Script to run regular training gluonCV with and without Data Extension Library (DALI).:
To launch training use scripts/classification/imagenet/test.sh, which will install DALI.
Before launching this script you can set following environment variables:
MODEL, NUM_TRAINING_SAMPLES, NUM_EPOCHS, DATA_BACKEND, TRAIN_DATA_DIR
If some (or all) of these variables are NOT set, the script will use their default values:

  export MODEL=resnet18_v1
  export NUM_TRAINING_SAMPLES=1281167
  export NUM_EPOCHS=3
  export NUM_GPUS=0
  export DATA_BACKEND='mxnet'
  export TRAIN_DATA_DIR=~/.mxnet/datasets/imagenet

To launch this script with DALI, the environment variable DATA_BACKEND should be set to dali-gpu OR dali-cpu

Following charts shows
the improvement of performance when DALI is used:

and the accuracy achieved for 3 epochs: