks which image preprocessing method is available (_prepare_image_like_inputs for transformers >= 4.53.0 or _prepare_input_images for older versions) and uses the appropriate one.

Key Changes

• Added runtime detection logic to check for available preprocessing methods using hasattr()
• Replaced direct calls to _prepare_image_like_inputs with the dynamically selected prepare_fn
• Added error handling when neither method is available

---

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

[Copilot]

The backward compatibility check for _prepare_input_images appears unnecessary since pyproject.toml specifies transformers>=4.53.0,<5.0.0 which only supports _prepare_image_like_inputs. Consider removing the fallback logic or updating the version constraint if older transformers versions need to be supported.

Suggested change

	# transformers >= 4.53.0: uses _prepare_image_like_inputs instead of _prepare_input_images
	# Check which method is available for compatibility
	if hasattr(self, '_prepare_image_like_inputs'):
	prepare_fn = self._prepare_image_like_inputs
	elif hasattr(self, '_prepare_input_images'):
	prepare_fn = self._prepare_input_images
	else:
	raise AttributeError("Neither _prepare_image_like_inputs nor _prepare_input_images method found")
	# transformers >= 4.53.0: uses _prepare_image_like_inputs
	prepare_fn = self._prepare_image_like_inputs

[Rayen023]

transformers >= 4.54.0 uses _prepare_image_like_inputs, while transformers <= 4.53.3 uses _prepare_input_images. This compatibility check prevents the AttributeError: 'Eagle25VLImageProcessorFast' object has no attribute '_prepare_image_like_inputs' when using transformers 4.53.x.