[Task] image-classification EP coverage investigation

Overview

Image classification models assign a category label to an input image. This task covers general-purpose ViT and CNN models as well as domain-specific variants (NSFW detection, age/gender classification, deepfake detection). All 11 passing models pass on at least one EP.

This issue focuses on the 3 models with incomplete EP coverage — two with VitisAI failures (BEiT, ConvNeXt) and one with a QNN failure (Swin-Large), suggesting architecture-specific compilation issues on each EP.

Agent Scenarios

• Content moderation agent: classify images for NSFW content, deepfake detection, or policy violations on-device before upload or sharing
• Age/gender verification agent: estimate demographics from images for access control or personalization workflows (fairface, gender-classification)
• Visual quality gate agent: route images to different processing pipelines based on content category (document vs. photo vs. diagram)

EP Coverage Status

Reference: Falconsai/nsfw_image_detection, dima806/fairface_age_image_detection, google/vit-base-patch16-224, apple/mobilevit-small, rizvandwiki/gender-classification, amunchet/rorshark-vit-base, buildborderless/CommunityForensics-DeepfakeDet-ViT, AdamCodd/vit-base-nsfw-detector, microsoft/resnet-50 already pass all 3 EPs.

Acceptance Criteria

• microsoft/beit-base-patch16-224-pt22k-ft22k passes all 3 EPs (investigate VitisAI failure)
• facebook/convnext-tiny-224 passes all 3 EPs (investigate VitisAI failure)
• microsoft/swin-large-patch4-window7-224 passes all 3 EPs (investigate QNN failure)