src/train.py

from datasets import load_dataset
import pandas as pd
import numpy as np
import os
import json
import torch
from torch.utils.data import Dataset, DataLoader
from transformers import DistilBertTokenizerFast, DistilBertForSequenceClassification
from transformers import Trainer, TrainingArguments, AdamW

model_name = "distilbert-base-uncased"

class USPTODataset(Dataset):
    def __init__(self, encodings, labels):
        self.encodings = encodings
        self.labels = labels
    def __getitem__(self, idx):
        item = {key: torch.tensor(val[idx]) for key, val in self.encodings.items()}
        item['labels'] = torch.tensor(self.labels[idx])
        return item
    def __len__(self):
        return len(self.labels)

def LoadDataset():
    print("=== LOADING THE DATASET ===")
    # Extracting the dataset, filtering only for Jan. 2016
    dataset_dict = load_dataset('HUPD/hupd',
        name='sample',
        data_files="https://huggingface.co/datasets/HUPD/hupd/blob/main/hupd_metadata_2022-02-22.feather", 
        icpr_label=None,
        train_filing_start_date='2016-01-01',
        train_filing_end_date='2016-01-21',
        val_filing_start_date='2016-01-22',
        val_filing_end_date='2016-01-31',
    )

    print("Separating between training and validation data")
    df_train = pd.DataFrame(dataset_dict['train'] )
    df_val = pd.DataFrame(dataset_dict['validation'] )

    
    print("=== PRE-PROCESSING THE DATASET ===")
    #We are interested in the following columns:
    # - Abstract
    # - Claims
    # - Decision <- our `y`
    # Let's preprocess them both out of our training and validation data
    # Also, consider that the "Decision" column has three types of values: "Accepted", "Rejected", and "Pending". To remove unecessary baggage, we will be only looking for "Accepted" and "Rejected".

    necessary_columns = ["abstract","claims","decision"]
    output_values = ['ACCEPTED','REJECTED']

    print("Dropping unused columns")
    trainFeaturesToDrop = [col for col in list(df_train.columns) if col not in necessary_columns]
    trainDF = df_train.dropna()
    trainDF.drop(columns=trainFeaturesToDrop, inplace=True)
    trainDF = trainDF[trainDF['decision'].isin(output_values)]
    valFeaturesToDrop = [col for col in list(df_val.columns) if col not in necessary_columns]
    valDF = df_val.dropna()
    valDF.drop(columns=valFeaturesToDrop, inplace=True)
    valDF = valDF[valDF['decision'].isin(output_values)]

    # We need to replace the values in the `decision` column to numerical representations. ]
    # We will set "ACCEPTED" as `1` and "REJECTED" as `0`.
    print("Replacing values in `decision` column")
    yKey = {"ACCEPTED":1,"REJECTED":0}
    trainDF2 = trainDF.replace({"decision": yKey})
    valDF2 = valDF.replace({"decision": yKey})

    # We combine the `abstract` and `claims` columns into a single `text` column. 
    # We also re-label the `decision` column to `label`.
    print("Combining columns and renaming `decision` to `label`")
    trainDF3 = trainDF2.rename(columns={'decision': 'label'})
    trainDF3['text'] = trainDF3['abstract'] + ' ' + trainDF3['claims']
    trainDF3.drop(columns=["abstract","claims"],inplace=True)

    valDF3 = valDF2.rename(columns={'decision': 'label'})
    valDF3['text'] = valDF3['abstract'] + ' ' + valDF3['claims']
    valDF3.drop(columns=["abstract","claims"],inplace=True)

    # We can grab the data for each column so that we have a list of values for training labels, 
    #   training texts, validation labels, and validation texts.
    print("Extracting label and text data from dataframes")
    trainData = {
        "labels":trainDF3["label"].tolist(),
        "text":trainDF3["text"].tolist()
    }
    valData = {
        "labels":valDF3["label"].tolist(),
        "text":valDF3["text"].tolist()
    }
    print(f'TRAINING:\t# labels: {len(trainData["labels"])}\t# texts: {len(trainData["text"])}')
    print(f'VALID:\t# labels: {len(valData["labels"])}\t# texts: {len(valData["text"])}')

    if not os.path.exists("./data"):
        os.makedirs('./data')

    with open("./data/train.json", "w") as outfile:
        json.dump(trainData, outfile, indent=2)
    with open("./data/val.json", "w") as outfile:
        json.dump(valData, outfile, indent=2)

    return trainData, valData

def main():
    trainDataPath = "./data/train.json"
    valDataPath = "./data/val.json"
    trainData = None
    valData = None

    if os.path.exists(trainDataPath) and os.path.exists(valDataPath):
        ftrain = open(trainDataPath)
        trainData = json.load(ftrain)
        ftrain.close()
        fval = open(valDataPath)
        valData = json.load(fval)
        fval.close()
    else:
        trainData, valData = LoadDataset()
    
    print(len(trainData["labels"]), len(trainData["text"]), len(valData["labels"]), len(valData["text"]))

    tokenizer = DistilBertTokenizerFast.from_pretrained(model_name)
    train_encodings = tokenizer(trainData["text"], truncation=True, padding=True)
    val_encodings = tokenizer(valData["text"], truncation=True, padding=True)

    train_dataset = USPTODataset(train_encodings, trainData["labels"])
    val_dataset = USPTODataset(val_encodings, valData["labels"])

    train_args = TrainingArguments(
        output_dir="./results",
        num_train_epochs=2,
        per_device_train_batch_size=16,
        per_device_eval_batch_size=64,
        warmup_steps=500,
        learning_rate=5e-5,
        weight_decay=0.01,
        logging_dir="./logs",
        logging_steps=10
    )

    model = DistilBertForSequenceClassification.from_pretrained(model_name)
    trainer = Trainer(
        model=model,
        args=train_args,
        train_dataset=train_dataset,
        eval_dataset=val_dataset
    )
    trainer.train()

if __name__ == "__main__":
    main()