model.py

import torch
from transformers import BertModel


class LanguageIdentifier(torch.nn.Module):
    def __init__(self):
        super().__init__()

        self.portuguese_bert = BertModel.from_pretrained(
            "neuralmind/bert-large-portuguese-cased")

        self.linear_layer = torch.nn.Sequential(
            torch.nn.Dropout(p=0.2),
            torch.nn.Linear(self.portuguese_bert.config.hidden_size, 1),
        )

    def forward(self, input_ids, attention_mask):

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

        outputs = self.linear_layer(outputs)

        return outputs


class Ensembler(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    def forward(self, input_ids, attention_mask):
        outputs = []

        with torch.no_grad():
            for domain in ['politics', 'news', 'law', 'social_media', 'literature', 'web']:
                specialist = LanguageIdentifier()
                
                specialist.load_state_dict(torch.load(f"{domain}.pt", map_location=self.device))
                
                specialist.eval()

                specialist.to(self.device)

                outputs.append(specialist(input_ids, attention_mask))

                # Remove the specialist from the GPU
                specialist.cpu()
                del specialist
                torch.cuda.empty_cache()
                
                
            outputs = torch.cat(outputs, dim=1)

            return torch.mean(outputs, dim=1).unsqueeze(1)