Некорректное поведение при обучении нейронной сети PyTorch

Question

Необходимо написать нейронку, которая будет распознавать цифру на изображении. В настоящее время выглядит так. что чем больше ошибка. тем больше точность. Почему так - загадка. Исходный размер: 276х144.

class DigitRecognition(nn.Module):
    def __init__(self):
        super().__init__()
        self.__forward_propagation = nn.Sequential(
                nn.Conv2d(in_channels=3, out_channels=8, kernel_size=1, stride=1, padding=1),
                nn.ReLU(),
                nn.MaxPool2d(kernel_size=2, stride=2),
                
                nn.Flatten(),

                nn.Linear(in_features=280, out_features=32),
                nn.Linear(in_features=32, out_features=16),
                nn.Linear(in_features=16, out_features=10),

                nn.Sigmoid(),
            )
        
    def forward(self, X) -> torch.Tensor:
        X = self.__forward_propagation(X)
        return X

    @staticmethod
    def accuracy_fn(y_true, y_pred) -> float:
        correct: int = 0
        for pred, true in zip(y_pred, y_true):
            if torch.argmax(y_true[0]).item() == torch.argmax(y_pred[0]).item():
                correct += 1
        acc: float = correct / len(y_true) *100
        return acc


EPOCHS: int = 100
lr: float = 0.0001

mean_train_error: float = 0.
mean_test_error: float = 0.

train_errors: float = 0.
test_errors: float = 0.

train_acc_errors: float = 0.
test_acc_errors: float = 0.

train_acc: float = 0.
test_acc: float = 0.


f: DigitRecognition = DigitRecognition()
loss_fn = nn.CrossEntropyLoss()

optim = torch.optim.Adam(f.parameters(), lr)

f.to(DEVICE)


for EPOCH in range(EPOCHS):
    f.train()
    for batch in train_dataloader:

        X: torch.Tensor = batch[0].to(DEVICE)
        y_pred = f(X).to(DEVICE)

        y_true_digit: torch.Tensor = batch[1]
        y_true_canvas: torch.Tensor = torch.zeros(y_pred.shape[0], 10)
        y_true: torch.Tensor = y_true_canvas.scatter_(1, y_true_digit.unsqueeze(1), 1.)

        y_true: torch.Tensor = y_true.to(DEVICE)

        error: float = loss_fn(y_pred, y_true)
        train_acc = f.accuracy_fn(y_true, y_pred)

        train_errors += error
        train_acc_errors += train_acc

        error.backward()

        optim.zero_grad()
        optim.step()

    f.eval()
    with torch.inference_mode():
        for batch in test_dataloader:

            X: torch.Tensor = batch[0].to(DEVICE)
            y_pred = f(X).to(DEVICE)

            y_true_digit: torch.Tensor = batch[1]
            y_true_canvas: torch.Tensor = torch.zeros(y_pred.shape[0], 10)
            y_true: torch.Tensor = y_true_canvas.scatter_(1, y_true_digit.unsqueeze(1), 1.)

            y_true: torch.Tensor = y_true.to(DEVICE)

            error: float = loss_fn(y_pred, y_true)
            test_acc = f.accuracy_fn(y_true, y_pred)

            test_errors += error
            test_acc_errors += test_acc
   

    mean_train_error: float = train_errors / BATCH_SIZE
    mean_test_error: float = test_errors / BATCH_SIZE

    mean_train_acc_errors: float = train_acc_errors / BATCH_SIZE
    mean_test_acc_errors: float = test_acc_errors / BATCH_SIZE

    print(f"epoch: {EPOCH}/Train. Error {mean_train_error}/Accuracy: {mean_train_acc_errors}/t ** Test. Error: {mean_test_error}/Accuracy: {mean_test_acc_errors}")

    mean_train_error: float = 0.
    mean_test_error: float = 0.

    train_errors: float = 0.
    test_errors: float = 0.

    train_acc_errors: float = 0.
    test_acc_errors: float = 0.

    mean_train_acc_errors: float = 0.
    mean_test_acc_errors: float = 0.

Google Colab Notebook

файлы: https://drive.google.com/drive/folders/12nhWJo73gHUXWa0vg38cCOlz_aKHesKN?usp=sharing

пример вывода:

смущает, что чем ниже ошибка, тем ниже точность, должно ведь быть наоборот?

БЛОГ НА HUSL

Некорректное поведение при обучении нейронной сети PyTorch

Ответы (0 шт):