TR-HASH-Vision-v6-1M-VOC

This repository contains the validated TR-Hash detector checkpoint.

TR-Hash Vision v6 is a compact anchor-free detector built on an ID-hash-routed MoE vision tower. Spatial-token IDs select expert parameter subspaces while shared attention preserves contextual feature mixing. The detector combines native hierarchical P3/P4/P5 features, shifted-window attention, a lightweight PAN, an optional P2 small-object path, dynamic one-to-many assignment with STAL, decoupled LTRB/DFL regression, unified QFL quality-class scores, and an optional one-to-one NMS-free inference branch. The realized model has 0.97M parameters and uses 640 px inputs on Pascal VOC 2007+2012.

Evaluation

Inference branch	mAP50	mAP50-95	AP small	AP medium	AP large	Precision	Recall	Best F1	Best confidence
O2M + NMS	0.3565	0.1357	0.0980	0.0562	0.1490	0.0154	0.8540	0.4621	0.322
NMS-free	0.1709	0.0789	0.0836	0.0449	0.0850	0.0111	0.8716	0.2766	0.258

These metrics use all 4,952 Pascal VOC validation images. model.predict() uses the NMS-free branch by default; pass nms_free=False to evaluate or serve the higher-accuracy O2M branch with class-aware NMS. Precision and recall in the table use the deliberately low evaluation threshold; Best F1 and Best confidence report the calibrated operating point found on validation.

Training curves

The main figure reports full-validation AP, size-specific AP, calibration and the learning-rate schedule.

Raw regression losses are split into Mosaic-active and Mosaic-disabled panels because disabling Mosaic changes the input distribution. The plotted values are not inverted or normalized across phases. The scheduled combined objective is intentionally omitted from the graphs because its box and one-to-one weights increase during training, so its absolute value is not comparable across steps. The complete weighted objective and every raw component remain available in metrics.jsonl.

Qualitative predictions

Each sample is shown as ground truth, O2M + NMS, then NMS-free. This qualitative grid is diagnostic rather than a substitute for the full validation metrics.

Inference

from PIL import Image
import torch

from complexity.generative.detection import (
    load_detector_from_hub,
    preprocess_detector_image,
    restore_detector_boxes,
)

device = "cuda" if torch.cuda.is_available() else "cpu"
model = load_detector_from_hub("AETHORIA-AI/TR-HASH-Vision-v6-1M-VOC", device=device)
pixels, metadata = preprocess_detector_image(
    Image.open("image.jpg"), model.config.image_size
)

with torch.inference_mode():
    prediction = model.predict(pixels[None].to(device))[0]
prediction["boxes"] = restore_detector_boxes(
    prediction["boxes"].cpu(), metadata
)

Class IDs are listed in class_names.json. Boxes returned by predict are normalized xyxy coordinates until restore_detector_boxes maps them to source pixels.

Training

• Dataset: Pascal VOC 2007+2012
• Data: 16,551 training and 4,952 validation images
• Budget: 50 epochs / 12,950 optimizer steps
• Hardware: 4× NVIDIA GeForce RTX 5060 Ti 16 GB
• Precision: BF16 with NCCL DDP and fused CUDA/Triton routed kernels
• Optimizer: MuSGD
• Routed-expert LR multiplier: 1.5x
• Backbone initialization: TR-Hash Vision v6 ImageNet-1K pretraining
• Training: EMA, multi-resolution, Mosaic, MixUp, Copy-Paste and random erasing
• Framework: Complexity Framework

Release files

• ema.safetensors — recommended full detector weights
• model.safetensors — final non-EMA detector weights
• ema_tower.safetensors and tower.safetensors — reusable vision tower
• validation.json and validation_nms_free.json — full validation results
• metrics.jsonl — 647 training logs and 10 full validation records
• training_summary.json — compact machine-readable run summary
• training_curves.png, loss_curves.png and prediction_grid.jpg — verified release figures
• SHA256SUMS — integrity hashes for all Safetensors files

Limitations

This is a research checkpoint under CC BY-NC 4.0. Validate accuracy, calibration, latency, and failure modes on your own target domain. A training configuration or active run is not evidence of accuracy; release claims require a realized checkpoint and an explicit evaluation protocol.