tjmg.py

import torch
import torch.nn as nn
import torch.nn.functional as F
from torchvision import transforms
from PIL import Image
from safetensors.torch import load_file
from huggingface_hub import hf_hub_download

# Definição da arquitetura da rede neural
class CaptchaCNN(nn.Module):
    def __init__(self, input_dim, output_ndigits, output_vocab_size, dropout=(0.25, 0.5), dense_units=200, vocab=None):
        super().__init__()
        self.input_dim = input_dim
        self.output_ndigits = output_ndigits
        self.output_vocab_size = output_vocab_size
        self.vocab = vocab

        self.batchnorm0 = nn.BatchNorm2d(3)
        self.conv1 = nn.Conv2d(3, 32, kernel_size=3)
        self.batchnorm1 = nn.BatchNorm2d(32)
        self.conv2 = nn.Conv2d(32, 64, kernel_size=3)
        self.batchnorm2 = nn.BatchNorm2d(64)
        self.conv3 = nn.Conv2d(64, 64, kernel_size=3)
        self.batchnorm3 = nn.BatchNorm2d(64)
        self.dropout1 = nn.Dropout(dropout[0])
        self.dropout2 = nn.Dropout(dropout[1])

        # Cálculo das dimensões após as camadas convolucionais
        def calc_dim(x):
            for _ in range(3):
                x = (x - 2) // 2
            return x

        conv_h = calc_dim(input_dim[0])
        conv_w = calc_dim(input_dim[1])
        fc1_in_features = conv_h * conv_w * 64

        self.fc1 = nn.Linear(fc1_in_features, dense_units)
        self.batchnorm_dense = nn.BatchNorm1d(dense_units)
        self.fc2 = nn.Linear(dense_units, output_vocab_size * output_ndigits)

    def forward(self, x):
        x = self.batchnorm0(x)
        x = F.relu(self.batchnorm1(F.max_pool2d(self.conv1(x), 2)))
        x = F.relu(self.batchnorm2(F.max_pool2d(self.conv2(x), 2)))
        x = F.relu(self.batchnorm3(F.max_pool2d(self.conv3(x), 2)))

        x = torch.flatten(x, start_dim=1)
        x = self.dropout1(x)
        x = F.relu(self.batchnorm_dense(self.fc1(x)))
        x = self.dropout2(x)
        x = self.fc2(x)
        x = x.view(-1, self.output_ndigits, self.output_vocab_size)
        return x


# Classe principal para carregar o modelo e fazer previsões
class TJMG:
    def __init__(self, repo_id="julio/captcha", filename="captcha_model.safetensors"):
        # Hiperparâmetros do modelo
        self.input_dim = (40, 110)
        self.output_vocab_size = 10
        self.vocab = [str(x) for x in range(10)]
        self.output_ndigits = 5
        self.dropout = (0.25, 0.5)
        self.dense_units = 200

        # Baixar o modelo do Hugging Face
        model_path = hf_hub_download(repo_id=repo_id, filename=filename)

        # Inicializar o modelo
        self.model = CaptchaCNN(
            input_dim=self.input_dim,
            output_ndigits=self.output_ndigits,
            output_vocab_size=self.output_vocab_size,
            dropout=self.dropout,
            dense_units=self.dense_units,
            vocab=self.vocab
        )
        
        # Carregar os pesos do modelo
        state_dict = load_file(model_path)
        self.model.load_state_dict(state_dict)
        self.model.eval()

        # Transformação da imagem
        self.transform = transforms.Compose([
            transforms.Resize(self.input_dim),
            transforms.ToTensor(),
        ])

    def predict(self, image_path):
        """
        Faz a previsão de um CAPTCHA.
        Args:
            image_path (str): Caminho da imagem do CAPTCHA.
        Returns:
            str: Texto previsto para o CAPTCHA.
        """
        image = Image.open(image_path).convert('RGB')
        image = self.transform(image).unsqueeze(0)

        with torch.no_grad():
            logits = self.model(image)

        preds = torch.argmax(logits, dim=2).squeeze().tolist()
        predicted_label = ''.join([self.vocab[i] for i in preds])
        return predicted_label