notebooks/daedra.py

import pandas as pd
import numpy as np
import torch
import os
from typing import List, Union
from transformers import AutoTokenizer, Trainer, AutoModelForSequenceClassification, TrainingArguments, DataCollatorWithPadding, pipeline, AutoModel
from datasets import load_dataset, Dataset, DatasetDict
import shap
import wandb
import evaluate
import logging

os.environ["TOKENIZERS_PARALLELISM"] = "false"

device: str = 'cuda' if torch.cuda.is_available() else 'cpu'

SEED: int = 42

BATCH_SIZE: int = 16
EPOCHS: int = 3
SUBSAMPLING: float = 0.1

# WandB configuration
os.environ["WANDB_PROJECT"] = "DAEDRA multiclass model training" 
os.environ["WANDB_LOG_MODEL"] = "checkpoint"  # log all model checkpoints
os.environ["WANDB_NOTEBOOK_NAME"] = "DAEDRA.ipynb"

dataset = load_dataset("chrisvoncsefalvay/vaers-outcomes")

if SUBSAMPLING < 1:
    _ = DatasetDict()
    for each in dataset.keys():
        _[each] = dataset[each].shuffle(seed=SEED).select(range(int(len(dataset[each]) * SUBSAMPLING)))

    dataset = _

accuracy = evaluate.load("accuracy")
precision, recall = evaluate.load("precision"), evaluate.load("recall")
f1 = evaluate.load("f1")

def compute_metrics(eval_pred):
    predictions, labels = eval_pred
    predictions = np.argmax(predictions, axis=1)
    return {
        'accuracy': accuracy.compute(predictions=predictions, references=labels)["accuracy"],
        'precision_macroaverage': precision.compute(predictions=predictions, references=labels, average='macro')["precision"],
        'precision_microaverage': precision.compute(predictions=predictions, references=labels, average='micro')["precision"],
        'recall_macroaverage': recall.compute(predictions=predictions, references=labels, average='macro')["recall"],
        'recall_microaverage': recall.compute(predictions=predictions, references=labels, average='micro')["recall"],
        'f1_microaverage': f1.compute(predictions=predictions, references=labels, average='micro')["f1"]
    }

label_map = {i: label for i, label in enumerate(dataset["test"].features["label"].names)}

def train_from_model(model_ckpt: str, push: bool = False):
    print(f"Initialising training based on {model_ckpt}...")

    print("Tokenising...")
    tokenizer = AutoTokenizer.from_pretrained(model_ckpt)

    cols = dataset["train"].column_names
    cols.remove("label")
    ds_enc = dataset.map(lambda x: tokenizer(x["text"], truncation=True, max_length=512), batched=True, remove_columns=cols)

    print("Loading model...")
    try:
        model = AutoModelForSequenceClassification.from_pretrained(model_ckpt, 
                                                                    num_labels=len(dataset["test"].features["label"].names), 
                                                                    id2label=label_map, 
                                                                    label2id={v:k for k,v in label_map.items()})
    except OSError:
        model = AutoModelForSequenceClassification.from_pretrained(model_ckpt, 
                                                            num_labels=len(dataset["test"].features["label"].names), 
                                                            id2label=label_map, 
                                                            label2id={v:k for k,v in label_map.items()},
                                                            from_tf=True)


    args = TrainingArguments(
        output_dir="vaers",
        evaluation_strategy="steps",
        eval_steps=100,
        save_strategy="epoch",
        learning_rate=2e-5,
        per_device_train_batch_size=BATCH_SIZE,
        per_device_eval_batch_size=BATCH_SIZE,
        num_train_epochs=EPOCHS,
        weight_decay=.01,
        logging_steps=1,
        run_name=f"daedra-minisample-comparison-{SUBSAMPLING}",
        report_to=["wandb"])

    trainer = Trainer(
            model=model,
            args=args,
            train_dataset=ds_enc["train"],
            eval_dataset=ds_enc["test"],
            tokenizer=tokenizer,
            compute_metrics=compute_metrics)
    
    if SUBSAMPLING != 1.0:
        wandb_tag: List[str] = [f"subsample-{SUBSAMPLING}"]
    else:
        wandb_tag: List[str] = [f"full_sample"]

    wandb_tag.append(f"batch_size-{BATCH_SIZE}")
    wandb_tag.append(f"base:{model_ckpt}")
        
    if "/" in model_ckpt:
        sanitised_model_name = model_ckpt.split("/")[1]
    else:
        sanitised_model_name = model_ckpt

    wandb.init(name=f"daedra_{SUBSAMPLING}-{sanitised_model_name}", tags=wandb_tag, magic=True)

    print("Starting training...")

    trainer.train()

    print("Training finished.")

    wandb.finish()

if __name__ == "__main__":
    wandb.finish()

    for mname in (
        #"dmis-lab/biobert-base-cased-v1.2",
        "emilyalsentzer/Bio_ClinicalBERT",
        "bert-base-uncased",
        "distilbert-base-uncased"
    ):
        print(f"Now training on subsample with {mname}...")
        train_from_model(mname)