[WWB] Fix loading the tokenizer for VLMs

[l-bat]

No description provided.

[Copilot]

Pull Request Overview

This PR fixes the tokenizer loading for visual-language models in the WWB benchmark and adjusts the resolution parameters for image generation in the text-to-image evaluator.

• For visual-text tasks, the tokenizer is now obtained from the processor when available.
• In the text-to-image evaluator, the width and height parameters have been swapped to better reflect the intended image dimensions.

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
tools/who_what_benchmark/whowhatbench/wwb.py	Use processor.tokenizer when available for visual-text tasks
tools/who_what_benchmark/whowhatbench/text2image_evaluator.py	Swap resolution indices for width and height in image generation

Comments suppressed due to low confidence (2)

tools/who_what_benchmark/whowhatbench/wwb.py:421

• The updated logic now prioritizes processor.tokenizer for visual-text tasks. Please ensure that the processor is reliably instantiated so that the conditional fallback to load_tokenizer(args) is not frequently triggered.

            tokenizer = processor.tokenizer if hasattr(processor, "tokenizer") else load_tokenizer(args)

tools/who_what_benchmark/whowhatbench/text2image_evaluator.py:124

• The resolution parameters have been swapped for width and height. Please verify that using self.resolution[1] for width and self.resolution[0] for height aligns with the expected image dimensions throughout the codebase.

                    width=self.resolution[1],