Skip to content

v1.1: held-out region and background models

Latest

Choose a tag to compare

@Shafwansafi06 Shafwansafi06 released this 11 Nov 11:54
· 1 commit to main since this release

Four ResNet-50 models (448 px) trained on the held-out cluster splits of Experiments 3 and 4. Same recipe as v1.0: SGD (Nesterov, momentum 0.9, lr 0.01, head lr x10, weight decay 1e-4), cosine schedule, 30 epochs, torchvision ImageNet weights (resnet50.tv_in1k), float16 mixed precision. Best checkpoint by val accuracy; weights only.

File Split Mode Epoch Val Held-out test SHA-256
resnet50_448_geography_full.pth Geography (iNaturalist, 156 species) full image 28 81.94% 62.68% (western N. America, 18,953) 527404323a41a8cc331489c29dedaacd483adcb102b58cee58c67f113ee9e703
resnet50_448_bgsplit_full.pth Background clusters (CUB) full image 25 85.66% 83.96% (ground / grass, 2,350) 9501f27a4e4a7927c7fdf5d5de76e61404a03752564a1011717f136719b43dd7
resnet50_448_bgsplit_crop.pth Background clusters (CUB) bird crop 23 86.93% 85.49% d0449c02294d400aca99b1467b3aa9fb0348a793fbb85545c0b2a2f473ab49b2
resnet50_448_bgsplit_birdshift.pth Background clusters (CUB) BirdShift 30 86.65% 84.30% 810a41f863992b8a70016ee9ec34df878d3d63a111f8c82c7bf81855ec202228

These models are trained on the split manifests in data/splits/, not on CUB's official train/test split, so they should not be scored on the official CUB test set (the background split puts about half of the official test images into training). The best model on the official split is still birdshift_resnet50_448_cub.pth in v1.0 (full-image, 84.55%).

Load: BirdResNet(num_classes=200, pretrained=False).load_state_dict(ck["model_state_dict"]). Each file also stores mode, image_size and manifest. Evaluate with scripts/eval_test.py --manifest <split csv> --splits val test (copy the file to checkpoints/<run>/best.pth). Full-image and BirdShift take Resize(512) + CenterCrop(448); the crop model takes the ground-truth box crop, square-padded and resized to 448.