HealthLearning

Running Project on HPC with Jupyter Notebook

# Login to greene
ssh greene

# Switch to Log-4
ssh log-4

# Request CPU/GPU node
srun --account=csci_ga_2565_0001 --partition=interactive --pty /bin/bash (CPU)
                                            OR
srun --gres=gpu:1 --account=csci_ga_2565_0001 --partition=n1s8-t4-1 --time=04:00:00 --pty /bin/bash (GPU)

# Load singularity image and enviorment
singularity exec --nv --bind /scratch --bind /share/apps --overlay /share/apps/pytorch/1.8.1/pytorch-1.8.1.sqf:ro /share/apps/images/cuda11.1.1-cudnn8-devel-ubuntu20.04.sif /bin/bash

source /ext3/env.sh

# Assign a port
port=$(shuf -i 10000-65500 -n 1)

# Remember this port
echo $port

# Remote forward this port to greene
ssh -N -f -R $port:localhost:$port log-1.hpc.nyu.edu
ssh -N -f -R $port:localhost:$port log-2.hpc.nyu.edu
ssh -N -f -R $port:localhost:$port log-3.hpc.nyu.edu

# Start notebook on this port
jupyter notebook --no-browser --port $port --notebook-dir=$(pwd)

# On a separate terminal window
ssh -L <port>:localhost:<port> <net_id>@greene

---

Language Modelling

  sbatch albert_train.sbatch

Usage

from transforms import ClinicalSynonymSubstitution,Compose

SCI_SM_transform=ClinicalSynonymSubstitution(substitution_probability=0.4,p=0.3,scispacy_entity_model="en_core_sci_sm")
BIONLP13CG_transform=ClinicalSynonymSubstitution(substitution_probability=0.7,p=0.7,scispacy_entity_model="en_ner_bionlp13cg_md")
BC5CDR_transform=ClinicalSynonymSubstitution(substitution_probability=0.7,p=0.7,scispacy_entity_model="en_ner_bc5cdr_md")

composite_transform=Compose(
        [
            SCI_SM_transform,
            BIONLP13CG_transform,
            BC5CDR_transform
        ]
)


composite_transform=Compose(
        [
            SCI_LG_transform, # For linking 2.78M unique concepts from UMLS, but not accurate yet, synonym candidates are generated without taking context information into account.
            BIONLP13CG_transform, # For cancer and genetics
            BC5CDR_transform # For chemicals and diseases
        ]
)

text="Patient has elevated BUN"
composite_transform(text)

Model Weights

from transformers import AutoTokenizer, AutoModelForMaskedLM
  
tokenizer = AutoTokenizer.from_pretrained("Vasudev/discharge_albert")

model = AutoModelForMaskedLM.from_pretrained("Vasudev/discharge_albert")