GeoNUSAF Stage 4 - SegNeXt-T - arm R0 - block fold 1

Kathmandu Valley land-use segmentation, 6 classes, ignore_index=255.

field value
architecture MSCAN-T encoder + LightHamHead decoder (SegNeXt, NeurIPS 2022)
arch_sig c615877e4f3e
encoder init ImageNet-1K, 100.0% of tensors loaded
params 4.23 M
head fuse stride 8 (stages [1, 2, 3])
NMF rank / steps R=16, 6 train / 7 eval, rand_init=True
arm R0 (real only)
train pairs 804 (804 real + 0 synthetic)
synthetic source sugam24/geonusaf-stage3-fakepairs-block-fold1, fake_pairs_R1
split block fold 1, train_sha1 eb0aebdd9919
schedule 6000 steps, warmup 500, cosine 6000
lr (head/encoder) 0.0006 / 6e-05
class weights from real
seed 42
best step 5800
val mIoU 0.5684
val mF1 0.7070
val OA 0.8366
val kappa 0.7046

Per-class (validation, 136 real fold-1 tiles)

class IoU F1
Residential 0.8576 0.9234
Road 0.4546 0.6250
River 0.4594 0.6296
Forest 0.7461 0.8546
UnusedLand 0.3085 0.4715
Agricultural 0.5845 0.7378

Validation contains no synthetic pixels in either arm.

best.pt holds model_state (EMA weights when EMA is on), arch, arch_sig, the run cfg and metrics. Rebuild with segnext_model.py from this same repo. Model code derives from Visual-Attention-Network/SegNeXt (Apache-2.0).

Not comparable to the part-1 fold-1 SegNeXt run: that one had frozen augmentation (the persistent_workers bug), determinism on, MIN_DELTA=0.0, and an epoch-based schedule.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support