Added __init__.py files to help importlib.util.find_spec in gin resource reader to work…

[ultrons]

If the moe configs are used with t5x with something like:

python3 $HOME/t5x/t5x/train.py \ --gin_search_paths="$HOME/flaxformer" \ --gin_file=$HOME/my_config.gin \ --gin_file=$HOME/flaxformer/flaxformer/t5x/configs/moe/models/switch_base.gin \ --gin.MODEL_PARALLEL_SUBMESH="[2,1,1,1]" \ --gin.NUM_ENCODER_SPARSE_LAYERS=2 \ --gin.NUM_DECODER_SPARSE_LAYERS=2 \ --gin.MIXTURE_OR_TASK_NAME="'wikipedia_20190301.en_v003_unsupervised'" \ --gin.MIXTURE_OR_TASK_MODULE="'t5.data.tasks'" \ --gin.MODEL_DIR="'${MODEL_DIR}'" \ --gin.TASK_FEATURE_LENGTHS="{'inputs': 1024, 'targets': 1024}" \ --gin.TRAIN_STEPS=60 \ --gin.DROPOUT_RATE=0 \ --alsologtostderr \

The include lines in flaxformer/flaxformer/t5x/configs/moe/models/switch_base.gin won't work as is. Since gin resource reader relies on the python importlib to get spec object and without init file in these directories spec.origin would None and the script will crash.

[github-actions[bot]]

Unfortunately, we cannot accept contributions to the Flaxformer repo at this time. Please file issues as needed though!