Some scripts to test ImageNet reading speed.

• With tensorpack.dataflow (pure python loader):

Augmentations=[GoogleNetResize, Lighting, Flip]: reach 5.6k image/s on a DGX1.

python benchmark-dataflow.py /path/to/imagenet --batch 128 --name train --aug resizeAndLighting

• With tf.data:

Augmentations=[GoogleNetResize, Flip]: 11k image/s on a DGX1. As a reference, a DGX1 with 8 V100s can train ResNet-50 in fp32 at 2.6k image/s.

python benchmark-tfdata.py /path/to/imagenet --name train --batch 128