def test_get_train_valid_restitutes_shape(self):
Xtr = np.arange(7840).reshape(10, 28, 28)
Ytr = np.ones(10)
some_dataset = ImageDataset(Xtr, Ytr, None, None)
(Xtr, Ytr), (Xval, Yval) = some_dataset.get_train_valid(3, False, False, False)
assert Xtr.shape[1:] == (28, 28)
assert Xval.shape[1:] == (28, 28)
def test_get_train_valid_centered_reduced(self):
Xtr_orig = np.arange(2).reshape(2, 1, 1)
# Xtr_orig = [[[0]], [[1]]] => mean = .5, variance = .25
Ytr_orig = np.ones(2)
some_dataset = ImageDataset(Xtr_orig, Ytr_orig, None, None)
valid = 0
(Xtr, Ytr), (Xval, Yval) = some_dataset.get_train_valid(valid, center=True, reduce=True, shuffle=False)
assert Xtr[0,0,0] == -1
assert Xtr[1,0,0] == 1
def test_get_train_valid_valid_as_float(self):
Xtr = np.arange(7840).reshape(10, 28, 28)
Ytr = np.ones(10)
some_dataset = ImageDataset(Xtr, Ytr, None, None)
valid = .35
(Xtr, Ytr), (Xval, Yval) = some_dataset.get_train_valid(valid, False, False, False)
assert Xtr.shape[0] == 7
assert Ytr.shape[0] == 7
assert Xval.shape[0] == 3
assert Yval.shape[0] == 3
def test_get_train_valid_shuffled(self):
Xtr_orig = np.arange(10).reshape(10, 1, 1)*10
Ytr_orig = -np.arange(10)
some_dataset = ImageDataset(Xtr_orig, Ytr_orig, None, None)
valid = 0
(Xtr, Ytr), (Xval, Yval) = some_dataset.get_train_valid(valid, center=False, reduce=False, shuffle=True)
assert not all(x_orig[0,0] == x[0,0] for x_orig, x in zip(Xtr_orig, Xtr))
indices = (np.argwhere(Xtr == x[0,0])[0][0] for x in Xtr_orig)
assert all(y == Ytr[idx] for idx, y in zip(indices, Ytr_orig))
def test_get_train_valid_reduced(self):
# Variance formula: E[(X - E[X])²]
Xtr_orig = np.arange(2).reshape(2, 1, 1)
# Xtr_orig = [[[0]], [[1]]] => variance = .25
Ytr_orig = np.ones(2)
some_dataset = ImageDataset(Xtr_orig, Ytr_orig, None, None)
valid = 0
(Xtr, Ytr), (Xval, Yval) = some_dataset.get_train_valid(valid, center=False, reduce=True, shuffle=False)
# We find the multiplicative factor "a" by setting the variance to 1:
# E[(a*X - E[a*X])²] = 1
# [(0 - a/2)² + (1 - a/2)²] = 1
# a = 2
assert Xtr[0,0,0] == 0
assert Xtr[1,0,0] == 2