Dataset condensation is proposed in paper Dataset Condensation with Gradient Matcing. It aims to condense a large traing set into a small synthetic set such tahat model trained on the small synthetic set would obtain comparable testing performance to that trained on large training set.
Our paper propose attention masking to input images so that model can learn only important visual parts of the image. Future work can further stretch out to background removal or unbiasing. Our paper is accepted to IPIU2022.
Install packages in the requirements. Attention mask can be downloaded here.
python main.py --dataset CIFAR10 --model AlexCifarNet --ipc 10
# --ipc (images/class): 1, 10, 20, 30, 40, 50
python main.py --dataset CIFAR10 --model AlexCifarNet --ipc 10 --method DSA --init real --dsa_strategy color_crop_cutout_flip_scale_rotate
# --ipc (images/class): 1, 10, 20, 30, 40, 50
| DD | DC | Ours | |
|---|---|---|---|
| 1 img/cls | - | 24.2 | 24.5 |
| 10 img/cls | 36.8 | 39.1 | 40.1 |
Table 1: Testing accuracies (%) of AlexNet trained from scratch on 1 or 10 synthetic image(s)/class.