Image as Set of Points - ICLR'23 [Oral, Top5%]
by Xu Ma*, Yuqian Zhou*, Huan Wang, Can Qin, Bin Sun, Chang Liu, Yun Fu.
Clustering Process (from scratch to last epoch):
TO DO (Mar 9):
- Fix a small bug in Line 170 and re-train all checkpoints.
- update the checkpoints (conv1x1 -> nn.linear, shape doesn't match)
- add large model and checkpoints
- release codes/ checkpoints for CoC without region partition (re-trained with updated codes, get better results)
- release the visualization script.
Image Classification
1. Requirements
torch>=1.7.0; torchvision>=0.8.0; pyyaml; timm; einops; apex-amp (if you want to use fp16);
data prepare: ImageNet with the following folder structure, you can extract ImageNet by this script.
โimagenet/
โโโtrain/
โ โโโ n01440764
โ โ โโโ n01440764_10026.JPEG
โ โ โโโ n01440764_10027.JPEG
โ โ โโโ ......
โ โโโ ......
โโโval/
โ โโโ n01440764
โ โ โโโ ILSVRC2012_val_00000293.JPEG
โ โ โโโ ILSVRC2012_val_00002138.JPEG
โ โ โโโ ......
โ โโโ ......
2. Pre-trained Context Cluster Models
We update the new checkpoints and logs (fix bugs, FC implementation). Feel free to download.
Model | #params | Image resolution | Top1 Acc | Throughtput | Download |
---|---|---|---|---|---|
ContextCluster-tiny | 5.3M | 224 | 71.9 | 518.4 | [checkpoint & logs] |
ContextCluster-tiny2 | 5.3M | 224 | 72.0 | 510.8 | [checkpoint & logs] |
ContextCluster-tiny_plain (w/o region partition) | 5.3M | 224 | 72.8 | - | [checkpoint] |
ContextCluster-small | 14.0M | 224 | 77.8 | 513.0 | [checkpoint & logs] |
ContextCluster-medium | 27.9M | 224 | 81.2 | 325.2 | [checkpoint & logs] |
3. Validation
To evaluate our Context Cluster models, run:
MODEL=coc_tiny #{tiny, tiny2 small, medium}
python3 validate.py /path/to/imagenet --model $MODEL -b 128 --checkpoint {/path/to/checkpoint}
4. Train
We show how to train Context Cluster on 8 GPUs. The relation between learning rate and batch size is lr=bs/1024*1e-3. For convenience, assuming the batch size is 1024, then the learning rate is set as 1e-3 (for batch size of 1024, setting the learning rate as 2e-3 sometimes sees better performance).
MODEL=coc_tiny # coc variants
DROP_PATH=0.1 # drop path rates
python3 -m torch.distributed.launch --nproc_per_node=8 train.py --data_dir /dev/shm/imagenet --model $MODEL -b 128 --lr 1e-3 --drop-path $DROP_PATH --amp
5. Clustering Visualization
We provide a script to visualize the clustering results of CoC for a given stage, block, head.
Different layers/heads will present different clustering patterns.
# Use example (generated image will saved to images/cluster_vis/{model}):
python cluster_visualize.py --image {path_to_image} --model {model} --checkpoint {path_to_checkpoint} --stage {stage} --block {block} --head {head}
See folder pointcloud for point cloud classification taks on ScanObjectNN.
See folder detection for Detection and instance segmentation tasks on COCO.
See folder segmentation for Semantic Segmentation task on ADE20K.
BibTeX
@inproceedings{ma2023image,
title={Image as Set of Points},
author={Xu Ma and Yuqian Zhou and Huan Wang and Can Qin and Bin Sun and Chang Liu and Yun Fu},
booktitle={The Eleventh International Conference on Learning Representations},
year={2023},
url={https://openreview.net/forum?id=awnvqZja69}
}
Acknowledgment
Our implementation is mainly based on the following codebases. We gratefully thank the authors for their wonderful works.
pointMLP, poolformer, pytorch-image-models, mmdetection, mmsegmentation.
License
The majority of Context Cluster is licensed under an Apache License 2.0