Autoregressive Generative Modeling on CIFAR-10 (test)
2.8BPDSparse Trans.
Evaluation results
| Method | Links | |
|---|---|---|
| Sparse Trans.Architecture=2D Attention, #params=59M2023.05 | 2.8 | |
| MULTIRESNETArchitecture=Convolution (no 2D bias), #params=38M2023.05 | 2.84 | |
| PixelSNAILArchitecture=2D Convolution + Attention, #params=46M2023.05 | 2.85 | |
| S4Architecture=State-Space Models + U-Net structure2023.05 | 2.85 | |
| Image Trans.Architecture=1D Attention2023.05 | 2.9 | |
| PixelCNN++Architecture=2D Convolution, #params=53M2023.05 | 2.92 | |
| PixelRNNArchitecture=RNN + 2D bias2023.05 | 3 | |
| Gated PixelCNNArchitecture=2D Convolution2023.05 | 3.03 | |
| PixelCNNArchitecture=2D Convolution2023.05 | 3.14 |