Steps for pre-trained REALM models

[hahazhar]

I have a question regarding the demo which was also mentioned in issue #67 . I have executed the following lines.

git clone https://github.com/google-research/language cd language

conda create --name orqa python=3.7 source activate orqa pip install -r language/orqa/requirements.txt

conda create --name orqa python=3.7 source activate orqa pip install -r language/orqa/requirements.txt

python -m language.orqa.utils.scann_utils_test

git clone https://github.com/facebookresearch/DrQA.git cd DrQA export DRQA_PATH=$(pwd) sh download.sh

After installing the gsutil, do I have to execute the following lines?

mkdir original_nq gsutil -m cp -R gs://natural_questions/v1.0 original_nq cd original_nq export ORIG_NQ_PATH=$(pwd)

The data is 45Gb and I do not have the space.
Or do i change it to "gsutil -m cp -R gs://realm-data/orqa_nq_model_from_realm" ?

or simply will suffice ?

python -m language.orqa.predict.orqa_demo \ --model_dir=gs://orqa-data/orqa_nq_model \ --port=8080