Activator/vision_onnx

sorush-80m ONNX Model Spec

Exported from checkpoint checkpoints/sorush-80m.ckpt (epoch 4, step 6930).

Overview

Property	Value
Architecture	DinoV2 ViT-B/14 + 4 register tokens + projection head
Precision	FP32
File	onnx_export/sorush-80m.onnx
Opset	17

Inputs

Name	Shape	Dtype	Description
imgs	(B, 3, 518, 518)	float32	RGB images, values in [0, 1]. B is dynamic batch size.
masks	(B, 518, 518)	bool	ROI attention mask. True = target region (segment of interest). Pixels outside the target region are strictly isolated in all attention layers to prevent feature contamination.

Output

Name	Shape	Dtype	Description
embeddings	(B, 128)	float32	Clustering features. Derived from the CLS token → projection head (768 → 1024 → ReLU → 128). Use with cosine similarity or Euclidean distance for clustering.

Notes

• LoRA weights (r=8, α=16, blocks 8–11) are merged into the backbone. No LoRA ops in the graph.
• CBP features are not used (training config: use_cbp_feature=False).
• Input images must be exactly 518×518 — the positional embedding interpolation path is not active at this resolution.
• masks can be all-True (no isolation) or partial (bidirectional isolation between target/background regions).