Additional setup steps required to get everything working

[jakpra]

Most of the information about how to setup inference and training runs is present in the readme, but there are a few gaps.
Here are the additional steps I had to take in order to make inference with a pretrained model work:

• pip install -r requirements.txt
• pip install pytorch_pretrained_bert
• pip install pycorenlp
• install CoreNLP
  • download from https://stanfordnlp.github.io/CoreNLP/download.html and follow the setup steps (add paths to CLASSPATH)
  • adjust classpath argument in run_standford_corenlp_server.sh (also there is a typo in the filename of this script)
• sh download_artifacts.sh
  • this is listed as a step for training a model, but it's also necessary if you just want to run inference
  • before running it for the first time, uncomment the commented lines (though I don't think GloVe is needed)
  • after running it, rename the config file to just config.json and the vocab file to just vocab.txt in bert-base-cased/
  • make sure bert-base-cased/ is one level above from the AMR-gs top-level directory; i.e., from within AMR-gs, the relative path should be ../bert-base-cased/ (unfortunately this relative path is currently expected by the pretrained model)
• make sure that all (train/dev/test) data is reachable from AMR-gs via data/AMR/amr_2.0 (or data/AMR/amr_1.0)
  • this will be done automatically if you run prepare_data.sh on one of the LDC AMR corpora, but if you use custom datasets you may have to do it manually, e.g., with symbolic links: ln -s actual/path/to/data path-to-AMR-gs/data/AMR/amr_2.0
• make sure the pretrained vocabulary files at, e.g., amr2.0.bert.gr/vocabs/ are reachable from AMR-gs at data/AMR/amr_2.0_reca/, e.g., with symbolic links ln -s amr2.0.bert.gr/vocabs/ path-to-AMR-gs/data/AMR/amr_2.0_reca/
• the machine learning code in parser/ expects a GPU/CUDA setup; if you want to run on CPU (but also in general) it is preferable to use the version of the files fixed in c795b88
• NOW you can proceed to follow the instructions in the readme for AMR Parsing with Pretrained Models