Updated: 2023/07/03
- CUDA 11.8.0
- CUDNN 8.6.0
- Python 3.8.10
- Tensorflow 2.12.0
- KerasCV 0.5.0 (f05494c1057c95cbf44abac3238afcf262a50431)
- KerasAug 0.5.6
pip install tensorflow==2.12.0 keras-aug==0.5.6 git+https://github.com/keras-team/keras-cv.git tensorflow-datasets tqdmcd benchmarks
# run with GPU (KerasAug vs. KerasCV)
TF_CPP_MIN_LOG_LEVEL=2 python run_gpu_benchmark.py- Intel i7-7700K
- NVIDIA GTX 1080 8GB
- Unit: FPS (frames per second)
- Metric: The mean of the fastest 80% of the 20 trials
- Image size: (640, 640, 3), float32
- Batch size: 128
- Graph mode (
@tf.function)
KerasAug is generally faster than KerasCV.
| Type | Layer | KerasAug | KerasCV | |
|---|---|---|---|---|
| Geometry | RandomHFlip | 2123 | 1956 | fair |
| RandomVFlip | 1871 | 1767 | fair | |
| RandomRotate | 1703 | 1723 | fair | |
| RandomAffine | 2578 | 2355 | fair | |
| RandomCropAndResize | 2664 | 213 | +1150% | |
| Resize (224, 224) | 2480 | 222 | +1017% | |
| Intensity | RandomBrightness | 3052 | 2768 | fair |
| RandomContrast* | 3099 | 2976 | fair | |
| RandomBrightnessContrast* | 2881 | 609 | +373% | |
| RandomColorJitter* | 2013 | 597 | +237% | |
| RandomGaussianBlur | 2345 | 203 | +1055% | |
| Invert | 2691 | X | ||
| Grayscale | 2917 | 3116 | fair | |
| Equalize | 196 | 139 | +41% | |
| AutoContrast | 3095 | 3025 | fair | |
| Posterize | 3033 | 2144 | fair | |
| Solarize | 3133 | 2972 | fair | |
| Sharpness | 2982 | 2833 | fair | |
| Regularization | RandomCutout | 2994 | 2795 | fair |
| RandomGridMask | 918 | 196 | +368% | |
| Mix | CutMix | 2967 | 2957 | fair |
| MixUp | 1897 | 1861 | fair | |
| Auto | AugMix | 79 | X (Error) | |
| RandAugment | 301 | 246 | +22% |
*: The implementation of contrast adjustment in KerasCV differs from that of KerasAug.