Some tests on full 50000 ImageNet validation set
using MobileNet model and ONNX runtime,
in different pre-processing modes:

-----------------------------------------------------------------------------
bilinear interpolation on resize,crop.875,model=mobilenet:

Accuracy top 1: 0.71226 (35613 of 50000)
Accuracy top 5: 0.89834 (44917 of 50000)
-----------------------------------------------------------------------------
bilinear interpolation on resize,crop.875,model=resnet:

Accuracy top 1: 0.7617 (38085 of 50000)
Accuracy top 5: 0.92866 (46433 of 50000)
-----------------------------------------------------------------------------

Creating single-image datasets (need disambiguation in filename, extra tags and extra path) :
    ck install package --tags=dataset,preprocessed,external_file --env.CK_IMAGE_FILE=~/Desktop/lenny_kite.JPG --extra_tags=lenny_kite --extra_path=_lenny_kite
    ck install package --tags=dataset,preprocessed,external_file --env.CK_IMAGE_FILE=~/Desktop/lenny_canopy.JPG --extra_tags=lenny_canopy --extra_path=_lenny_canopy

```
$ ck virtual env --tags=dataset,imagenet,raw
$ ck install package --tags=dataset,imagenet,preprocessed \
--env.CK_IMAGE_FILE=${CK_ENV_DATASET_IMAGENET_VAL}/ILSVRC2012_val_00002916.JPEG \
--extra_tags=toilet-paper --extra_path=_ILSVRC2012_val_00002916
$ ck run program:image-classification-onnx-py --cmd_key=preprocessed --dep_add_tags.images=toilet-paper
```