Created by: littletomatodonkey
Based on the same training strategy with ResNet50(Top1 Acc 76.5%, models size 99M), CSPResNet50 Top1 Acc is 77.17% on ImageNet 1k val(while the model size is just 86M).
- Note: resize short size and crop size are 256 for CSPResNet50 eval process.
Inference benchmark on T4 GPU
| Models | Crop Size | Resize Short Size | FP16 Batch Size=1 (ms) |
FP16 Batch Size=4 (ms) |
FP16 Batch Size=8 (ms) |
FP32 Batch Size=1 (ms) |
FP32 Batch Size=4 (ms) |
FP32 Batch Size=8 (ms) |
|---|---|---|---|---|---|---|---|---|
| CSPResNet50 | 256 | 256 | 3.08307 | 5.96324 | 9.86307 | 3.79766 | 9.24683 | 16.68879 |