TSViT — PASTIS24 checkpoints (5-fold)
Per-fold best.pth checkpoints (model weights only, state_dict) from training
TSViT on the PASTIS24 benchmark
(24x24-patch split of PASTIS),
reproducing Tarasiou et al., "ViTs for SITS: Vision Transformers for Satellite
Image Time Series", CVPR 2023.
Each checkpoint is the epoch with the best eval IoU during training for that fold,
using configs/PASTIS24/TSViT_fold<N>.yaml from the repo (architecture: TSViT,
dim=128, 4 temporal + 4 spatial transformer layers, ~1.7M params).
Files
| File | Fold | Test OA | Test mIoU |
|---|---|---|---|
tsvit_fold1_best.pth |
1 | 83.21 | 64.94 |
tsvit_fold2_best.pth |
2 | 84.17 | 66.96 |
tsvit_fold3_best.pth |
3 | 83.43 | 65.38 |
tsvit_fold4_best.pth |
4 | 83.08 | 63.47 |
tsvit_fold5_best.pth |
5 | 84.10 | 67.27 |
| 5-fold average | 83.60 | 65.60 | |
| Paper (5-fold average) | 83.4 | 65.4 |
Test-set metrics computed with masked cross-entropy loss (background class masked), overall accuracy (OA, pixel-averaged) and mean IoU (mIoU, class-averaged), matching the paper's evaluation protocol.
Usage
Each file is a raw PyTorch state_dict (not a full checkpoint dict). Load it into a
TSViT model instance from the TSViT repo:
import torch
from models.TSViT.TSViTdense import TSViT
model_config = {
'img_res': 24, 'max_seq_len': 60, 'num_channels': 11, 'num_classes': 19,
'patch_size': 2, 'dim': 128, 'temporal_depth': 4, 'spatial_depth': 4,
'heads': 4, 'pool': 'cls', 'dim_head': 32, 'scale_dim': 4,
'dropout': 0., 'emb_dropout': 0.,
}
net = TSViT(model_config)
net.load_state_dict(torch.load("tsvit_fold1_best.pth", map_location="cpu"))
net.eval()