def test_compare1(self):
rate = 0.4
x = np.random.rand(10, 2)
t = np.zeros((10)).astype(int)
layer = dezero.layers.Linear(2, 3, nobias=True)
layer.W.data = np.ones_like(layer.W.data)
optimizer = dezero.optimizers.SGD().setup(layer)
optimizer.add_hook(dezero.optimizers.WeightDecay(rate=rate))
layer.cleargrads()
y = layer(x)
y = F.softmax_cross_entropy(y, t)
y.backward()
optimizer.update()
W0 = layer.W.data.copy()
layer.W.data = np.ones_like(layer.W.data)
optimizer.hooks.clear()
layer.cleargrads()
y = layer(x)
y = F.softmax_cross_entropy(y, t) + rate / 2 * (layer.W**2).sum()
y.backward()
optimizer.update()
W1 = layer.W.data
self.assertTrue(array_allclose(W0, W1))
def test_forward1(self):
x = np.array([[-1, 0, 1, 2], [2, 0, 1, -1]], np.float32)
t = np.array([3, 0]).astype(np.int32)
y = F.softmax_cross_entropy(x, t)
y2 = CF.softmax_cross_entropy(x, t)
res = np.allclose(y.data, y2.data)
self.assertTrue(res)
def test_MNIST(self):
max_epoch = 5
batch_size = 100
hidden_size = 1000
train_set = dezero.datasets.MNIST(train=True)
test_set = dezero.datasets.MNIST(train=False)
train_loader = DataLoader(train_set, batch_size)
test_loader = DataLoader(test_set, batch_size, shuffle=False)
#model = MLP((hidden_size, 10))
model = MLP((hidden_size, hidden_size, 10), activation=F.relu)
optimizer = optimizers.SGD().setup(model)
if dezero.cuda.gpu_enable:
train_loader.to_gpu()
model.to_gpu()
for epoch in range(max_epoch):
sum_loss, sum_acc = 0, 0
for x, t in train_loader:
y = model(x)
loss = F.softmax_cross_entropy(y, t)
acc = F.accuracy(y, t)
model.cleargrads()
loss.backward()
optimizer.update()
sum_loss += float(loss.data) * len(t)
sum_acc += float(acc.data) * len(t)
print('epoch: {}'.format(epoch + 1))
print('train loss: {:.4f}, accuracy: {:.4f}'.format(
sum_loss / len(train_set), sum_acc / len(train_set)))
sum_loss, sum_acc = 0, 0
with dezero.no_grad():
for x, t in test_loader:
y = model(x)
loss = F.softmax_cross_entropy(y, t)
acc = F.accuracy(y, t)
sum_loss += float(loss.data) * len(t)
sum_acc += float(acc.data) * len(t)
print('test loss: {:.4f}, accuracy: {:.4f}'.format(
sum_loss / len(test_set), sum_acc / len(test_set)))
def test_softmax_cross_entropy(self):
x = Variable(np.array([[1.0, 1.0, 1.0], [3.0, 2.0, 1.0]]))
t = Variable(np.array([1, 0]))
y = softmax_cross_entropy(x, t)
p_1 = np.log(np.exp(1.0) / (3 * np.exp(1.0)))
p_2 = np.log(np.exp(3.0) / (np.exp(3.0) + np.exp(2.0) + np.exp(1.0)))
expected_ans = -(p_1 + p_2) / 2
assert_array_equal(y.data, np.array(expected_ans))
def test_SoftmaxCrossEntorpy(self):
max_epoch = 0
batch_size = 30
hidden_size = 10
lr = 1.0
train_set = Spiral(train=True)
test_set = Spiral(train=False)
train_loader = DataLoader(train_set, batch_size)
test_loader = DataLoader(test_set, batch_size, shuffle=False)
model = MLP((hidden_size, hidden_size, hidden_size, 3))
optimizer = optimizers.SGD(lr).setup(model)
for epoch in range(max_epoch):
sum_loss, sum_acc = 0, 0
for x, t in train_loader:
y = model(x)
loss = F.softmax_cross_entropy(y, t)
acc = F.accuracy(y, t)
model.cleargrads()
loss.backward()
optimizer.update()
sum_loss += float(loss.data) * len(t)
sum_acc += float(acc.data) * len(t)
print('epoch: {}'.format(epoch + 1))
print('train loss: {:.4f}, accuracy: {:.4f}'.format(
sum_loss / len(train_set), sum_acc / len(train_set)))
sum_loss, sum_acc = 0, 0
with dezero.no_grad():
for x, t in test_loader:
y = model(x)
loss = F.softmax_cross_entropy(y, t)
acc = F.accuracy(y, t)
sum_loss += float(loss.data) * len(t)
sum_acc += float(acc.data) * len(t)
print('test loss: {:.4f}, accuracy: {:.4f}'.format(
sum_loss / len(test_set), sum_acc / len(test_set)))
def test_forward(self):
x = Variable(
np.array([[0.2, -0.4], [0.3, 0.5], [1.3, -3.2], [2.1, 0.3]]))
t = np.array([1, 0, 1, 0])
y = F.softmax_cross_entropy(x, t)
y_expected = np.exp(x.data) / \
np.exp(x.data).sum(axis=1, keepdims=True)
y_expected = -np.log(y_expected)
y_expected = y_expected[np.arange(4), t].sum() / 4
assert_almost_equal(y.data, y_expected, decimal=5)
def test_backward(self):
x = Variable(
np.array([[0.2, -0.4], [0.3, 0.5], [1., -3.2], [2.1, 0.3]]))
t = np.array([1, 0, 1, 0])
t_onehot = np.eye(2)[t]
L = F.softmax_cross_entropy(x, t)
L.backward()
m = len(x.data)
y = np.exp(x.data) / np.exp(x.data).sum(axis=1, keepdims=True)
gx_expected = 1 / m * (y - t_onehot)
assert_almost_equal(x.grad.data, gx_expected)
optimizer = optimizers.SGD(lr).setup(model)
data_size = len(x)
max_iter = math.ceil(data_size / batch_size)
for epoch in range(max_epoch):
index = np.random.permutation(data_size)
sum_loss = 0
for i in range(max_iter):
batch_index = index[i * batch_size:(i + 1) * batch_size]
batch_x = x[batch_index]
batch_t = t[batch_index]
y = model(batch_x)
loss = F.softmax_cross_entropy(y, batch_t)
model.clear_grads()
loss.backward()
optimizer.update()
sum_loss += float(loss.data) * len(batch_t)
avg_loss = sum_loss / data_size
print('epoch %d, loss %.2f' % (epoch + 1, avg_loss))
# Plot boundary area the model predict
h = 0.001
x_min, x_max = x[:, 0].min() - .1, x[:, 0].max() + .1
y_min, y_max = x[:, 1].min() - .1, x[:, 1].max() + .1
xx, yy = np.meshgrid(np.arange(x_min, x_max, h), np.arange(y_min, y_max, h))
X = np.c_[xx.ravel(), yy.ravel()]
# /root/.dezero/t10k-images-idx3-ubyte.gz 경로에 있음
train_set = dezero.datasets.MNIST(train=True)
test_set = dezero.datasets.MNIST(train=False)
train_loader = DataLoader(train_set, batch_size)
test_loader = DataLoader(test_set, batch_size, shuffle=False)
# model = MLP((hidden_size, 10))
model = MLP((hidden_size, hidden_size, 10), activation=F.relu)
optimizer = SGD().setup(model)
for epoch in range(max_epoch):
sum_loss, sum_acc = 0, 0
for x, t in train_loader:
y = model(x)
loss = F.softmax_cross_entropy(y, t)
acc = F.accuracy(y, t)
model.cleargrads()
loss.backward()
optimizer.update()
sum_loss += float(loss.data) * len(t)
sum_acc += float(acc.data) * len(t)
print("epoch: {}".format(epoch + 1))
print("train loss: {:.4f}, accuracy: {:.4f}".format(
sum_loss / len(train_set), sum_acc / len(train_set)))
sum_loss, sum_acc = 0, 0
with dezero.no_grad(): # 기울기 불필요 모드
for x, t in test_loader: # 테스트용 미니배치 데이터
def test_backward3(self):
N, CLS_NUM = 100, 10
x = np.random.randn(N, CLS_NUM)
t = np.random.randint(0, CLS_NUM, (N, ))
f = lambda x: F.softmax_cross_entropy(x, t)
self.assertTrue(gradient_check(f, x))
def test_backward1(self):
x = np.array([[-1, 0, 1, 2], [2, 0, 1, -1]], np.float32)
t = np.array([3, 0]).astype(np.int32)
f = lambda x: F.softmax_cross_entropy(x, Variable(t))
self.assertTrue(gradient_check(f, x))
def test_dont_divide_by_zero(self):
x = np.array([[1., 0., 300000., 400000.]])
t = np.array([1])
L = F.softmax_cross_entropy(x, t)
self.assertFalse(np.isnan(L.data) or np.isinf(L.data))
def test_dont_overflow(self):
x = Variable(np.array([[1.0, 100.0, 10000.0]]))
t = np.array([0])
L = F.softmax_cross_entropy(x, t)
self.assertFalse(np.isnan(L.data) or np.isinf(L.data))
