Skip to content

Latest commit

 

History

History
89 lines (63 loc) · 2.68 KB

notes.md

File metadata and controls

89 lines (63 loc) · 2.68 KB

SBATCH

to request how much memory you want for a100 machine #SBATCH --constraint=a100_80gb

request a node foe debug ijob -A uva-dsa -p gpu --gres=gpu:a6000:4 -c 16 -t 70:00:00

check the long slurm name squeue --format="%.18i %.9P %.30j %.8u %.8T %.10M %.9l %.6D %R" --me

Heterogeneous_graph.py

  • bertEmbedding:
  • HeteroGraph: GNN architectures that contains knowledge information from protocol information

data_utils.py

  • EMSDataPipeline: wrapper class to tokenize narratives and create data loader for training/testing
    • .build_single_dataloader() return a dataloader that streams samples for training/testing
  • EMSDataset: wrapper class that tokenizes narrative and meta knowledge text. Used in EMSDataPipeline to create a data_loader object.

model.py

  • HGT (Heterogeneous Graph Transformer) that operates on graph object passed as argument

  • EMSMultiModel: This class integrates the label-wise attention with the BERT model. Below I have summarized the initialization fields (to the best of my understanding):

    • Backbone: backbone for BERT model to process narrative text
    • dp: dropout (not used anymore)
    • max_len: max sentence length (not used anymore, I think)
    • attn: specifies what type of attention to use in knowledge fusion
      • 'qkv-la': label-wise attention qkv
      • 'la': label-wise attention
      • 'sa': self-attention
    • cluster: whether the model is outputting grouped or ungrouped protocols
    • cls: whether or not the output of fusion through a fully connected layer (does this stand for classifier?)
    • graph: the GNN architecture (such as Heterograph) used for knowledge

  • classifier: MLP with 3 hidden layers. I believe this inputs the embeddings after they have been attended to

loss_fn.py

  • SimCLR
  • ResampleLoss
  • FocalLoss

eval_metrics.py

  • evaluate model performance and generate report json file

bad_case_analysis.py

  • an interface for evaluate bad cases

psuedo_label_generator.py

  • dataAug
  • knowledgeModel
  • EMSBertModel
  • comb_pipeline
  • dm_pipeline

counterfactual_reasoning.py

  • Doing counterfactual reasoning

trainer

  • trainer
  • tester
  • validator
  • pseudolabeler

visualize

  • Desription: a bunch of graphing code

TF-IDF.ipynb

  • generate a vector representation for protocols by calculate TF-IDF of signs & symptoms

Metamap Extractor.ipynb

  • extract signs & symptoms by using Metamap

GT-Label Extraction.ipynb

  • extract gt labels from RAA data
  • split into 3-folds train-val-test

Model Explanation.ipynb

  • explain model's prediction by using Integrated Gradient

N-gram for Text Classification.ipynb

  • use one-gram as features and train machine learning models
  • used as baseline