def test_calculate_loss_sets_final_output_derivative_on_success(self):
pred = np.asarray([[randint(0, 3) for i in range (2)] for j in range(3)])
# make sure there is no loss
lab = pred
n = Net(MSE)
n.calculate_loss(pred, lab)
assert (n.loss_derivative == 0).all()
def test_sigmoid(self):
rand = [[randint(-10, 10) for i in range(50)] for i in range(30)]
torch_func = TorchSigmoid()
my_func = Sigmoid(Net(MeanSquaredError))
torch_result = torch_func(torch.DoubleTensor(rand))
self.assertTrue((torch_result == torch.from_numpy(
my_func.call(rand)).double()).all())
def test_relu(self):
rand = [[randint(-10, 10) for i in range(50)] for i in range(30)]
relu = ReLu(Net(MeanSquaredError))
torch_result = F.relu(Variable(torch.DoubleTensor(rand)))
t = torch_result.data
i = torch.from_numpy(relu.call(rand))
for x, y in zip(t, i):
for i, j in zip(x, y):
assert round(i.item(), 10) == round(j.item(), 10)
def test_old_weights_are_updated_on_second_pass(self):
input = randint(5, 10)
output = randint(5, 10)
weights = np.asarray([[randint(0, 2) for j in range(input)] for i in range(output)])
n = Net(Sigmoid)
linear = Linear(n, input, output, weights)
linear.forward(input)
sh = np.asarray(linear.network.saved_weights).shape
linear.forward(input)
assert sh == np.asarray(linear.network.saved_weights).shape
def test_saves_inputs_outputs_and_weights_to_network(self):
input = randint(5, 10)
output = randint(5, 10)
weights = np.asarray([[randint(0, 2) for j in range(input)] for i in range(output)])
n = Net(Sigmoid)
linear = Linear(n, input, output, weights)
linear.forward(input)
assert len(linear.network.saved_inputs) > 0
assert len(linear.network.saved_outputs) > 0
assert len(linear.network.saved_weights) > 0
def test_softmax(self):
rand = np.asarray([[randint(0, 10) for i in range(2)]
for i in range(3)])
softmax = SoftMax(Net(MeanSquaredError))
torch_result = F.softmax(Variable(torch.DoubleTensor(rand)), dim=1)
t = torch_result.data
i = torch.from_numpy(softmax.call(rand))
for x, y in zip(t, i):
for i, j in zip(x, y):
assert round(i.item(), 10) == round(j.item(), 10)
def test_convolution_output_with_biases(self):
torch.random.manual_seed(42)
inp = np.arange(96).reshape(3, 2, 4, 4)
torch_layer = nn.Conv2d(in_channels=2, out_channels=2, kernel_size=2)
torch_output = torch_layer(torch.tensor(inp).float())
kernel, biases = list(torch_layer.parameters())
# create using kernel generated by pytorch
neurose_layer = Conv2D(Net(None),
input_channels=1,
kernel_amount=2,
kernel_size=2,
use_biases=True,
initial_weights=kernel.detach().numpy(),
initial_biases=biases.detach().numpy())
my_output = neurose_layer.forward_pass(inp)
self.assertTrue(np.allclose(my_output, torch_output.detach().numpy()))
def test_convolution_output_with_one_kernel_more_channels(self):
torch.random.manual_seed(42)
in_channels = 3
batch_size = 3
kernel_size = 2
kernel_amount = 1
inp = np.arange(36).reshape(batch_size, in_channels, 2, 2)
torch_layer = nn.Conv2d(in_channels=in_channels,
out_channels=kernel_amount,
kernel_size=kernel_size,
bias=False)
torch_output = torch_layer(torch.tensor(inp).float())
kernel = list(torch_layer.parameters())[0].detach().numpy()
# create using kernel generated by pytorch
neurose_layer = Conv2D(Net(None),
input_channels=in_channels,
kernel_amount=kernel_amount,
kernel_size=kernel_size,
initial_weights=kernel)
my_output = neurose_layer.forward_pass(inp)
self.assertTrue(np.allclose(my_output, torch_output.detach().numpy()))
def test_raises_error_if_biases_not_right_dimension(self):
input = randint(5, 10)
output = randint(5, 10)
# biases should be the dimension of output
biases = np.ndarray(output + 2)
self.assertRaises(ValueError, Linear, Net(Sigmoid), input, output, np.ndarray(0), biases)
def test_raises_error_if_weights_not_right_dimension(self):
input = randint(5, 10)
output = randint(5, 10)
weights = np.ndarray((output + 1, input))
self.assertRaises(ValueError, Linear, Net(Sigmoid), input, output, weights)
def test_biases_are_initialized_correctly(self):
input = randint(5, 10)
output = randint(5, 10)
biases = np.asarray([randint(1, 5) for i in range(output)])
linear = Linear(Net(Sigmoid), input, output, np.ndarray(0), biases)
self.assertTrue(np.array_equal(linear.biases, biases))
def test_weights_are_initialized_correctly(self):
input = randint(5, 10)
output = randint(5, 10)
weights = np.asarray([[randint(0, 2) for j in range(input)] for i in range(output)])
linear = Linear(Net(Sigmoid), input, output, weights)
self.assertTrue((linear.weights == weights).all())
def test_linear_layer_output_correct_shape(self):
input = randint(5, 10)
output = randint(5, 10)
a = [randint(0, 5) for i in range(input)]
linear = Linear(Net(Sigmoid), input, output)
self.assertTrue(len(linear.forward(a)) == output)
def test_update_weights_raises_error_if_backpropagate_not_called(self):
pred = np.asarray([[randint(0, 3) for i in range (2)] for j in range(3)])
lab = np.asarray([[randint(0, 3) for i in range (2)] for j in range(3)])
n = Net(MSE)
self.assertRaises(ValueError, n.update_weights)
def test_backpropagation_raises_error_if_calculate_loss_not_called(self):
pred = np.asarray([[randint(0, 3) for i in range (2)] for j in range(3)])
lab = np.asarray([[randint(0, 3) for i in range (2)] for j in range(3)])
n = Net(MSE)
self.assertRaises(ValueError, n.backpropagate)
def test_calculate_loss_raises_error_if_wrong_dimensions(self):
pred = np.asarray([[i for i in range (2)] for j in range(3)])
lab = np.asarray([[i for i in range (3)] for j in range(2)])
n = Net(MSE)
self.assertRaises(ValueError, n.calculate_loss, pred, lab)
