| set -e | |
| cd emergent_communication_at_scale | |
| mkdir -p emcom_datasets/ | |
| cd emcom_datasets | |
| wget https://storage.googleapis.com/dm_emcom_at_scale_dataset/byol_celeb_a2.tar.gz | |
| tar xf byol_celeb_a2.tar.gz | |
| wget https://storage.googleapis.com/dm_emcom_at_scale_dataset/byol_imagenet2012.tar.gz | |
| tar xf byol_imagenet2012.tar.gz | |
| cd .. | |
| cd .. | |
| # Python cannot find the CUDA libraries without manually inserting the conad | |
| # environment's /lib path into the LD_LIBRARY_PATH | |
| CONDA_LIB_DIR=$(which python | sed s,bin/python,lib,) | |
| export LD_LIBRARY_PATH=${LD_LIBRARY_PATH:+$LD_LIBRARY_PATH:}$CONDA_LIB_DIR | |
| # If this is unset, the code will OOM on an 11 GiB card, possibly due to jax | |
| # and TensorFlow both preallocating. | |
| export XLA_PYTHON_CLIENT_PREALLOCATE=false | |
| python helper.py | |
| for dir in checkpoint/*/; do | |
| target=../data/$(basename $dir) | |
| mkdir -p $target | |
| cp $dir/{corpus.jsonl,metadata.json} $target | |
| done | |