This repository hosts soft-equivariant vision models introduced in our paper. This repository contains a soft-equivariant VIT fine-tuned on ADE20k. It uses a linear segmentation head.

Usage

Note: All models require trust_remote_code=True because they use custom model classes.

Semantic Segmentation (ViT backbone)

python

from transformers import AutoModel, AutoConfig
import torch
import torch.nn.functional as F

model_id = "ashiq24/softeq-dinov2-base-ade20k-seg-c4-s0.8-sp1.0"

config = AutoConfig.from_pretrained(model_id, trust_remote_code=True)
model  = AutoModel.from_pretrained(model_id, trust_remote_code=True)
model.eval()

# Input size must match model training resolution (e.g., 512×512)
pixel_values = torch.randn(1, 3, 512, 512)

with torch.no_grad():
    outputs = model(pixel_values=pixel_values)

# outputs.logits shape: (1, num_labels, H, W) — already upsampled to input resolution
seg_map = outputs.logits.argmax(dim=1)   # (1, H, W) predicted label per pixel

Configuration Parameters

The SoftEqConfig class stores all architectural parameters. Key fields:

Parameter	Type	Description
`n_rotations`	`int`	Size of the discrete rotation group (e.g., `4` for C4, `720` for near-continuous)
`soft_thresholding`	`float`	Softness of the patch-embedding filter in `[0, 1]`; `0` = strict equivariance, `1` = no filter
`soft_thresholding_pos`	`float`	Softness of the positional-embedding filter in `[0, 1]`
`group_type`	`str`	Symmetry group: `"rotation"` or `"roto_reflection"`
`hard_mask`	`bool`	Use a hard (step-function) mask instead of exponential damping
`model_arch`	`str`	Architecture variant (see table above)
`pretrained_model`	`str`	HuggingFace identifier of the base backbone
`num_labels`	`int`	Number of output classes

Citation

If you use these models in your research, please cite:

bibtex

@article{rahman2026tunable,
  title={Tunable Soft Equivariance with Guarantees},
  author={Rahman, Md Ashiqur and Hao, Lim Jun and Jiang, Jeremiah and Lim, Teck-Yian and Yeh, Raymond A},
  journal={arXiv preprint arXiv:2603.26657},
  year={2026}
}