buildborderless

/

CommunityForensics-DeepfakeDet-ViT

like 15

Follow

Borderless R&D 18

Summary

This PR fixes the model after discovering the safetensors weights, config, and preprocessor were all incorrect. The model had been serving 500K+ monthly downloads with silently wrong inference results.

Root cause

The original HF port was hastily created for an internal proof-of-concept. The model.safetensors was converted from different weights than the training checkpoint. The config.json had ViT-Base values (12 heads, intermediate_size=3072) instead of ViT-Small (6 heads, intermediate_size=1536). The preprocessor squashed non-square images.

What changed

Critical fixes

• model.safetensors — regenerated from the authoritative training checkpoint pretrained_weights/model_v11_ViT_384_base_ckpt.pt
• config.json — num_classes 2→1, num_attention_heads 12→6, intermediate_size 3072→1536, removed redundant fields
• preprocessor_config.json — shortest_edge: 440 + do_center_crop (preserves aspect ratio)

ONNX v1.1

• 5 new variants exported from corrected weights (84 MB FP32 down to 16 MB Q4)
• Previous 8 variants archived in onnx_legacy/

Documentation

• README.md — rewritten with fix notice, breaking change warning, ONNX variant guide, v2 teaser
• CHANGELOG.md — detailed change log of every commit
• AGENTS.md — created for LLM-based coding agents

Housekeeping

• LICENSE added with original MIT copyright + Borderless contributions
• modeling_vit_classifier.py moved to scripts/ and deprecated
• model_legacy.safetensors + model_fixed.safetensors archived

Breaking change

• transformers >= 5.4.0 now required — older versions lack shortest_edge resize and will crash

Verification

Corrected model output matches the original timm training pipeline within 0.004%.

Full changelog

See CHANGELOG.md for the complete commit-by-commit breakdown.