Created by: shippingwang
This PR
- add high performance preprocessing in PaddleClas
- add DALI in PaddleClas
It Improves almost 200%~400% training speed in no mixup series model and 350% ~ 400% training speed in mixup series model
please install paddlepaddle-gpu compiled by GCC ≥5.4
official paddle whl package works, GCC ≥5.4 environment is still required though!
and install dali at first.
- Autoaugment
- RandAugment hasn't been supported yet.
run tools/run_dali.sh to start
- Performance
statistics time = 10th ~ 110th batch time / 100 on NVIDIA® Tesla® V100 '_vd' series use mixup preprocessing
| Model | batch size | cards | using cpu preprocess time(ms/batch) | using gpu preprocess time(ms/batch) |
|---|---|---|---|---|
| ResNet50 | 128 | 1 | 359.3 | 102.3 |
| 256 | 2 | 445.4 | 195.2 | |
| ResNet50_vd | 128 | 1 | 396.8 | 121.4 |
| 256 | 2 | 737.9 | 208.7 |
MobileNetV1 (71.30%) now _vd series ResNet18_vd 72.61