Inicializar pesos com **números inteiros aleatórios** não é recomendado na prática; esta abordagem é utilizada aqui apenas para fins desta tarefa.

Nota

import unittest
import importlib
import torch


def _dynamic(test, cond, ok, fail):
    if cond:
        test._testMethodName = ok
        test.assertTrue(True)
    else:
        test._testMethodName = fail
        test.fail(fail)


class TestUserCode(unittest.TestCase):

    # -------- manual seed --------
    def test_seed_set(self):
        with open("user_code.py") as f:
            src = f.read()

        cond = "torch.manual_seed" in src

        _dynamic(
            self,
            cond,
            "The manual seed is set correctly.",
            "You must set a manual seed using `torch.manual_seed()` before creating any tensors."
        )

    # -------- weights_layer1 --------
    def test_weights1_exists(self):
        import user_code
        cond = hasattr(user_code, "weights_layer1")
        _dynamic(
            self,
            cond,
            "`weights_layer1` is successfully defined.",
            "`weights_layer1` is missing. Assign your 3x4 uniform-random tensor to this variable."
        )

    def test_weights1_valid(self):
        import user_code
        try:
            t = user_code.weights_layer1
            cond = (
                isinstance(t, torch.Tensor)
                and t.shape == (3, 4)
                and torch.all(t >= 0)
                and torch.all(t < 1)
            )
        except:
            cond = False

        _dynamic(
            self,
            cond,
            "`weights_layer1` is a valid 3x4 tensor with values uniformly sampled from [0, 1).",
            "`weights_layer1` must be created with `torch.rand(3, 4)` so that all values lie between 0 and 1."
        )

    # -------- biases_layer1 --------
    def test_biases1_exists(self):
        import user_code
        cond = hasattr(user_code, "biases_layer1")
        _dynamic(
            self,
            cond,
            "`biases_layer1` is successfully defined.",
            "`biases_layer1` is missing. You must assign a 1x4 zero tensor to this variable."
        )

    def test_biases1_valid(self):
        import user_code
        try:
            t = user_code.biases_layer1
            cond = (
                isinstance(t, torch.Tensor)
                and t.shape == (1, 4)
                and torch.all(t == 0)
            )
        except:
            cond = False

        _dynamic(
            self,
            cond,
            "`biases_layer1` is correctly created as a 1x4 tensor of zeros.",
            "`biases_layer1` must be a 1x4 zero tensor created using `torch.zeros(1, 4)`."
        )

    # -------- weights_layer2 --------
    def test_weights2_exists(self):
        import user_code
        cond = hasattr(user_code, "weights_layer2")
        _dynamic(
            self,
            cond,
            "`weights_layer2` is successfully defined.",
            "`weights_layer2` is missing. Assign a 4x2 integer tensor created with `torch.randint()`."
        )

    def test_weights2_valid(self):
        import user_code
        try:
            t = user_code.weights_layer2
            cond = (
                isinstance(t, torch.Tensor)
                and t.shape == (4, 2)
                and torch.all(t >= -5)
                and torch.all(t <= 5)
                and torch.equal(t, t.int())
            )
        except:
            cond = False

        _dynamic(
            self,
            cond,
            "`weights_layer2` is a valid 4x2 integer tensor with values in the range [-5, 5].",
            "`weights_layer2` must be created using `torch.randint(-5, 5, (4, 2))` to generate integer weights in the specified range."
        )

    # -------- printing output --------
    def test_printed_output(self):
        import user_code

        import io, sys
        buf = io.StringIO()
        sys.stdout = buf

        importlib.reload(user_code)
        sys.stdout = sys.__stdout__

        cond = len(buf.getvalue().strip()) > 0

        _dynamic(
            self,
            cond,
            "All tensors are printed as required.",
            "No output detected. Ensure that you print all three tensors exactly as instructed."
        )


if __name__ == "__main__":
    unittest.main()


test_code.py

Transição de bibliotecas de alto nível para a flexibilidade do PyTorch. Domínio dos blocos fundamentais do deep learning ao trabalhar com tensores, a principal estrutura de dados do PyTorch. Exploração da diferenciação automática com gradientes, implementação do backpropagation e construção de redes neurais completas do zero. Aprendizado sobre como lidar com conjuntos de dados do mundo real e gerenciar todo o ciclo de treinamento – desde a inicialização dos pesos até a avaliação do modelo.

Desafio: Inicialização dos Pesos e Vieses do Modelo

Solução