DuoMamba weights and experiment logs

Weights and original experiment logs for DuoMamba, accepted to ACCV 2026. The tables below provide ImageNet-1K pretrained backbones, complete COCO Mask R-CNN and ADE20K UPerNet models, and links to the corresponding logs. Reported results are from the ACCV manuscript.

Results

ImageNet-1K classification

224 × 224 input, 300 epochs, trained from scratch.

Model	#Params	FLOPs	Top-1 Acc (%)	Weights	Logs
DuoMamba-T	28M	4.8G	84.0	Weights	Logs
DuoMamba-S	41M	7.2G	84.7	Weights	Logs
DuoMamba-B	91M	15.4G	85.4	Weights	Logs

COCO object detection and instance segmentation

Mask R-CNN with ImageNet-1K initialization. FLOPs are reported at 1280 × 800. AP^b denotes bounding-box AP; AP^m denotes mask AP.

1× schedule (12 epochs)

Backbone	AP^b	AP^b₅₀	AP^b₇₅	AP^m	AP^m₅₀	AP^m₇₅	#Params	FLOPs	Weights	Logs
DuoMamba-T	48.3	70.2	53.3	43.6	67.5	47.2	48M	301G	Weights	Logs
DuoMamba-S	49.7	71.7	54.6	44.5	68.7	47.7	61M	369G	Weights	Logs
DuoMamba-B	50.6	72.5	55.6	45.2	69.6	48.8	111M	564G	Weights	Logs

3× schedule (36 epochs) with multi-scale training

Backbone	AP^b	AP^b₅₀	AP^b₇₅	AP^m	AP^m₅₀	AP^m₇₅	#Params	FLOPs	Weights	Logs
DuoMamba-T	49.9	71.3	54.7	44.3	68.5	47.8	48M	301G	Weights	Logs
DuoMamba-S	51.1	72.3	56.0	45.3	69.5	49.1	61M	369G	Weights	Logs

ADE20K semantic segmentation

UPerNet, 512 × 512 training crops, 160k iterations, total batch size 16. SS/MS denote single-scale/multi-scale evaluation; reported FLOPs follow the manuscript's 2048 × 512 convention.

Backbone	mIoU (SS)	mIoU (MS)	#Params	FLOPs	Weights	Logs
DuoMamba-T	49.5	50.0	57M	978G	Weights	Logs
DuoMamba-S	50.3	51.3	70M	1048G	Weights	Logs
DuoMamba-B	52.1	52.7	122M	1250G	Weights	Logs

Checkpoint format and usage

The classification/duomamba_tiny_imagenet1k/duomamba_tiny.pth, classification/duomamba_small_imagenet1k/duomamba_small.pth, and classification/duomamba_base_imagenet1k/duomamba_base.pth files contain classification weights under model. They can be used for classification evaluation or backbone initialization with the matching configuration.

Files under detection/ and segmentation/ contain the complete task model under state_dict, plus basic epoch/iteration and dataset metadata under meta. Every model tensor is unchanged from its source checkpoint. Optimizer, scheduler, message-hub state, and the embedded training configuration are omitted. These files are for evaluation or initialization and do not provide optimizer state for resuming training.

Download the weights matching your variant and task schedule, then follow the classification, detection, or segmentation guide. Installation requires the CUDA dependencies in the repository instructions.

Files and integrity

Weights and original experiment logs are grouped by task and experiment. Each experiment directory contains its checkpoint and corresponding log together.

classification/
  duomamba_tiny_imagenet1k/      checkpoint + log_rank0.txt
  duomamba_small_imagenet1k/     checkpoint + log_rank0.txt
  duomamba_base_imagenet1k/      checkpoint + log_rank0.txt
detection/
  mask_rcnn_duomamba_tiny_coco_1x/    checkpoint + log
  mask_rcnn_duomamba_small_coco_1x/   checkpoint + log
  mask_rcnn_duomamba_base_coco_1x/    checkpoint + log
  mask_rcnn_duomamba_tiny_coco_3x/    checkpoint + log
  mask_rcnn_duomamba_small_coco_3x/   checkpoint + log
segmentation/
  upernet_duomamba_tiny_ade20k_160k/  checkpoint + log
  upernet_duomamba_small_ade20k_160k/ checkpoint + log
  upernet_duomamba_base_ade20k_160k/  checkpoint + log

Log contents are preserved unchanged. ARTIFACTS.json lists all published weights and logs, their source filenames and records sizes and SHA-256 hashes. SHA256SUMS includes all published weights and logs.

Licensing

A project-wide license for the original DuoMamba contributions has not yet been selected. See the repository notices for third-party portions.

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

PangS00oo
/

DuoMamba