AETHORIA-AI/TR-HASH-Vision-v8-2M-COCO

TR-HASH Vision v8 Nano — COCO

Final checkpoint from the first 50-epoch COCO pretraining run of the corrected 2.53M-parameter TR-HASH Vision v8 architecture.

TR-HASH Vision is a compact anchor-free detector with a hierarchical vision tower. Spatial-token identities route activations through two narrow experts among eight, while a wider shared SwiGLU path remains active for every token. The model was trained from random initialization without an external detector or classification backbone.

Architecture

Property	Value
Input resolution	640 × 640
Parameters	2.53M
Vision stages	2 / 2 / 3 blocks
Experts	8 stored, top-2 active
Shared / expert width	216 / 27
Detection scales	P2 / P3 / P4 / P5
Neck	normalized PAN, 2 repeats
Regression	LTRB + DFL, reg_max=16
Inference branches	O2M + NMS and one-to-one NMS-free

Evaluation

Official COCO evaluation on the 5,000-image val2017 split at 640 px. The checkpoint was selected at the final epoch by official COCO mAP50-95.

Example detection

Qualitative O2M inference example from the published checkpoint. This image is not part of the COCO evaluation set.

Plain-language context: the final O2M checkpoint reaches 16.59 COCO AP, or approximately 52% of the 32.2 AP YOLO26 reference used for this comparison. TR-HASH reached this point in one 50-epoch run from random initialization. This is a progress indicator rather than a strict apples-to-apples efficiency claim: training recipes, initialization and compute budgets differ.

O2M + NMS — recommended branch

Metric	Score
COCO AP50-95	0.1659
COCO AP50	0.2781
COCO AP75	0.1714
AP small	0.0947
AP medium	0.1796
AP large	0.2429
AR100	0.3435
Best F1 / confidence	0.4094 / 0.190

One-to-one NMS-free — experimental branch

Metric	Score
COCO AP50-95	0.0806
COCO AP50	0.1208
COCO AP75	0.0882
AP small	0.0577
AP medium	0.1093
AP large	0.1214
AR100	0.3506
Best F1 / confidence	0.2326 / 0.134

The O2M branch is the recommended inference path. NMS-free reaches similar recall but still trails in score calibration and average precision. Expand the panels above for branch-level results; the raw evaluator outputs remain in validation.json and validation_nms_free.json.

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-v8-2M-COCO",
    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),
        nms_free=False,
        confidence_threshold=0.19,
    )[0]
prediction["boxes"] = restore_detector_boxes(
    prediction["boxes"].cpu(), metadata
)

Class IDs are listed in class_names.json. The reported O2M metrics use the official COCO evaluator; validation.json contains the complete result.

Files

• ema.safetensors: recommended EMA detector weights;
• model.safetensors: non-EMA detector weights;
• training_state.pt: optimizer, scheduler and exact training state;
• tower.safetensors and ema_tower.safetensors: standalone vision tower;
• config.json: complete v8 architecture configuration;
• validation.json: official O2M + NMS COCO metrics;
• validation_nms_free.json: official NMS-free COCO metrics;
• metrics.csv: complete pretraining and validation history;
• class_names.json: the 80 COCO classes.

Both inference branches are contained in the same weights. The two metric files record their independent evaluations.

Training

• Dataset: COCO 2017
• Initialization: random
• Epochs: 50
• Optimizer: MuSGD
• Precision: BF16
• EMA: enabled
• External pretrained weights: none

Supervised clean-image refinement is planned as a separate follow-up phase and is not included in the metrics above.

Limitations

This is a research checkpoint, not a production release. Accuracy remains below mature production detectors, and the NMS-free branch requires further calibration. Validate accuracy, latency and failure modes on your own target domain before deployment.