PEIRA: pretrained checkpoints
Checkpoints for PEIRA: Learning Predictive Encoders through Inter-View Regressor Alignment
(Arbel*, Terver*, Ponce, NeurIPS 2026; arXiv:2605.17671).
Project page: peira-ssl.github.io.
Code: github.com/Basile-Terv/PEIRA.
Loss: pip install peira-loss (github.com/MichaelArbel/peira-loss).
Each file is the ResNet-50 backbone only (no projector), pretrained for 100 epochs on
ImageNet-1K (batch 2048, LARS), for linear probing and transfer; these are the baseline and PEIRA
rows of Table 1 of the paper. A file is a torch.save dict with model_state_dict (backbone
weights), epoch (0-indexed, so 99), the training config and a metadata dict (method, projector dims
used in training, seed, epochs, offline top-1). Top-1 is the offline linear probe (VICReg protocol: SGD, 100
epochs, batch 1024) on frozen backbone features.
Files
| Method | Projector hidden | k (projector output) | Seed | ImageNet-1K top-1 | Weights |
|---|---|---|---|---|---|
| SIGReg | 2048 | 128 | 1 | 66.46 | in1k_rn50_sigreg_seed1.pth |
| SIGReg | 2048 | 128 | 1000 | 66.38 | in1k_rn50_sigreg_seed1000.pth |
| SIGReg | 2048 | 128 | 10000 | 65.94 | in1k_rn50_sigreg_seed10000.pth |
| VICReg | 8192 | 8192 | 1 | 68.80 | in1k_rn50_vicreg_seed1.pth |
| VICReg | 8192 | 8192 | 1000 | 68.71 | in1k_rn50_vicreg_seed1000.pth |
| VICReg | 8192 | 8192 | 10000 | 68.90 | in1k_rn50_vicreg_seed10000.pth |
| PEIRA | 8192 | 256 | 1 | 66.51 | in1k_rn50_peira_k256_seed1.pth |
| PEIRA | 8192 | 256 | 1000 | 66.32 | in1k_rn50_peira_k256_seed1000.pth |
| PEIRA | 8192 | 256 | 10000 | 66.51 | in1k_rn50_peira_k256_seed10000.pth |
| PEIRA | 8192 | 1024 | 1 | 69.23 | in1k_rn50_peira_k1024_seed1.pth |
| PEIRA | 8192 | 1024 | 1000 | 69.12 | in1k_rn50_peira_k1024_seed1000.pth |
| PEIRA | 8192 | 1024 | 10000 | 69.15 | in1k_rn50_peira_k1024_seed10000.pth |
PEIRA recipe: EMA Σ (Algorithm 1), λ=0.1, η 0.99→0.95, grad clip 1.0. Three-seed means (± std over seeds): k=256 reaches 66.45 ± 0.11 on ImageNet-1K (70.89 ± 0.18 average transfer); k=1024 reaches 69.17 ± 0.06 (71.09 ± 0.24 average transfer).
Usage
Standalone, the backbone is a torchvision ResNet-50 with fc = Identity():
import torch, torchvision
from huggingface_hub import hf_hub_download
ckpt = hf_hub_download("peira-ssl/PEIRA", "in1k_rn50_peira_k1024_seed1.pth")
sd = torch.load(ckpt, map_location="cpu", weights_only=True)["model_state_dict"]
backbone = torchvision.models.resnet50()
backbone.fc = torch.nn.Identity()
backbone.load_state_dict({k.removeprefix("backbone.backbone."): v for k, v in sd.items()
if k.startswith("backbone.backbone.")})
backbone.eval()
with torch.no_grad():
features = backbone(torch.randn(1, 3, 224, 224)) # [1, 2048]
With the code repository installed, probe.py takes these
files directly, e.g. for the transfer probe:
python probe.py --ckpt <ckpt> --datasets lejepa8 --transfer-root $EBJEPA_DSETS/transfer --out peira.json
License
The weights are released under Apache-2.0, like the code. They were trained on ImageNet-1K, whose terms of access apply.
Citation
@inproceedings{arbel2026peira,
title = {{PEIRA}: Learning Predictive Encoders through Inter-View Regressor Alignment},
author = {Arbel, Michael and Terver, Basile and Ponce, Jean},
booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
year = {2026},
eprint = {2605.17671},
archivePrefix = {arXiv},
}