def test_2d_to_1d():
min_x, max_x, gradient, intercept, n_samples = 10, 30, 1.5, 20, 300
X_1 = np.random.uniform(min_x, max_x, n_samples)[:, np.newaxis]
X_2_exact = gradient * X_1 + intercept
flat_noise = np.random.standard_normal(size=(n_samples, 1))
X_2_noise = flat_noise * (10 - X_1) * (X_1 - 30) / 40
X_2 = X_2_exact + X_2_noise
X_train = np.concatenate([X_1, X_2], axis=1)
pca_model = PCA(n_components=1)
pca_model.fit(features=X_train)
assert pca_model.n_components_ == 1
assert pca_model.n_samples_ == n_samples
assert pca_model.n_features_ == 2
assert pca_model.explained_variance_.shape == (1, )
assert pca_model.explained_variance_[0] >= 80
assert pca_model.explained_variance_ratio_.shape == (1, )
assert pca_model.explained_variance_ratio_[0] >= 0.98
assert pca_model.components_.shape == (1, 2)
principle = pca_model.components_[0]
assert principle[1] / principle[0] == pytest.approx(gradient, abs=0.2)
assert principle[1]**2 + principle[0]**2 == pytest.approx(1)
X_transform = pca_model.transform(features=X_train)
assert X_transform.shape == (n_samples, 1)
assert np.mean(X_transform) == pytest.approx(0)
def pca_and_pair_plot():
from endochrone.stats.scaling import FeatureScaling
from endochrone.decomposition import PCA
global x
if test_fit:
# TODO: make this part of PCA
# We need to remove features with no variation, otherwise scaling & PCA
# will fail
features_to_keep = [
i for i, feat in enumerate(x.T) if len(np.unique(feat)) > 1
]
x = x[:, features_to_keep]
# first we rescale to stop large floats dominating categories
scale_model = FeatureScaling(method='z_score')
scaled_x = scale_model.fit_and_transform(features=x)
N = 3
pcam_min = PCA(n_components=N)
pcam_min.fit(features=scaled_x)
pca_x = pcam_min.transform(features=scaled_x)
labels = (list(range(N)) + ['species'])
df = pd.DataFrame(np.hstack([pca_x, y[:, np.newaxis]]), columns=labels)
sns.pairplot(df, hue='species', height=1.5)
plt.show()
def test_zero_components_specified():
min_x, max_x, gradient, intercept, n_samples = 10, 30, 4, 20, 300
X_1 = np.random.uniform(min_x, max_x, n_samples)[:, np.newaxis]
X_2_exact = gradient * X_1 + intercept
flat_noise = np.random.standard_normal(size=(n_samples, 1))
X_2_noise = flat_noise * (10 - X_1) * (X_1 - 30) / 40
X_2 = X_2_exact + X_2_noise
X_train = np.concatenate([X_1, X_2], axis=1)
pca_model = PCA()
pca_model.fit(features=X_train)
assert pca_model.n_components_ == 2
assert pca_model.n_samples_ == n_samples
assert pca_model.n_features_ == 2
assert pca_model.explained_variance_.shape == (2, )
assert pca_model.explained_variance_ratio_.shape == (2, )
assert sum(pca_model.explained_variance_ratio_) == pytest.approx(1)
assert pca_model.components_.shape == (2, 2)
def iris_pca():
from mpl_toolkits import mplot3d # noqa: F401
from endochrone.decomposition import PCA
fig = plt.figure(facecolor="w", figsize=(14, 7))
i_target = iris['target']
pcam_2 = PCA(n_components=2)
red_i_data_2 = pcam_2.fit_and_transform(features=i_data)
var_sum_2 = np.abs(np.sum(pcam_2.explained_variance_ratio_))
title_2 = '%s components, capturing %.4f%% variation' % (2, var_sum_2*100)
X_2 = red_i_data_2[:, 0]
Y_2 = red_i_data_2[:, 1]
ax_2 = fig.add_subplot(1, 2, 1, title=title_2)
ax_2.scatter(X_2, Y_2, c=i_target, s=3, marker='d', cmap='cool')
# Compare to 3 component reduction
pcam_3 = PCA(n_components=3)
red_i_data_3 = pcam_3.fit_and_transform(features=i_data)
var_sum_3 = np.abs(np.sum(pcam_3.explained_variance_ratio_))
title_3 = '%s components, capturing %.4f%% variation' % (2, var_sum_3*100)
X_3 = red_i_data_3[:, 0]
Y_3 = red_i_data_3[:, 1]
Z_3 = red_i_data_3[:, 2]
ax_3 = fig.add_subplot(1, 2, 2, projection='3d', title=title_3)
ax_3.scatter3D(X_3, Y_3, Z_3, c=i_target, s=3, marker='d', cmap='cool')
plt.show()
def test_6d_to_2d():
n_samples = 300
min_x, max_x, gradient_1, gradient_2, intercept = 10, 30, 1.5, 0.6, 20
X_1 = np.random.uniform(min_x, max_x, n_samples)[:, np.newaxis]
X_2 = np.random.uniform(min_x, max_x, n_samples)[:, np.newaxis]
noise = 5 * np.random.standard_normal(size=(n_samples, 1))
X_3 = gradient_1 * X_1 + intercept + noise
X_4 = gradient_2 * X_2 + intercept + noise
X_5 = X_3 + X_2
X_6 = X_4 + X_1
X_train = np.concatenate([X_1, X_2, X_3, X_4, X_5, X_6], axis=1)
pca_model = PCA(n_components=2)
pca_model.fit(features=X_train)
assert pca_model.n_components_ == 2
assert pca_model.n_samples_ == n_samples
assert pca_model.n_features_ == 6
assert pca_model.explained_variance_.shape == (2, )
assert pca_model.explained_variance_ratio_.shape == (2, )
assert np.abs(np.sum(pca_model.explained_variance_ratio_)) > 0.92
assert pca_model.components_.shape == (2, 6)
def test_accuracy_of_inversion():
n_features, n_samples = 6, 300
X_train = np.random.rand(n_samples, n_features)
pcam = PCA(n_components=n_features)
pcam.fit(features=X_train)
act = pcam.reverse(features=pcam.transform(features=X_train))
assert np.all(act == pytest.approx(X_train))
def with_pca():
from endochrone.stats.scaling import FeatureScaling
from endochrone.decomposition import PCA
global x
if test_fit:
# TODO: make this part of PCA
# We need to remove features with no variation, otherwise scaling & PCA
# will fail
features_to_keep = [
i for i, feat in enumerate(x.T) if len(np.unique(feat)) > 1
]
x = x[:, features_to_keep]
# first we rescale to stop large floats dominating categories
scale_model = FeatureScaling(method='z_score')
scaled_x = scale_model.fit_and_transform(features=x)
# Then build a test PCA model so we can figure out how many components we
# want
pcam_test = PCA()
pcam_test.fit(features=scaled_x)
cutoff = 0.97 # i.e. we want to retain this % of variance
# TODO: want to be able to do this with a single PCA model
n_comp = np.argmax(np.cumsum(pcam_test.explained_variance_ratio_) > cutoff)
# Optionally see how our variance increases as n_comp increases
show_pca_variances = False
if show_pca_variances:
plt.plot(range(54), np.cumsum(pcam_test.explained_variance_ratio_))
plt.show()
# Transform according to the above PCA
pcam_forest = PCA(n_components=n_comp)
pca_x = pcam_forest.fit_and_transform(features=scaled_x)
xtrain, xtest, ytrain, ytest = train_test_split(pca_x, y)
# Now build our RF model and fit it
t0 = time.process_time()
forest_model = RandomForest(64, max_tree_depth=15)
forest_model.fit(xtrain, ytrain, debug=True)
t1 = time.process_time()
print('train time: ', t1 - t0)
# print(forest_model.trees[0])
ypred = forest_model.predict(xtest)
metrics = mcm(ytest, ypred)
print('\nmetrics for test set')
print('macro_precision', metrics.macro_precision)
print('macro_recall', metrics.macro_recall)
print('macro_f1', metrics.macro_f1_score)
print('micro_f1', metrics.micro_f1_score)
train_pred = forest_model.predict(xtrain)
metrics = mcm(ytrain, train_pred)
print('\nmetrics for training set')
print('macro_precision', metrics.macro_precision)
print('macro_recall', metrics.macro_recall)
print('macro_f1', metrics.macro_f1_score)
print('micro_f1', metrics.micro_f1_score)
"""Example output:
gridspec_kw=dict(hspace=0.1, wspace=0.1))
for i, ax in enumerate(axes.flat):
ax.imshow(digits.images[i], cmap='binary', interpolation='nearest')
ax.text(0.05,
0.05,
str(digits.target[i]),
transform=ax.transAxes,
color='g')
plt.show()
Xtrain, Xtest, Ytrain, Ytest = train_test_split(digits.data, digits.target)
# Test run to figure out how many components we should keep
# TODO should be able to do this with a single model
pcam_test = PCA()
pcam_test.fit(features=Xtrain)
cutoff = 0.97 # i.e. we want to retain this % of variance
n_comp = np.argmax(np.cumsum(pcam_test.explained_variance_ratio_) > cutoff)
# Now reduce our features with this PCA model
pcam = PCA(n_components=n_comp)
pca_Xtrain = pcam.fit_and_transform(features=Xtrain)
pca_Xtest = pcam.transform(features=Xtest)
# Try KNN to see if we're any good at classifying
knn_model = KNearest()
knn_model.fit(features=pca_Xtrain, targets=Ytrain)
ypred = knn_model.predict(features=pca_Xtest)
acc = accuracy_score(Ytest, ypred) * 100
print("cut-off: %s \nn_comp: %s \naccuracy: %0.4f%%" % (cutoff, n_comp, acc))