embeddings.py

import torch
import logging
from pathlib import Path
import os
from transformers import BertModel, BertForTokenClassification
import pandas as pd
import evaluate
from datasets import load_dataset
from transformers import BertTokenizerFast
from torch.utils.data import DataLoader
from tqdm import tqdm


def tokenize(dataset):
    BERT_MAX_LEN = 512

    tokenizer = BertTokenizerFast.from_pretrained(
        "neuralmind/bert-base-portuguese-cased", max_length=BERT_MAX_LEN)

    dataset = dataset.map(lambda example: tokenizer(
        example["text"], truncation=True, padding="max_length", max_length=BERT_MAX_LEN))

    return dataset


def create_dataloader(dataset, shuffle=True):
    return DataLoader(dataset, batch_size=8, shuffle=shuffle, num_workers=8, drop_last=True)


CURRENT_PATH = Path(__file__).parent

logging.basicConfig(level=logging.INFO,
                    format='%(asctime)s %(levelname)s %(message)s', datefmt='%H:%M:%S', filename=os.path.join(CURRENT_PATH, 'out', 'debug_embeddings.txt'), filemode='w')


class LanguageIdentifer(torch.nn.Module):
    def __init__(self, mode='horizontal_stacking', pos_layers_to_freeze=0, bertimbau_layers_to_freeze=0):
        super().__init__()

        self.labels = ['pt-PT', 'pt-BR']

        self.portuguese_model = BertModel.from_pretrained(
            "neuralmind/bert-base-portuguese-cased")

        self.portuguese_pos_tagging_model = BertForTokenClassification.from_pretrained(
            "lisaterumi/postagger-portuguese")

        for layer in range(bertimbau_layers_to_freeze):
            for name, param in self.portuguese_model.named_parameters():
                if f".{layer}" in name:
                    print(f"Freezing Layer {name} of Bertimbau")
                    param.requires_grad = False

        for layer in range(pos_layers_to_freeze):
            for name, param in self.portuguese_pos_tagging_model.named_parameters():
                if f".{layer}" in name:
                    print(f"Freezing Layer {name} of POS")
                    param.requires_grad = False

        self.portuguese_pos_tagging_model.classifier = torch.nn.Identity()
        self.mode = mode

        if self.mode == 'horizontal_stacking':
            self.linear = self.common_network(torch.nn.Linear(
                self.portuguese_pos_tagging_model.config.hidden_size + self.portuguese_model.config.hidden_size, 512))
        elif self.mode == 'bertimbau_only' or self.mode == 'pos_only' or self.mode == 'vertical_sum':
            self.linear = self.common_network(torch.nn.Linear(
                self.portuguese_model.config.hidden_size, 512))
        else:
            raise NotImplementedError

    def common_network(self, custom_linear):
        return torch.nn.Sequential(
            custom_linear,
            torch.nn.ReLU(),
            torch.nn.Dropout(0.2),
            torch.nn.Linear(512, 1),
        )

    def forward(self, input_ids, attention_mask):

        #(Batch_Size,Sequence Length, Hidden_Size)
        outputs_bert = self.portuguese_model(
            input_ids=input_ids, attention_mask=attention_mask).last_hidden_state[:, 0, :]

        #(Batch_Size,Sequence Length, Hidden_Size)
        outputs_pos = self.portuguese_pos_tagging_model(
            input_ids=input_ids, attention_mask=attention_mask).logits[:, 0, :]

        if self.mode == 'horizontal_stacking':
            outputs = torch.cat((outputs_bert, outputs_pos), dim=1)
        elif self.mode == 'bertimbau_only':
            outputs = outputs_bert
        elif self.mode == 'pos_only':
            outputs = outputs_pos
        elif self.mode == 'vertical_sum':
            outputs = outputs_bert + outputs_pos
            outputs = torch.nn.functional.normalize(outputs, p=2, dim=1)

        return self.linear(outputs)


def load_models():
    models = []

    for domain in ['politics', 'news', 'law', 'social_media', 'literature', 'web']:
        logging.info(f"Loading {domain} model...")

        model = LanguageIdentifer(mode='pos_only')
        model.load_state_dict(torch.load(os.path.join(
            CURRENT_PATH, 'models', 'embeddings', f'{domain}.pt')))
        
        models.append({
            'model': model,
            'train_domain': domain,
        })

    return models


def benchmark(model, debug=False):

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

    df_result = pd.DataFrame(
        columns=['train_domain', 'test_domain', 'accuracy', 'f1', 'precision', 'recall'])

    train_domain = model['train_domain']

    model = model['model']

    model.to(device)

    model.eval()

    for test_domain in ['politics', 'news', 'law', 'social_media', 'literature', 'web']:
        dataset = load_dataset(
            'arubenruben/Portuguese_Language_Identification', test_domain, split='test')

        if debug:
            logging.info("Debug mode: using only 100 samples")
            dataset = dataset.shuffle().select(range(100))
        else:
            dataset = dataset.shuffle().select(range(min(50_000, len(dataset))))

        dataset = tokenize(dataset)
        dataset = create_dataloader(dataset)

        y = []

        with torch.no_grad():
            for batch in tqdm(dataset):
                input_ids = batch['input_ids'].to(device)
                attention_mask = batch['attention_mask'].to(device)

                y.extend(model(input_ids, attention_mask).cpu().detach().numpy())

        y = [1 if y_ > 0.5 else 0 for y_ in y]

        accuracy = evaluate.load('accuracy').compute(
            predictions=y, references=dataset['label'])['accuracy']
        f1 = evaluate.load('f1').compute(
            predictions=y, references=dataset['label'])['f1']
        precision = evaluate.load('precision').compute(
            predictions=y, references=dataset['label'])['precision']
        recall = evaluate.load('recall').compute(
            predictions=y, references=dataset['label'])['recall']

        df_result = pd.concat([df_result, pd.DataFrame({
            'train_domain': [train_domain],
            'test_domain': [test_domain],
            'accuracy': [accuracy],
            'f1': [f1],
            'precision': [precision],
            'recall': [recall],
        })], ignore_index=True)

    return df_result


def test():
    DEBUG = False
    
    models = load_models()
    
    df_results = pd.DataFrame(
        columns=['train_domain', 'test_domain', 'accuracy', 'f1', 'precision', 'recall'])

    for model in models:
        logging.info(f"Train Domain {model['train_domain']}...")

        df_results = pd.concat(
            [df_results, benchmark(model, debug=DEBUG)], ignore_index=True)

        logging.info("Saving Results...")

        df_results.to_json(os.path.join(CURRENT_PATH, 'out', 'embeddings.json'),
                           orient='records', indent=4, force_ascii=False)