model_utils.py

import pandas as pd
import torch
import pickle
import torch.nn as nn
from sklearn.preprocessing import LabelEncoder
from sklearn.model_selection import train_test_split
from transformers import BertTokenizer, BertModel
from torch.optim import AdamW
from tqdm import tqdm

TEXT_COLUMN = 'Sanction_Context'
LABEL_COLUMNS = [
    'Red_Flag_Reason', 'Maker_Action', 'Escalation_Level',
    'Risk_Category', 'Risk_Drivers', 'Investigation_Outcome'
]

PRETRAINED_MODEL_NAME = 'bert-base-uncased'
MAX_LEN = 128
BATCH_SIZE = 16
EPOCHS = 1
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")

class BertMultiOutput(nn.Module):
    def __init__(self, num_labels_per_output):
        super().__init__()
        self.bert = BertModel.from_pretrained(PRETRAINED_MODEL_NAME)
        self.dropout = nn.Dropout(0.3)
        self.classifiers = nn.ModuleList([
            nn.Linear(self.bert.config.hidden_size, n_labels)
            for n_labels in num_labels_per_output
        ])
    def forward(self, input_ids, attention_mask):
        outputs = self.bert(input_ids=input_ids, attention_mask=attention_mask)
        pooled_output = self.dropout(outputs.pooler_output)
        return [classifier(pooled_output) for classifier in self.classifiers]

def train_and_save_model(csv_path, output_path='app/bert_model.pkl'):
    df = pd.read_csv(csv_path)
    X = df[TEXT_COLUMN].tolist()
    y = df[LABEL_COLUMNS]

    label_encoders = {}
    y_encoded = pd.DataFrame()
    for col in LABEL_COLUMNS:
        le = LabelEncoder()
        y_encoded[col] = le.fit_transform(y[col])
        label_encoders[col] = le

    X_train, _, y_train, _ = train_test_split(X, y_encoded, test_size=0.2, random_state=42)
    tokenizer = BertTokenizer.from_pretrained(PRETRAINED_MODEL_NAME)

    def tokenize_texts(texts):
        return tokenizer(texts, padding=True, truncation=True, max_length=MAX_LEN, return_tensors="pt")

    train_encodings = tokenize_texts(X_train)
    labels = [torch.tensor(y_train[col].values) for col in LABEL_COLUMNS]

    num_labels_list = [len(le.classes_) for le in label_encoders.values()]
    model = BertMultiOutput(num_labels_list).to(DEVICE)
    optimizer = AdamW(model.parameters(), lr=2e-5)
    loss_fn = nn.CrossEntropyLoss()

    model.train()
    for epoch in range(EPOCHS):
        for i in tqdm(range(0, len(X_train), BATCH_SIZE)):
            input_ids = train_encodings['input_ids'][i:i+BATCH_SIZE].to(DEVICE)
            attention_mask = train_encodings['attention_mask'][i:i+BATCH_SIZE].to(DEVICE)
            batch_labels = [label[i:i+BATCH_SIZE].to(DEVICE) for label in labels]

            optimizer.zero_grad()
            outputs = model(input_ids, attention_mask)
            loss = sum([loss_fn(o, l) for o, l in zip(outputs, batch_labels)])
            loss.backward()
            optimizer.step()

    model_bundle = {
        'model_state_dict': model.state_dict(),
        'tokenizer': tokenizer,
        'label_encoders': label_encoders
    }
    with open(output_path, 'wb') as f:
        pickle.dump(model_bundle, f)

def load_model(path='app/bert_model.pkl'):
    with open(path, 'rb') as f:
        bundle = pickle.load(f)
    tokenizer = bundle['tokenizer']
    label_encoders = bundle['label_encoders']
    num_labels_list = [len(le.classes_) for le in label_encoders.values()]
    model = BertMultiOutput(num_labels_list).to(DEVICE)
    model.load_state_dict(bundle['model_state_dict'])
    model.eval()
    return model, tokenizer, label_encoders