Table 4 Performance scores for various baselines on UltraMNIST21 benchmark. Here Acc*@24 GB % denotes the accuracy score in % obtained for full utilization of 24 GB GPU memory. Further, Acc@24 GB % denotes the best accuracy score for the GPU utilization of 24 GB memory or even less (e.g., 16 GB, 11 GB, etc).
From: An UltraMNIST classification benchmark to train CNNs for very large images
Method | Backbone | Input size | Acc.*@24 GB (%) | Acc.@24 GB (%) |
|---|---|---|---|---|
RN-256 | ResNet-50 | 256 × 256 | 5.2 | 8.9 |
RN-512 | ResNet-50 | 512 × 512 | 24.8 | 24.8 |
ENB0-256 | EfficientNet-B0 | 256 × 256 | 20.4 | 20.4 |
ENB0-512 | EfficientNet-B0 | 512 × 512 | 49.9 | 49.9 |
ENB0-1024 | EfficientNet-B0 | 1024 × 1024 | 52.3 | 52.3 |
ENB3-256 | EfficientNet-B3 | 256 × 256 | 30.3 | 30.3 |
ENB3-512 | EfficientNet-B3 | 512 × 512 | 67.5 | 69.1 |
MBNet-demir | MobileNet-V3 | 1280 × 1280 | — | 86.2 |
