TR-HASH Vision v6

Compact hierarchical token-routed vision

Research preview — no model weights or release metrics are published yet. ImageNet-1K pretraining is active. The COCO detector and its validated checkpoints will follow only after the full evaluation protocol is complete.

TR-HASH Vision v6 is a compact hierarchical vision architecture built around deterministic spatial token routing. It preserves the TR-HASH Mixture-of-Experts identity while producing native multi-scale features for classification and object detection.

Release status

Stage	Dataset	Resolution	Status
Backbone pretraining	ImageNet-1K	224 px	In progress
Detector training	COCO 2017	512–640 px multi-scale	Planned after pretraining
Same-protocol evaluation	COCO val2017	640 px	Pending
Standalone `trhash` bundle	Validated checkpoint	—	Pending

No Pascal VOC v0.3 result is being promoted as a v6 result. Previous experiments were architecture-development runs and are not release metrics.

Architecture

Component	Vision v6
Backbone	Hierarchical TR-HASH tower with native P3/P4/P5 features
Attention	Local shifted windows with flexible 2D positional interpolation
Routing	4 deterministic experts, top-2 active per spatial identity
Neck	Lightweight additive PAN with optional P2 small-object path
Detection head	Decoupled quality/classification and local LTRB/DFL regression
Assignment	Dynamic one-to-many assignment with STAL small-object supervision
End-to-end path	Optional one-to-one branch for NMS-free inference
Training	EMA, Mosaic, MixUp, Copy-Paste, random erasing and multi-resolution
Runtime	PyTorch, fused Triton on CUDA, portable MPS/CPU fallback

Model size

ImageNet-1K pretraining classifier: 0.76M parameters
Transferable hierarchical tower: 0.63M parameters
Target 80-class COCO detector: approximately 1M parameters

The final detector count will be reported from the realized checkpoint rather than inferred from the design target.

Training

The active pretraining recipe uses the complete ImageNet-1K training split, four-GPU DistributedDataParallel, AdamW, a 1.5× routed-expert learning-rate multiplier, bfloat16, fused TR-HASH CUDA/Triton kernels and exact resumable checkpoints including optimizer, scheduler, cursor and distributed RNG state.

The detector stage will transfer the best ImageNet-1K tower into the v6 COCO configuration and train at 512–640 px with PAN, P2, STAL, DFL/QFL, strong detection augmentation and EMA.

Evaluation

There are intentionally no mAP claims yet. The release will report:

COCO mAP50-95 and mAP50;
AP small, medium and large;
precision, recall and calibrated operating point;
parameters, latency, throughput and peak memory;
matched 640 px comparisons under the same dataset and evaluation protocol.

Inference

Inference instructions will be added when validated v6 weights are uploaded. The research framework is used for architecture development and training; the separate trhash runtime is the intended user-facing path for prediction, validation, export and serving.

Resources

Limitations

This is a research preview under CC BY-NC 4.0, not a production model release. Architecture implementation and an active training run are not evidence of accuracy. Validate accuracy, calibration, latency and failure modes on the intended target domain after the final checkpoint is published.

Downloads last month: -; Downloads are not tracked for this model. How to track