sonjaaits / thesis-code

thesis-code

Setup

Setup using Conda (Anaconda / Miniconda)

It's best to create a custom environment first:

conda create -n ENV_NAME
conda activate ENV_NAME
conda install python==3.7

This will create an empty environment and install Python 3.7 together with the corresponding version of pip. We will then use that version of pip to install the requirements:

pip install -r requirements.txt

It's important to get this right, since BERT requires TensorFlow 1.15, which in its turn requires Python/pip 3.7 (not 3.8).

Understanding the pipeline

The pipeline consists of several steps, which need not all be rerun every time.

• Step 1 is to fetch and save the data: for this purpose either downloader or cord_loader is used.
  • For downloader, the input is a list of newline-separated PubMed IDs.
  • For cord_loader, the input is the metadata.csv file from inside the .tar.gz files in the CORD-19 Historical Releases (this seems unavailable for early releases).
• Step 2 is sentencer which processes the data further for use by the models.
• Step 3 is ner, named-entity recognition.
• Step 4 is re, relationship extraction.
• Optional step: metrics will create metrics such as F1-score for the NER model.
• Optional step: analysis will analyse the NER results to find co-occurrences.

Running the pipeline

Open the config.json file in the root directory and un-ignore the steps you want to run by setting them to false. Then, make sure that input and output file names align. Here's a nice little chart to help you understand (A-H are file names).

(A)———[downloader]———.                                       .——[analysis]———(E)
                      |———(C)———[sentencer]———(D)———[ner]———|
(B)———[cord_loader]——'                                       '—————[re]——————(F)

(G)———[metrics]———(H)  (independent)

Then run the script by doing: python main.py

Converting BioBERT (TensorFlow) to ONNX

First make sure to install tf2onnx:

pip install -U tf2onnx

Then convert your (exported) TensorFlow model:

python -m tf2onnx.convert --saved-model ./PATH_TO_MODEL_DIR/ --output ./OUT_PATH/model_name.onnx

Creating a symlink to a model

ln -s [absolute path to model] [path to link]

Download

BioBERT-Base fine-tuned ONNX-model with vocabulary - fine-tuned on BC5CDR-chem dataset