File size: 4,984 Bytes
5e1c670
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
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)