CIFAR-5m is a dataset of ~6 million synthetic CIFAR-10-like images (RGB 32 x 32px). This dataset was used in the Deep Bootstrap paper.
It was generated by sampling the DDPM generative model of Ho et al., which was trained on the CIFAR-10 train set. The unconditional images were then labeled by a 98.5% accurate Big-Transfer model. Specifically, we used the pretrained BiT-M-R152x2 model, fine-tuned to CIFAR-10.
CIFAR-5m is available publicly on Google Cloud Storage as 6 npz files,
accessible at gs://gresearch/cifar5m/part{i}.npz
for i
in {0,...,5}
.
Or via HTTP, e.g. https://storage.cloud.google.com/gresearch/cifar5m/part0.npz.
The distribution of CIFAR-5m is of course not identical to CIFAR-10, but is close for research purposes. The following tables are baselines of training a network on 50K samples of either dataset (CIFAR-5m, CIFAR-10), and testing on both datasets.
ResNet18 trained with standard data-augmentation:
Trained On | Test Error On | → |
---|---|---|
↓ | CIFAR-10 | CIFAR-5m |
CIFAR-10 | 0.050 | 0.096 |
CIFAR-5m | 0.110 | 0.106 |
WideResNet28-10 trained with cutout augmentation:
Trained On | Test Error On | → |
---|---|---|
↓ | CIFAR-10 | CIFAR-5m |
CIFAR-10 | 0.032 | 0.091 |
CIFAR-5m | 0.088 | 0.097 |