Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -1,3 +1,50 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: cc-by-nc-4.0
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-nc-4.0
|
| 3 |
+
library_name: timm
|
| 4 |
+
tags:
|
| 5 |
+
- icc
|
| 6 |
+
- convnextv2
|
| 7 |
+
- routing
|
| 8 |
+
- retrieval
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# ICC Weights (ConvNeXt V2)
|
| 12 |
+
|
| 13 |
+
This repository provides the ICC (Image Complexity Classifier) weights used in
|
| 14 |
+
"Image Complexity-Aware Adaptive Retrieval for Efficient Vision-Language Models".
|
| 15 |
+
|
| 16 |
+
Paper: https://arxiv.org/abs/2512.15372
|
| 17 |
+
Code: https://github.com/MikelWL/ICAR
|
| 18 |
+
|
| 19 |
+
The ICC is a ConvNeXt V2 classifier fine-tuned from an ImageNet-22K pretrained
|
| 20 |
+
checkpoint. It is used to route images between early-exit and full-path
|
| 21 |
+
inference in ICAR.
|
| 22 |
+
|
| 23 |
+
## Contents
|
| 24 |
+
|
| 25 |
+
This repository ships a single file at the repo root:
|
| 26 |
+
|
| 27 |
+
- `icc.pt`
|
| 28 |
+
|
| 29 |
+
## Usage
|
| 30 |
+
|
| 31 |
+
Point ICAR to the ICC weights with `--icc-checkpoint`:
|
| 32 |
+
|
| 33 |
+
```bash
|
| 34 |
+
python scripts/evaluate_mixed_preprocessed.py \
|
| 35 |
+
--config icar/configs/coco.yaml \
|
| 36 |
+
--checkpoint checkpoints/icar_coco/layer_12/latest_checkpoint.pt \
|
| 37 |
+
--base-dataset mscoco \
|
| 38 |
+
--base-data-root /path/to/coco-images \
|
| 39 |
+
--laion-data-root /path/to/laion_coco_100k \
|
| 40 |
+
--complexity-scores /path/to/laion_coco_100k_metadata/complexity_scores.json \
|
| 41 |
+
--early-exit-layer 12 \
|
| 42 |
+
--use-icc-routing \
|
| 43 |
+
--icc-checkpoint /path/to/icc.pt
|
| 44 |
+
```
|
| 45 |
+
|
| 46 |
+
## License
|
| 47 |
+
|
| 48 |
+
The ICC weights are derived from ImageNet-pretrained ConvNeXt V2 models, which
|
| 49 |
+
are licensed under CC-BY-NC. Please use these weights for non-commercial
|
| 50 |
+
research purposes and provide appropriate attribution.
|