---
license: mit
library_name: timm
pipeline_tag: image-classification
tags:
- image-classification
- vision-transformer
- adaptive-inference
- imagenet-1k
- progresvit
- arxiv:2609.03216
datasets:
- ILSVRC/imagenet-1k
metrics:
- accuracy
---

# ProgResViT DeiT-S (192 → 240)

ImageNet-1K EMA weights for **ProgResViT DeiT-S (192 → 240)** from [ProgResViT: Progressive Resolution and Width for Adaptive Vision Transformers](https://arxiv.org/abs/2609.03216).

- Paper: [arXiv:2609.03216](https://arxiv.org/abs/2609.03216)
- Hugging Face paper: [2609.03216](https://huggingface.co/papers/2609.03216)
- Code: [ds-kiel/ProgResViT](https://github.com/ds-kiel/ProgResViT)
- Training: standard
- Full-path top-1 accuracy: 82.206%
- Full-path compute: 6.267 GMACs

## Usage

Run from the cloned ProgResViT repository root, or add the repository to `PYTHONPATH`:

```python
import torch
from timm.models import create_model

model = create_model("hf-hub:NCPS/progresvit-deit-s-192-240-imagenet1k", pretrained=True)
model.eval()

x = torch.randn(1, 3, 240, 240)
with torch.inference_mode():
    logits, stage = model(x, threshold=0.35)

print(logits.shape)  # (1, 1000)
print(stage)         # 0 = 192 px / 3 heads; 1 = 240 px / 6 heads
```

The entropy threshold controls routing. Lower values send more images to the second round; higher values exit more images after the first round.

## ImageNet-1K validation

Download the release checkpoint into `checkpoints/`:

```python
from huggingface_hub import hf_hub_download

hf_hub_download(
    repo_id="NCPS/progresvit-deit-s-192-240-imagenet1k",
    filename="progresvit_192_240.pth.tar",
    local_dir="checkpoints",
)
```

Then run the repository evaluator:

```bash
python validate.py /path/to/imagenet/val \
  --config 192_240 \
  --checkpoint checkpoints/progresvit_192_240.pth.tar
```