#!/usr/bin/env python
# coding: utf-8

# **Chapter 8 – Dimensionality Reduction**
# 
# _This notebook contains all the sample code and solutions to the exercises in chapter 8._
# 
# <table align="left">
#   <td>
#     <a target="_blank" href="https://colab.research.google.com/github/ageron/handson-ml/blob/master/08_dimensionality_reduction.ipynb"><img src="https://www.tensorflow.org/images/colab_logo_32px.png" />Run in Google Colab</a>
#   </td>
# </table>

# **Warning**: this is the code for the 1st edition of the book. Please visit https://github.com/ageron/handson-ml2 for the 2nd edition code, with up-to-date notebooks using the latest library versions.

# # Setup

# First, let's make sure this notebook works well in both python 2 and 3, import a few common modules, ensure MatplotLib plots figures inline and prepare a function to save the figures:

# In[1]:


# To support both python 2 and python 3
from __future__ import division, print_function, unicode_literals

# Common imports
import numpy as np
import os

# to make this notebook's output stable across runs
np.random.seed(42)

# To plot pretty figures
get_ipython().run_line_magic('matplotlib', 'inline')
import matplotlib as mpl
import matplotlib.pyplot as plt
mpl.rc('axes', labelsize=14)
mpl.rc('xtick', labelsize=12)
mpl.rc('ytick', labelsize=12)

# Where to save the figures
PROJECT_ROOT_DIR = "."
CHAPTER_ID = "unsupervised_learning"
IMAGES_PATH = os.path.join(PROJECT_ROOT_DIR, "images", CHAPTER_ID)
os.makedirs(IMAGES_PATH, exist_ok=True)

def save_fig(fig_id, tight_layout=True, fig_extension="png", resolution=300):
    path = os.path.join(IMAGES_PATH, fig_id + "." + fig_extension)
    print("Saving figure", fig_id)
    if tight_layout:
        plt.tight_layout()
    plt.savefig(path, format=fig_extension, dpi=resolution)


# # Projection methods
# Build 3D dataset:

# In[2]:


np.random.seed(4)
m = 60
w1, w2 = 0.1, 0.3
noise = 0.1

angles = np.random.rand(m) * 3 * np.pi / 2 - 0.5
X = np.empty((m, 3))
X[:, 0] = np.cos(angles) + np.sin(angles)/2 + noise * np.random.randn(m) / 2
X[:, 1] = np.sin(angles) * 0.7 + noise * np.random.randn(m) / 2
X[:, 2] = X[:, 0] * w1 + X[:, 1] * w2 + noise * np.random.randn(m)


# ## PCA using SVD decomposition

# Note: the `svd()` function returns `U`, `s` and `Vt`, where `Vt` is equal to $\mathbf{V}^T$, the transpose of the matrix $\mathbf{V}$. Earlier versions of the book mistakenly said that it returned `V` instead of `Vt`. Also, Equation 8-1 should actually contain $\mathbf{V}$ instead of $\mathbf{V}^T$, like this:
# 
# $
# \mathbf{V} =
# \begin{pmatrix}
#   \mid & \mid & & \mid \\
#   \mathbf{c_1} & \mathbf{c_2} & \cdots & \mathbf{c_n} \\
#   \mid & \mid & & \mid
# \end{pmatrix}
# $

# In[3]:


X_centered = X - X.mean(axis=0)
U, s, Vt = np.linalg.svd(X_centered)
c1 = Vt.T[:, 0]
c2 = Vt.T[:, 1]


# In[4]:


m, n = X.shape

S = np.zeros(X_centered.shape)
S[:n, :n] = np.diag(s)


# In[5]:


np.allclose(X_centered, U.dot(S).dot(Vt))


# In[6]:


W2 = Vt.T[:, :2]
X2D = X_centered.dot(W2)


# In[7]:


X2D_using_svd = X2D


# ## PCA using Scikit-Learn

# With Scikit-Learn, PCA is really trivial. It even takes care of mean centering for you:

# In[8]:


from sklearn.decomposition import PCA

pca = PCA(n_components = 2)
X2D = pca.fit_transform(X)


# In[9]:


X2D[:5]


# In[10]:


X2D_using_svd[:5]


# Notice that running PCA multiple times on slightly different datasets may result in different results. In general the only difference is that some axes may be flipped. In this example, PCA using Scikit-Learn gives the same projection as the one given by the SVD approach, except both axes are flipped:

# In[11]:


np.allclose(X2D, -X2D_using_svd)


# Recover the 3D points projected on the plane (PCA 2D subspace).

# In[12]:


X3D_inv = pca.inverse_transform(X2D)


# Of course, there was some loss of information during the projection step, so the recovered 3D points are not exactly equal to the original 3D points:

# In[13]:


np.allclose(X3D_inv, X)


# We can compute the reconstruction error:

# In[14]:


np.mean(np.sum(np.square(X3D_inv - X), axis=1))


# The inverse transform in the SVD approach looks like this:

# In[15]:


X3D_inv_using_svd = X2D_using_svd.dot(Vt[:2, :])


# The reconstructions from both methods are not identical because Scikit-Learn's `PCA` class automatically takes care of reversing the mean centering, but if we subtract the mean, we get the same reconstruction:

# In[16]:


np.allclose(X3D_inv_using_svd, X3D_inv - pca.mean_)


# The `PCA` object gives access to the principal components that it computed:

# In[17]:


pca.components_


# Compare to the first two principal components computed using the SVD method:

# In[18]:


Vt[:2]


# Notice how the axes are flipped.

# Now let's look at the explained variance ratio:

# In[19]:


pca.explained_variance_ratio_


# The first dimension explains 84.2% of the variance, while the second explains 14.6%.

# By projecting down to 2D, we lost about 1.1% of the variance:

# In[20]:


1 - pca.explained_variance_ratio_.sum()


# Here is how to compute the explained variance ratio using the SVD approach (recall that `s` is the diagonal of the matrix `S`):

# In[21]:


np.square(s) / np.square(s).sum()


# Next, let's generate some nice figures! :)

# Utility class to draw 3D arrows (copied from http://stackoverflow.com/questions/11140163)

# In[22]:


from matplotlib.patches import FancyArrowPatch
from mpl_toolkits.mplot3d import proj3d

class Arrow3D(FancyArrowPatch):
    def __init__(self, xs, ys, zs, *args, **kwargs):
        FancyArrowPatch.__init__(self, (0,0), (0,0), *args, **kwargs)
        self._verts3d = xs, ys, zs

    def draw(self, renderer):
        xs3d, ys3d, zs3d = self._verts3d
        xs, ys, zs = proj3d.proj_transform(xs3d, ys3d, zs3d, renderer.M)
        self.set_positions((xs[0],ys[0]),(xs[1],ys[1]))
        FancyArrowPatch.draw(self, renderer)


# Express the plane as a function of x and y.

# In[23]:


axes = [-1.8, 1.8, -1.3, 1.3, -1.0, 1.0]

x1s = np.linspace(axes[0], axes[1], 10)
x2s = np.linspace(axes[2], axes[3], 10)
x1, x2 = np.meshgrid(x1s, x2s)

C = pca.components_
R = C.T.dot(C)
z = (R[0, 2] * x1 + R[1, 2] * x2) / (1 - R[2, 2])


# Plot the 3D dataset, the plane and the projections on that plane.

# In[24]:


from mpl_toolkits.mplot3d import Axes3D

fig = plt.figure(figsize=(6, 3.8))
ax = fig.add_subplot(111, projection='3d')

X3D_above = X[X[:, 2] > X3D_inv[:, 2]]
X3D_below = X[X[:, 2] <= X3D_inv[:, 2]]

ax.plot(X3D_below[:, 0], X3D_below[:, 1], X3D_below[:, 2], "bo", alpha=0.5)

ax.plot_surface(x1, x2, z, alpha=0.2, color="k")
np.linalg.norm(C, axis=0)
ax.add_artist(Arrow3D([0, C[0, 0]],[0, C[0, 1]],[0, C[0, 2]], mutation_scale=15, lw=1, arrowstyle="-|>", color="k"))
ax.add_artist(Arrow3D([0, C[1, 0]],[0, C[1, 1]],[0, C[1, 2]], mutation_scale=15, lw=1, arrowstyle="-|>", color="k"))
ax.plot([0], [0], [0], "k.")

for i in range(m):
    if X[i, 2] > X3D_inv[i, 2]:
        ax.plot([X[i][0], X3D_inv[i][0]], [X[i][1], X3D_inv[i][1]], [X[i][2], X3D_inv[i][2]], "k-")
    else:
        ax.plot([X[i][0], X3D_inv[i][0]], [X[i][1], X3D_inv[i][1]], [X[i][2], X3D_inv[i][2]], "k-", color="#505050")
    
ax.plot(X3D_inv[:, 0], X3D_inv[:, 1], X3D_inv[:, 2], "k+")
ax.plot(X3D_inv[:, 0], X3D_inv[:, 1], X3D_inv[:, 2], "k.")
ax.plot(X3D_above[:, 0], X3D_above[:, 1], X3D_above[:, 2], "bo")
ax.set_xlabel("$x_1$", fontsize=18)
ax.set_ylabel("$x_2$", fontsize=18)
ax.set_zlabel("$x_3$", fontsize=18)
ax.set_xlim(axes[0:2])
ax.set_ylim(axes[2:4])
ax.set_zlim(axes[4:6])

# Note: If you are using Matplotlib 3.0.0, it has a bug and does not
# display 3D graphs properly.
# See https://github.com/matplotlib/matplotlib/issues/12239
# You should upgrade to a later version. If you cannot, then you can
# use the following workaround before displaying each 3D graph:
# for spine in ax.spines.values():
#     spine.set_visible(False)

save_fig("dataset_3d_plot")
plt.show()


# In[25]:


fig = plt.figure()
ax = fig.add_subplot(111, aspect='equal')

ax.plot(X2D[:, 0], X2D[:, 1], "k+")
ax.plot(X2D[:, 0], X2D[:, 1], "k.")
ax.plot([0], [0], "ko")
ax.arrow(0, 0, 0, 1, head_width=0.05, length_includes_head=True, head_length=0.1, fc='k', ec='k')
ax.arrow(0, 0, 1, 0, head_width=0.05, length_includes_head=True, head_length=0.1, fc='k', ec='k')
ax.set_xlabel("$z_1$", fontsize=18)
ax.set_ylabel("$z_2$", fontsize=18, rotation=0)
ax.axis([-1.5, 1.3, -1.2, 1.2])
ax.grid(True)
save_fig("dataset_2d_plot")


# # Manifold learning
# Swiss roll:

# In[26]:


from sklearn.datasets import make_swiss_roll
X, t = make_swiss_roll(n_samples=1000, noise=0.2, random_state=42)


# In[27]:


axes = [-11.5, 14, -2, 23, -12, 15]

fig = plt.figure(figsize=(6, 5))
ax = fig.add_subplot(111, projection='3d')

ax.scatter(X[:, 0], X[:, 1], X[:, 2], c=t, cmap=plt.cm.hot)
ax.view_init(10, -70)
ax.set_xlabel("$x_1$", fontsize=18)
ax.set_ylabel("$x_2$", fontsize=18)
ax.set_zlabel("$x_3$", fontsize=18)
ax.set_xlim(axes[0:2])
ax.set_ylim(axes[2:4])
ax.set_zlim(axes[4:6])

save_fig("swiss_roll_plot")
plt.show()


# In[28]:


plt.figure(figsize=(11, 4))

plt.subplot(121)
plt.scatter(X[:, 0], X[:, 1], c=t, cmap=plt.cm.hot)
plt.axis(axes[:4])
plt.xlabel("$x_1$", fontsize=18)
plt.ylabel("$x_2$", fontsize=18, rotation=0)
plt.grid(True)

plt.subplot(122)
plt.scatter(t, X[:, 1], c=t, cmap=plt.cm.hot)
plt.axis([4, 15, axes[2], axes[3]])
plt.xlabel("$z_1$", fontsize=18)
plt.grid(True)

save_fig("squished_swiss_roll_plot")
plt.show()


# In[29]:


from matplotlib import gridspec

axes = [-11.5, 14, -2, 23, -12, 15]

x2s = np.linspace(axes[2], axes[3], 10)
x3s = np.linspace(axes[4], axes[5], 10)
x2, x3 = np.meshgrid(x2s, x3s)

fig = plt.figure(figsize=(6, 5))
ax = plt.subplot(111, projection='3d')

positive_class = X[:, 0] > 5
X_pos = X[positive_class]
X_neg = X[~positive_class]
ax.view_init(10, -70)
ax.plot(X_neg[:, 0], X_neg[:, 1], X_neg[:, 2], "y^")
ax.plot_wireframe(5, x2, x3, alpha=0.5)
ax.plot(X_pos[:, 0], X_pos[:, 1], X_pos[:, 2], "gs")
ax.set_xlabel("$x_1$", fontsize=18)
ax.set_ylabel("$x_2$", fontsize=18)
ax.set_zlabel("$x_3$", fontsize=18)
ax.set_xlim(axes[0:2])
ax.set_ylim(axes[2:4])
ax.set_zlim(axes[4:6])

save_fig("manifold_decision_boundary_plot1")
plt.show()

fig = plt.figure(figsize=(5, 4))
ax = plt.subplot(111)

plt.plot(t[positive_class], X[positive_class, 1], "gs")
plt.plot(t[~positive_class], X[~positive_class, 1], "y^")
plt.axis([4, 15, axes[2], axes[3]])
plt.xlabel("$z_1$", fontsize=18)
plt.ylabel("$z_2$", fontsize=18, rotation=0)
plt.grid(True)

save_fig("manifold_decision_boundary_plot2")
plt.show()

fig = plt.figure(figsize=(6, 5))
ax = plt.subplot(111, projection='3d')

positive_class = 2 * (t[:] - 4) > X[:, 1]
X_pos = X[positive_class]
X_neg = X[~positive_class]
ax.view_init(10, -70)
ax.plot(X_neg[:, 0], X_neg[:, 1], X_neg[:, 2], "y^")
ax.plot(X_pos[:, 0], X_pos[:, 1], X_pos[:, 2], "gs")
ax.set_xlabel("$x_1$", fontsize=18)
ax.set_ylabel("$x_2$", fontsize=18)
ax.set_zlabel("$x_3$", fontsize=18)
ax.set_xlim(axes[0:2])
ax.set_ylim(axes[2:4])
ax.set_zlim(axes[4:6])

save_fig("manifold_decision_boundary_plot3")
plt.show()

fig = plt.figure(figsize=(5, 4))
ax = plt.subplot(111)

plt.plot(t[positive_class], X[positive_class, 1], "gs")
plt.plot(t[~positive_class], X[~positive_class, 1], "y^")
plt.plot([4, 15], [0, 22], "b-", linewidth=2)
plt.axis([4, 15, axes[2], axes[3]])
plt.xlabel("$z_1$", fontsize=18)
plt.ylabel("$z_2$", fontsize=18, rotation=0)
plt.grid(True)

save_fig("manifold_decision_boundary_plot4")
plt.show()


# # PCA

# In[30]:


angle = np.pi / 5
stretch = 5
m = 200

np.random.seed(3)
X = np.random.randn(m, 2) / 10
X = X.dot(np.array([[stretch, 0],[0, 1]])) # stretch
X = X.dot([[np.cos(angle), np.sin(angle)], [-np.sin(angle), np.cos(angle)]]) # rotate

u1 = np.array([np.cos(angle), np.sin(angle)])
u2 = np.array([np.cos(angle - 2 * np.pi/6), np.sin(angle - 2 * np.pi/6)])
u3 = np.array([np.cos(angle - np.pi/2), np.sin(angle - np.pi/2)])

X_proj1 = X.dot(u1.reshape(-1, 1))
X_proj2 = X.dot(u2.reshape(-1, 1))
X_proj3 = X.dot(u3.reshape(-1, 1))

plt.figure(figsize=(8,4))
plt.subplot2grid((3,2), (0, 0), rowspan=3)
plt.plot([-1.4, 1.4], [-1.4*u1[1]/u1[0], 1.4*u1[1]/u1[0]], "k-", linewidth=1)
plt.plot([-1.4, 1.4], [-1.4*u2[1]/u2[0], 1.4*u2[1]/u2[0]], "k--", linewidth=1)
plt.plot([-1.4, 1.4], [-1.4*u3[1]/u3[0], 1.4*u3[1]/u3[0]], "k:", linewidth=2)
plt.plot(X[:, 0], X[:, 1], "bo", alpha=0.5)
plt.axis([-1.4, 1.4, -1.4, 1.4])
plt.arrow(0, 0, u1[0], u1[1], head_width=0.1, linewidth=5, length_includes_head=True, head_length=0.1, fc='k', ec='k')
plt.arrow(0, 0, u3[0], u3[1], head_width=0.1, linewidth=5, length_includes_head=True, head_length=0.1, fc='k', ec='k')
plt.text(u1[0] + 0.1, u1[1] - 0.05, r"$\mathbf{c_1}$", fontsize=22)
plt.text(u3[0] + 0.1, u3[1], r"$\mathbf{c_2}$", fontsize=22)
plt.xlabel("$x_1$", fontsize=18)
plt.ylabel("$x_2$", fontsize=18, rotation=0)
plt.grid(True)

plt.subplot2grid((3,2), (0, 1))
plt.plot([-2, 2], [0, 0], "k-", linewidth=1)
plt.plot(X_proj1[:, 0], np.zeros(m), "bo", alpha=0.3)
plt.gca().get_yaxis().set_ticks([])
plt.gca().get_xaxis().set_ticklabels([])
plt.axis([-2, 2, -1, 1])
plt.grid(True)

plt.subplot2grid((3,2), (1, 1))
plt.plot([-2, 2], [0, 0], "k--", linewidth=1)
plt.plot(X_proj2[:, 0], np.zeros(m), "bo", alpha=0.3)
plt.gca().get_yaxis().set_ticks([])
plt.gca().get_xaxis().set_ticklabels([])
plt.axis([-2, 2, -1, 1])
plt.grid(True)

plt.subplot2grid((3,2), (2, 1))
plt.plot([-2, 2], [0, 0], "k:", linewidth=2)
plt.plot(X_proj3[:, 0], np.zeros(m), "bo", alpha=0.3)
plt.gca().get_yaxis().set_ticks([])
plt.axis([-2, 2, -1, 1])
plt.xlabel("$z_1$", fontsize=18)
plt.grid(True)

save_fig("pca_best_projection")
plt.show()


# # MNIST compression

# In[31]:


import urllib.request
try:
    from sklearn.datasets import fetch_openml
    mnist = fetch_openml('mnist_784', version=1, as_frame=False)
    mnist.target = mnist.target.astype(np.int64)
except ImportError:
    from sklearn.datasets import fetch_mldata
    mnist = fetch_mldata('MNIST original')


# In[32]:


from sklearn.model_selection import train_test_split

X = mnist["data"]
y = mnist["target"]

X_train, X_test, y_train, y_test = train_test_split(X, y)


# In[33]:


pca = PCA()
pca.fit(X_train)
cumsum = np.cumsum(pca.explained_variance_ratio_)
d = np.argmax(cumsum >= 0.95) + 1


# In[34]:


d


# In[35]:


pca = PCA(n_components=0.95)
X_reduced = pca.fit_transform(X_train)


# In[36]:


pca.n_components_


# In[37]:


np.sum(pca.explained_variance_ratio_)


# In[38]:


pca = PCA(n_components = 154)
X_reduced = pca.fit_transform(X_train)
X_recovered = pca.inverse_transform(X_reduced)


# In[39]:


def plot_digits(instances, images_per_row=5, **options):
    size = 28
    images_per_row = min(len(instances), images_per_row)
    images = [instance.reshape(size,size) for instance in instances]
    n_rows = (len(instances) - 1) // images_per_row + 1
    row_images = []
    n_empty = n_rows * images_per_row - len(instances)
    images.append(np.zeros((size, size * n_empty)))
    for row in range(n_rows):
        rimages = images[row * images_per_row : (row + 1) * images_per_row]
        row_images.append(np.concatenate(rimages, axis=1))
    image = np.concatenate(row_images, axis=0)
    plt.imshow(image, cmap = mpl.cm.binary, **options)
    plt.axis("off")


# In[40]:


plt.figure(figsize=(7, 4))
plt.subplot(121)
plot_digits(X_train[::2100])
plt.title("Original", fontsize=16)
plt.subplot(122)
plot_digits(X_recovered[::2100])
plt.title("Compressed", fontsize=16)

save_fig("mnist_compression_plot")


# In[41]:


X_reduced_pca = X_reduced


# ## Incremental PCA

# In[42]:


from sklearn.decomposition import IncrementalPCA

n_batches = 100
inc_pca = IncrementalPCA(n_components=154)
for X_batch in np.array_split(X_train, n_batches):
    print(".", end="") # not shown in the book
    inc_pca.partial_fit(X_batch)

X_reduced = inc_pca.transform(X_train)


# In[43]:


X_recovered_inc_pca = inc_pca.inverse_transform(X_reduced)


# In[44]:


plt.figure(figsize=(7, 4))
plt.subplot(121)
plot_digits(X_train[::2100])
plt.subplot(122)
plot_digits(X_recovered_inc_pca[::2100])
plt.tight_layout()


# In[45]:


X_reduced_inc_pca = X_reduced


# Let's compare the results of transforming MNIST using regular PCA and incremental PCA. First, the means are equal: 

# In[46]:


np.allclose(pca.mean_, inc_pca.mean_)


# But the results are not exactly identical. Incremental PCA gives a very good approximate solution, but it's not perfect:

# In[47]:


np.allclose(X_reduced_pca, X_reduced_inc_pca)


# ### Using `memmap()`

# Let's create the `memmap()` structure and copy the MNIST data into it. This would typically be done by a first program:

# In[48]:


filename = "my_mnist.data"
m, n = X_train.shape

X_mm = np.memmap(filename, dtype='float32', mode='write', shape=(m, n))
X_mm[:] = X_train


# Now deleting the `memmap()` object will trigger its Python finalizer, which ensures that the data is saved to disk.

# In[49]:


del X_mm


# Next, another program would load the data and use it for training:

# In[50]:


X_mm = np.memmap(filename, dtype="float32", mode="readonly", shape=(m, n))

batch_size = m // n_batches
inc_pca = IncrementalPCA(n_components=154, batch_size=batch_size)
inc_pca.fit(X_mm)


# In[51]:


rnd_pca = PCA(n_components=154, svd_solver="randomized", random_state=42)
X_reduced = rnd_pca.fit_transform(X_train)


# ## Time complexity

# Let's time regular PCA against Incremental PCA and Randomized PCA, for various number of principal components:

# In[52]:


import time

for n_components in (2, 10, 154):
    print("n_components =", n_components)
    regular_pca = PCA(n_components=n_components)
    inc_pca = IncrementalPCA(n_components=n_components, batch_size=500)
    rnd_pca = PCA(n_components=n_components, random_state=42, svd_solver="randomized")

    for pca in (regular_pca, inc_pca, rnd_pca):
        t1 = time.time()
        pca.fit(X_train)
        t2 = time.time()
        print("    {}: {:.1f} seconds".format(pca.__class__.__name__, t2 - t1))


# Now let's compare PCA and Randomized PCA for datasets of different sizes (number of instances):

# In[53]:


times_rpca = []
times_pca = []
sizes = [1000, 10000, 20000, 30000, 40000, 50000, 70000, 100000, 200000, 500000]
for n_samples in sizes:
    X = np.random.randn(n_samples, 5)
    pca = PCA(n_components = 2, svd_solver="randomized", random_state=42)
    t1 = time.time()
    pca.fit(X)
    t2 = time.time()
    times_rpca.append(t2 - t1)
    pca = PCA(n_components = 2)
    t1 = time.time()
    pca.fit(X)
    t2 = time.time()
    times_pca.append(t2 - t1)

plt.plot(sizes, times_rpca, "b-o", label="RPCA")
plt.plot(sizes, times_pca, "r-s", label="PCA")
plt.xlabel("n_samples")
plt.ylabel("Training time")
plt.legend(loc="upper left")
plt.title("PCA and Randomized PCA time complexity ")


# And now let's compare their performance on datasets of 2,000 instances with various numbers of features:

# In[54]:


times_rpca = []
times_pca = []
sizes = [1000, 2000, 3000, 4000, 5000, 6000]
for n_features in sizes:
    X = np.random.randn(2000, n_features)
    pca = PCA(n_components = 2, random_state=42, svd_solver="randomized")
    t1 = time.time()
    pca.fit(X)
    t2 = time.time()
    times_rpca.append(t2 - t1)
    pca = PCA(n_components = 2)
    t1 = time.time()
    pca.fit(X)
    t2 = time.time()
    times_pca.append(t2 - t1)

plt.plot(sizes, times_rpca, "b-o", label="RPCA")
plt.plot(sizes, times_pca, "r-s", label="PCA")
plt.xlabel("n_features")
plt.ylabel("Training time")
plt.legend(loc="upper left")
plt.title("PCA and Randomized PCA time complexity ")


# # Kernel PCA

# In[55]:


X, t = make_swiss_roll(n_samples=1000, noise=0.2, random_state=42)


# In[56]:


from sklearn.decomposition import KernelPCA

rbf_pca = KernelPCA(n_components = 2, kernel="rbf", gamma=0.04)
X_reduced = rbf_pca.fit_transform(X)


# In[57]:


from sklearn.decomposition import KernelPCA

lin_pca = KernelPCA(n_components = 2, kernel="linear", fit_inverse_transform=True)
rbf_pca = KernelPCA(n_components = 2, kernel="rbf", gamma=0.0433, fit_inverse_transform=True)
sig_pca = KernelPCA(n_components = 2, kernel="sigmoid", gamma=0.001, coef0=1, fit_inverse_transform=True)

y = t > 6.9

plt.figure(figsize=(11, 4))
for subplot, pca, title in ((131, lin_pca, "Linear kernel"), (132, rbf_pca, "RBF kernel, $\gamma=0.04$"), (133, sig_pca, "Sigmoid kernel, $\gamma=10^{-3}, r=1$")):
    X_reduced = pca.fit_transform(X)
    if subplot == 132:
        X_reduced_rbf = X_reduced
    
    plt.subplot(subplot)
    #plt.plot(X_reduced[y, 0], X_reduced[y, 1], "gs")
    #plt.plot(X_reduced[~y, 0], X_reduced[~y, 1], "y^")
    plt.title(title, fontsize=14)
    plt.scatter(X_reduced[:, 0], X_reduced[:, 1], c=t, cmap=plt.cm.hot)
    plt.xlabel("$z_1$", fontsize=18)
    if subplot == 131:
        plt.ylabel("$z_2$", fontsize=18, rotation=0)
    plt.grid(True)

save_fig("kernel_pca_plot")
plt.show()


# In[58]:


plt.figure(figsize=(6, 5))

X_inverse = rbf_pca.inverse_transform(X_reduced_rbf)

ax = plt.subplot(111, projection='3d')
ax.view_init(10, -70)
ax.scatter(X_inverse[:, 0], X_inverse[:, 1], X_inverse[:, 2], c=t, cmap=plt.cm.hot, marker="x")
ax.set_xlabel("")
ax.set_ylabel("")
ax.set_zlabel("")
ax.set_xticklabels([])
ax.set_yticklabels([])
ax.set_zticklabels([])

save_fig("preimage_plot", tight_layout=False)
plt.show()


# In[59]:


X_reduced = rbf_pca.fit_transform(X)

plt.figure(figsize=(11, 4))
plt.subplot(132)
plt.scatter(X_reduced[:, 0], X_reduced[:, 1], c=t, cmap=plt.cm.hot, marker="x")
plt.xlabel("$z_1$", fontsize=18)
plt.ylabel("$z_2$", fontsize=18, rotation=0)
plt.grid(True)


# In[60]:


from sklearn.model_selection import GridSearchCV
from sklearn.linear_model import LogisticRegression
from sklearn.pipeline import Pipeline

clf = Pipeline([
        ("kpca", KernelPCA(n_components=2)),
        ("log_reg", LogisticRegression(solver="liblinear"))
    ])

param_grid = [{
        "kpca__gamma": np.linspace(0.03, 0.05, 10),
        "kpca__kernel": ["rbf", "sigmoid"]
    }]

grid_search = GridSearchCV(clf, param_grid, cv=3)
grid_search.fit(X, y)


# In[61]:


print(grid_search.best_params_)


# In[62]:


rbf_pca = KernelPCA(n_components = 2, kernel="rbf", gamma=0.0433,
                    fit_inverse_transform=True)
X_reduced = rbf_pca.fit_transform(X)
X_preimage = rbf_pca.inverse_transform(X_reduced)


# In[63]:


from sklearn.metrics import mean_squared_error

mean_squared_error(X, X_preimage)


# # LLE

# In[64]:


X, t = make_swiss_roll(n_samples=1000, noise=0.2, random_state=41)


# In[65]:


from sklearn.manifold import LocallyLinearEmbedding

lle = LocallyLinearEmbedding(n_components=2, n_neighbors=10, random_state=42)
X_reduced = lle.fit_transform(X)


# In[66]:


plt.title("Unrolled swiss roll using LLE", fontsize=14)
plt.scatter(X_reduced[:, 0], X_reduced[:, 1], c=t, cmap=plt.cm.hot)
plt.xlabel("$z_1$", fontsize=18)
plt.ylabel("$z_2$", fontsize=18)
plt.axis([-0.065, 0.055, -0.1, 0.12])
plt.grid(True)

save_fig("lle_unrolling_plot")
plt.show()


# # MDS, Isomap and t-SNE

# In[67]:


from sklearn.manifold import MDS

mds = MDS(n_components=2, random_state=42)
X_reduced_mds = mds.fit_transform(X)


# In[68]:


from sklearn.manifold import Isomap

isomap = Isomap(n_components=2)
X_reduced_isomap = isomap.fit_transform(X)


# In[69]:


from sklearn.manifold import TSNE

tsne = TSNE(n_components=2, random_state=42)
X_reduced_tsne = tsne.fit_transform(X)


# In[70]:


from sklearn.discriminant_analysis import LinearDiscriminantAnalysis

lda = LinearDiscriminantAnalysis(n_components=2)
X_mnist = mnist["data"]
y_mnist = mnist["target"]
lda.fit(X_mnist, y_mnist)
X_reduced_lda = lda.transform(X_mnist)


# In[71]:


titles = ["MDS", "Isomap", "t-SNE"]

plt.figure(figsize=(11,4))

for subplot, title, X_reduced in zip((131, 132, 133), titles,
                                     (X_reduced_mds, X_reduced_isomap, X_reduced_tsne)):
    plt.subplot(subplot)
    plt.title(title, fontsize=14)
    plt.scatter(X_reduced[:, 0], X_reduced[:, 1], c=t, cmap=plt.cm.hot)
    plt.xlabel("$z_1$", fontsize=18)
    if subplot == 131:
        plt.ylabel("$z_2$", fontsize=18, rotation=0)
    plt.grid(True)

save_fig("other_dim_reduction_plot")
plt.show()


# In[72]:


def learned_parameters(model):
    return [m for m in dir(model)
            if m.endswith("_") and not m.startswith("_")]


# # Extra Material – Clustering

# ## Introduction – Classification _vs_ Clustering

# In[73]:


from sklearn.datasets import load_iris


# In[74]:


data = load_iris()
X = data.data
y = data.target
data.target_names


# In[75]:


plt.figure(figsize=(9, 3.5))

plt.subplot(121)
plt.plot(X[y==0, 2], X[y==0, 3], "yo", label="Iris-Setosa")
plt.plot(X[y==1, 2], X[y==1, 3], "bs", label="Iris-Versicolor")
plt.plot(X[y==2, 2], X[y==2, 3], "g^", label="Iris-Virginica")
plt.xlabel("Petal length", fontsize=14)
plt.ylabel("Petal width", fontsize=14)
plt.legend(fontsize=12)

plt.subplot(122)
plt.scatter(X[:, 2], X[:, 3], c="k", marker=".")
plt.xlabel("Petal length", fontsize=14)
plt.tick_params(labelleft=False)

save_fig("classification_vs_clustering_diagram")
plt.show()


# A Gaussian mixture model (explained below) can actually separate these clusters pretty well (using all 4 features: petal length & width, and sepal length & width).

# In[76]:


from sklearn.mixture import GaussianMixture


# In[77]:


y_pred = GaussianMixture(n_components=3, random_state=42).fit(X).predict(X)
mapping = np.array([2, 0, 1])
y_pred = np.array([mapping[cluster_id] for cluster_id in y_pred])


# In[78]:


plt.plot(X[y_pred==0, 2], X[y_pred==0, 3], "yo", label="Cluster 1")
plt.plot(X[y_pred==1, 2], X[y_pred==1, 3], "bs", label="Cluster 2")
plt.plot(X[y_pred==2, 2], X[y_pred==2, 3], "g^", label="Cluster 3")
plt.xlabel("Petal length", fontsize=14)
plt.ylabel("Petal width", fontsize=14)
plt.legend(loc="upper left", fontsize=12)
plt.show()


# In[79]:


np.sum(y_pred==y)


# In[80]:


np.sum(y_pred==y) / len(y_pred)


# ## K-Means

# Let's start by generating some blobs:

# In[81]:


from sklearn.datasets import make_blobs


# In[82]:


blob_centers = np.array(
    [[ 0.2,  2.3],
     [-1.5 ,  2.3],
     [-2.8,  1.8],
     [-2.8,  2.8],
     [-2.8,  1.3]])
blob_std = np.array([0.4, 0.3, 0.1, 0.1, 0.1])


# In[83]:


X, y = make_blobs(n_samples=2000, centers=blob_centers,
                  cluster_std=blob_std, random_state=7)


# Now let's plot them:

# In[84]:


def plot_clusters(X, y=None):
    plt.scatter(X[:, 0], X[:, 1], c=y, s=1)
    plt.xlabel("$x_1$", fontsize=14)
    plt.ylabel("$x_2$", fontsize=14, rotation=0)


# In[85]:


plt.figure(figsize=(8, 4))
plot_clusters(X)
save_fig("blobs_diagram")
plt.show()


# ### Fit and Predict

# Let's train a K-Means clusterer on this dataset. It will try to find each blob's center and assign each instance to the closest blob:

# In[86]:


from sklearn.cluster import KMeans


# In[87]:


k = 5
kmeans = KMeans(n_clusters=k, random_state=42)
y_pred = kmeans.fit_predict(X)


# Each instance was assigned to one of the 5 clusters:

# In[88]:


y_pred


# In[89]:


y_pred is kmeans.labels_


# And the following 5 _centroids_ (i.e., cluster centers) were estimated:

# In[90]:


kmeans.cluster_centers_


# Note that the `KMeans` instance preserves the labels of the instances it was trained on. Somewhat confusingly, in this context, the _label_ of an instance is the index of the cluster that instance gets assigned to:

# In[91]:


kmeans.labels_


# Of course, we can predict the labels of new instances:

# In[92]:


X_new = np.array([[0, 2], [3, 2], [-3, 3], [-3, 2.5]])
kmeans.predict(X_new)


# ### Decision Boundaries

# Let's plot the model's decision boundaries. This gives us a _Voronoi diagram_:

# In[93]:


def plot_data(X):
    plt.plot(X[:, 0], X[:, 1], 'k.', markersize=2)

def plot_centroids(centroids, weights=None, circle_color='w', cross_color='k'):
    if weights is not None:
        centroids = centroids[weights > weights.max() / 10]
    plt.scatter(centroids[:, 0], centroids[:, 1],
                marker='o', s=30, linewidths=8,
                color=circle_color, zorder=10, alpha=0.9)
    plt.scatter(centroids[:, 0], centroids[:, 1],
                marker='x', s=50, linewidths=50,
                color=cross_color, zorder=11, alpha=1)

def plot_decision_boundaries(clusterer, X, resolution=1000, show_centroids=True,
                             show_xlabels=True, show_ylabels=True):
    mins = X.min(axis=0) - 0.1
    maxs = X.max(axis=0) + 0.1
    xx, yy = np.meshgrid(np.linspace(mins[0], maxs[0], resolution),
                         np.linspace(mins[1], maxs[1], resolution))
    Z = clusterer.predict(np.c_[xx.ravel(), yy.ravel()])
    Z = Z.reshape(xx.shape)

    plt.contourf(Z, extent=(mins[0], maxs[0], mins[1], maxs[1]),
                cmap="Pastel2")
    plt.contour(Z, extent=(mins[0], maxs[0], mins[1], maxs[1]),
                linewidths=1, colors='k')
    plot_data(X)
    if show_centroids:
        plot_centroids(clusterer.cluster_centers_)

    if show_xlabels:
        plt.xlabel("$x_1$", fontsize=14)
    else:
        plt.tick_params(labelbottom=False)
    if show_ylabels:
        plt.ylabel("$x_2$", fontsize=14, rotation=0)
    else:
        plt.tick_params(labelleft=False)


# In[94]:


plt.figure(figsize=(8, 4))
plot_decision_boundaries(kmeans, X)
save_fig("voronoi_diagram")
plt.show()


# Not bad! Some of the instances near the edges were probably assigned to the wrong cluster, but overall it looks pretty good.

# ### Hard Clustering _vs_ Soft Clustering

# Rather than arbitrarily choosing the closest cluster for each instance, which is called _hard clustering_, it might be better measure the distance of each instance to all 5 centroids. This is what the `transform()` method does:

# In[95]:


kmeans.transform(X_new)


# You can verify that this is indeed the Euclidian distance between each instance and each centroid:

# In[96]:


np.linalg.norm(np.tile(X_new, (1, k)).reshape(-1, k, 2) - kmeans.cluster_centers_, axis=2)


# ### K-Means Algorithm

# The K-Means algorithm is one of the fastest clustering algorithms, but also one of the simplest:
# * First initialize $k$ centroids randomly: $k$ distinct instances are chosen randomly from the dataset and the centroids are placed at their locations.
# * Repeat until convergence (i.e., until the centroids stop moving):
#     * Assign each instance to the closest centroid.
#     * Update the centroids to be the mean of the instances that are assigned to them.

# The `KMeans` class applies an optimized algorithm by default. To get the original K-Means algorithm (for educational purposes only), you must set `init="random"`, `n_init=1`and `algorithm="full"`. These hyperparameters will be explained below.

# Let's run the K-Means algorithm for 1, 2 and 3 iterations, to see how the centroids move around:

# In[97]:


kmeans_iter1 = KMeans(n_clusters=5, init="random", n_init=1,
                     algorithm="full", max_iter=1, random_state=1)
kmeans_iter2 = KMeans(n_clusters=5, init="random", n_init=1,
                     algorithm="full", max_iter=2, random_state=1)
kmeans_iter3 = KMeans(n_clusters=5, init="random", n_init=1,
                     algorithm="full", max_iter=3, random_state=1)
kmeans_iter1.fit(X)
kmeans_iter2.fit(X)
kmeans_iter3.fit(X)


# And let's plot this:

# In[98]:


plt.figure(figsize=(10, 8))

plt.subplot(321)
plot_data(X)
plot_centroids(kmeans_iter1.cluster_centers_, circle_color='r', cross_color='w')
plt.ylabel("$x_2$", fontsize=14, rotation=0)
plt.tick_params(labelbottom=False)
plt.title("Update the centroids (initially randomly)", fontsize=14)

plt.subplot(322)
plot_decision_boundaries(kmeans_iter1, X, show_xlabels=False, show_ylabels=False)
plt.title("Label the instances", fontsize=14)

plt.subplot(323)
plot_decision_boundaries(kmeans_iter1, X, show_centroids=False, show_xlabels=False)
plot_centroids(kmeans_iter2.cluster_centers_)

plt.subplot(324)
plot_decision_boundaries(kmeans_iter2, X, show_xlabels=False, show_ylabels=False)

plt.subplot(325)
plot_decision_boundaries(kmeans_iter2, X, show_centroids=False)
plot_centroids(kmeans_iter3.cluster_centers_)

plt.subplot(326)
plot_decision_boundaries(kmeans_iter3, X, show_ylabels=False)

save_fig("kmeans_algorithm_diagram")
plt.show()


# ### K-Means Variability

# In the original K-Means algorithm, the centroids are just initialized randomly, and the algorithm simply runs a single iteration to gradually improve the centroids, as we saw above.
# 
# However, one major problem with this approach is that if you run K-Means multiple times (or with different random seeds), it can converge to very different solutions, as you can see below:

# In[99]:


def plot_clusterer_comparison(clusterer1, clusterer2, X, title1=None, title2=None):
    clusterer1.fit(X)
    clusterer2.fit(X)

    plt.figure(figsize=(10, 3.2))

    plt.subplot(121)
    plot_decision_boundaries(clusterer1, X)
    if title1:
        plt.title(title1, fontsize=14)

    plt.subplot(122)
    plot_decision_boundaries(clusterer2, X, show_ylabels=False)
    if title2:
        plt.title(title2, fontsize=14)


# In[100]:


kmeans_rnd_init1 = KMeans(n_clusters=5, init="random", n_init=1,
                         algorithm="full", random_state=11)
kmeans_rnd_init2 = KMeans(n_clusters=5, init="random", n_init=1,
                         algorithm="full", random_state=19)

plot_clusterer_comparison(kmeans_rnd_init1, kmeans_rnd_init2, X,
                          "Solution 1", "Solution 2 (with a different random init)")

save_fig("kmeans_variability_diagram")
plt.show()


# ### Inertia

# To select the best model, we will need a way to evaluate a K-Mean model's performance. Unfortunately, clustering is an unsupervised task, so we do not have the targets. But at least we can measure the distance between each instance and its centroid. This is the idea behind the _inertia_ metric:

# In[101]:


kmeans.inertia_


# As you can easily verify, inertia is the sum of the squared distances between each training instance and its closest centroid:

# In[102]:


X_dist = kmeans.transform(X)
np.sum(X_dist[np.arange(len(X_dist)), kmeans.labels_]**2)


# The `score()` method returns the negative inertia. Why negative? Well, it is because a predictor's `score()` method must always respect the "_great is better_" rule.

# In[103]:


kmeans.score(X)


# ### Multiple Initializations

# So one approach to solve the variability issue is to simply run the K-Means algorithm multiple times with different random initializations, and select the solution that minimizes the inertia. For example, here are the inertias of the two "bad" models shown in the previous figure:

# In[104]:


kmeans_rnd_init1.inertia_


# In[105]:


kmeans_rnd_init2.inertia_


# As you can see, they have a higher inertia than the first "good" model we trained, which means they are probably worse.

# When you set the `n_init` hyperparameter, Scikit-Learn runs the original algorithm `n_init` times, and selects the solution that minimizes the inertia. By default, Scikit-Learn sets `n_init=10`.

# In[106]:


kmeans_rnd_10_inits = KMeans(n_clusters=5, init="random", n_init=10,
                              algorithm="full", random_state=11)
kmeans_rnd_10_inits.fit(X)


# As you can see, we end up with the initial model, which is certainly the optimal K-Means solution (at least in terms of inertia, and assuming $k=5$).

# In[107]:


plt.figure(figsize=(8, 4))
plot_decision_boundaries(kmeans_rnd_10_inits, X)
plt.show()


# ### K-Means++

# Instead of initializing the centroids entirely randomly, it is preferable to initialize them using the following algorithm, proposed in a [2006 paper](https://goo.gl/eNUPw6) by David Arthur and Sergei Vassilvitskii:
# * Take one centroid $c_1$, chosen uniformly at random from the dataset.
# * Take a new center $c_i$, choosing an instance $\mathbf{x}_i$ with probability: $D(\mathbf{x}_i)^2$ / $\sum\limits_{j=1}^{m}{D(\mathbf{x}_j)}^2$ where $D(\mathbf{x}_i)$ is the distance between the instance $\mathbf{x}_i$ and the closest centroid that was already chosen. This probability distribution ensures that instances that are further away from already chosen centroids are much more likely be selected as centroids.
# * Repeat the previous step until all $k$ centroids have been chosen.

# The rest of the K-Means++ algorithm is just regular K-Means. With this initialization, the K-Means algorithm is much less likely to converge to a suboptimal solution, so it is possible to reduce `n_init` considerably. Most of the time, this largely compensates for the additional complexity of the initialization process.

# To set the initialization to K-Means++, simply set `init="k-means++"` (this is actually the default):

# In[108]:


KMeans()


# In[109]:


good_init = np.array([[-3, 3], [-3, 2], [-3, 1], [-1, 2], [0, 2]])
kmeans = KMeans(n_clusters=5, init=good_init, n_init=1, random_state=42)
kmeans.fit(X)
kmeans.inertia_


# ### Accelerated K-Means

# The K-Means algorithm can be significantly accelerated by avoiding many unnecessary distance calculations: this is achieved by exploiting the triangle inequality (given three points A, B and C, the distance AC is always such that AC ≤ AB + BC) and by keeping track of lower and upper bounds for distances between instances and centroids (see this [2003 paper](https://www.aaai.org/Papers/ICML/2003/ICML03-022.pdf) by Charles Elkan for more details).

# To use Elkan's variant of K-Means, just set `algorithm="elkan"`. Note that it does not support sparse data, so by default, Scikit-Learn uses `"elkan"` for dense data, and `"full"` (the regular K-Means algorithm) for sparse data.

# In[110]:


get_ipython().run_line_magic('timeit', '-n 50 KMeans(algorithm="elkan").fit(X)')


# In[111]:


get_ipython().run_line_magic('timeit', '-n 50 KMeans(algorithm="full").fit(X)')


# ### Mini-Batch K-Means

# Scikit-Learn also implements a variant of the K-Means algorithm that supports mini-batches (see [this paper](http://www.eecs.tufts.edu/~dsculley/papers/fastkmeans.pdf)):

# In[112]:


from sklearn.cluster import MiniBatchKMeans


# In[113]:


minibatch_kmeans = MiniBatchKMeans(n_clusters=5, random_state=42)
minibatch_kmeans.fit(X)


# In[114]:


minibatch_kmeans.inertia_


# If the dataset does not fit in memory, the simplest option is to use the `memmap` class, just like we did for incremental PCA:

# In[115]:


filename = "my_mnist.data"
m, n = 50000, 28*28
X_mm = np.memmap(filename, dtype="float32", mode="readonly", shape=(m, n))


# In[116]:


minibatch_kmeans = MiniBatchKMeans(n_clusters=10, batch_size=10, random_state=42)
minibatch_kmeans.fit(X_mm)


# If your data is so large that you cannot use `memmap`, things get more complicated. Let's start by writing a function to load the next batch (in real life, you would load the data from disk):

# In[117]:


def load_next_batch(batch_size):
    return X[np.random.choice(len(X), batch_size, replace=False)]


# Now we can train the model by feeding it one batch at a time. We also need to implement multiple initializations and keep the model with the lowest inertia:

# In[118]:


np.random.seed(42)


# In[119]:


k = 5
n_init = 10
n_iterations = 100
batch_size = 100
init_size = 500  # more data for K-Means++ initialization
evaluate_on_last_n_iters = 10

best_kmeans = None

for init in range(n_init):
    minibatch_kmeans = MiniBatchKMeans(n_clusters=k, init_size=init_size)
    X_init = load_next_batch(init_size)
    minibatch_kmeans.partial_fit(X_init)

    minibatch_kmeans.sum_inertia_ = 0
    for iteration in range(n_iterations):
        X_batch = load_next_batch(batch_size)
        minibatch_kmeans.partial_fit(X_batch)
        if iteration >= n_iterations - evaluate_on_last_n_iters:
            minibatch_kmeans.sum_inertia_ += minibatch_kmeans.inertia_

    if (best_kmeans is None or
        minibatch_kmeans.sum_inertia_ < best_kmeans.sum_inertia_):
        best_kmeans = minibatch_kmeans


# In[120]:


best_kmeans.score(X)


# Mini-batch K-Means is much faster than regular K-Means:

# In[121]:


get_ipython().run_line_magic('timeit', 'KMeans(n_clusters=5).fit(X)')


# In[122]:


get_ipython().run_line_magic('timeit', 'MiniBatchKMeans(n_clusters=5).fit(X)')


# That's *much* faster! However, its performance is often lower (higher inertia), and it keeps degrading as _k_ increases. Let's plot the inertia ratio and the training time ratio between Mini-batch K-Means and regular K-Means:

# In[123]:


from timeit import timeit


# In[124]:


times = np.empty((100, 2))
inertias = np.empty((100, 2))
for k in range(1, 101):
    kmeans = KMeans(n_clusters=k, random_state=42)
    minibatch_kmeans = MiniBatchKMeans(n_clusters=k, random_state=42)
    print("\r{}/{}".format(k, 100), end="")
    times[k-1, 0] = timeit("kmeans.fit(X)", number=10, globals=globals())
    times[k-1, 1]  = timeit("minibatch_kmeans.fit(X)", number=10, globals=globals())
    inertias[k-1, 0] = kmeans.inertia_
    inertias[k-1, 1] = minibatch_kmeans.inertia_


# In[125]:


plt.figure(figsize=(10,4))

plt.subplot(121)
plt.plot(range(1, 101), inertias[:, 0], "r--", label="K-Means")
plt.plot(range(1, 101), inertias[:, 1], "b.-", label="Mini-batch K-Means")
plt.xlabel("$k$", fontsize=16)
#plt.ylabel("Inertia", fontsize=14)
plt.title("Inertia", fontsize=14)
plt.legend(fontsize=14)
plt.axis([1, 100, 0, 100])

plt.subplot(122)
plt.plot(range(1, 101), times[:, 0], "r--", label="K-Means")
plt.plot(range(1, 101), times[:, 1], "b.-", label="Mini-batch K-Means")
plt.xlabel("$k$", fontsize=16)
#plt.ylabel("Training time (seconds)", fontsize=14)
plt.title("Training time (seconds)", fontsize=14)
plt.axis([1, 100, 0, 6])
#plt.legend(fontsize=14)

save_fig("minibatch_kmeans_vs_kmeans")
plt.show()


# ### Finding the optimal number of clusters

# What if the number of clusters was set to a lower or greater value than 5?

# In[126]:


kmeans_k3 = KMeans(n_clusters=3, random_state=42)
kmeans_k8 = KMeans(n_clusters=8, random_state=42)

plot_clusterer_comparison(kmeans_k3, kmeans_k8, X, "$k=3$", "$k=8$")
save_fig("bad_n_clusters_diagram")
plt.show()


# Ouch, these two models don't look great. What about their inertias?

# In[127]:


kmeans_k3.inertia_


# In[128]:


kmeans_k8.inertia_


# No, we cannot simply take the value of $k$ that minimizes the inertia, since it keeps getting lower as we increase $k$. Indeed, the more clusters there are, the closer each instance will be to its closest centroid, and therefore the lower the inertia will be. However, we can plot the inertia as a function of $k$ and analyze the resulting curve:

# In[129]:


kmeans_per_k = [KMeans(n_clusters=k, random_state=42).fit(X)
                for k in range(1, 10)]
inertias = [model.inertia_ for model in kmeans_per_k]


# In[130]:


plt.figure(figsize=(8, 3.5))
plt.plot(range(1, 10), inertias, "bo-")
plt.xlabel("$k$", fontsize=14)
plt.ylabel("Inertia", fontsize=14)
plt.annotate('Elbow',
             xy=(4, inertias[3]),
             xytext=(0.55, 0.55),
             textcoords='figure fraction',
             fontsize=16,
             arrowprops=dict(facecolor='black', shrink=0.1)
            )
plt.axis([1, 8.5, 0, 1300])
save_fig("inertia_vs_k_diagram")
plt.show()


# As you can see, there is an elbow at $k=4$, which means that less clusters than that would be bad, and more clusters would not help much and might cut clusters in half. So $k=4$ is a pretty good choice. Of course in this example it is not perfect since it means that the two blobs in the lower left will be considered as just a single cluster, but it's a pretty good clustering nonetheless.

# In[131]:


plot_decision_boundaries(kmeans_per_k[4-1], X)
plt.show()


# Another approach is to look at the _silhouette score_, which is the mean _silhouette coefficient_ over all the instances. An instance's silhouette coefficient is equal to $(b - a)/\max(a, b)$ where $a$ is the mean distance to the other instances in the same cluster (it is the _mean intra-cluster distance_), and $b$ is the _mean nearest-cluster distance_, that is the mean distance to the instances of the next closest cluster (defined as the one that minimizes $b$, excluding the instance's own cluster). The silhouette coefficient can vary between -1 and +1: a coefficient close to +1 means that the instance is well inside its own cluster and far from other clusters, while a coefficient close to 0 means that it is close to a cluster boundary, and finally a coefficient close to -1 means that the instance may have been assigned to the wrong cluster.

# Let's plot the silhouette score as a function of $k$:

# In[132]:


from sklearn.metrics import silhouette_score


# In[133]:


silhouette_score(X, kmeans.labels_)


# In[134]:


silhouette_scores = [silhouette_score(X, model.labels_)
                     for model in kmeans_per_k[1:]]


# In[135]:


plt.figure(figsize=(8, 3))
plt.plot(range(2, 10), silhouette_scores, "bo-")
plt.xlabel("$k$", fontsize=14)
plt.ylabel("Silhouette score", fontsize=14)
plt.axis([1.8, 8.5, 0.55, 0.7])
save_fig("silhouette_score_vs_k_diagram")
plt.show()


# As you can see, this visualization is much richer than the previous one: in particular, although it confirms that $k=4$ is a very good choice, but it also underlines the fact that $k=5$ is quite good as well.

# An even more informative visualization is given when you plot every instance's silhouette coefficient, sorted by the cluster they are assigned to and by the value of the coefficient. This is called a _silhouette diagram_:

# In[136]:


from sklearn.metrics import silhouette_samples
from matplotlib.ticker import FixedLocator, FixedFormatter

plt.figure(figsize=(11, 9))

for k in (3, 4, 5, 6):
    plt.subplot(2, 2, k - 2)
    
    y_pred = kmeans_per_k[k - 1].labels_
    silhouette_coefficients = silhouette_samples(X, y_pred)

    padding = len(X) // 30
    pos = padding
    ticks = []
    for i in range(k):
        coeffs = silhouette_coefficients[y_pred == i]
        coeffs.sort()

        color = mpl.cm.Spectral(i / k)
        plt.fill_betweenx(np.arange(pos, pos + len(coeffs)), 0, coeffs,
                          facecolor=color, edgecolor=color, alpha=0.7)
        ticks.append(pos + len(coeffs) // 2)
        pos += len(coeffs) + padding

    plt.gca().yaxis.set_major_locator(FixedLocator(ticks))
    plt.gca().yaxis.set_major_formatter(FixedFormatter(range(k)))
    if k in (3, 5):
        plt.ylabel("Cluster")
    
    if k in (5, 6):
        plt.gca().set_xticks([-0.1, 0, 0.2, 0.4, 0.6, 0.8, 1])
        plt.xlabel("Silhouette Coefficient")
    else:
        plt.tick_params(labelbottom=False)

    plt.axvline(x=silhouette_scores[k - 2], color="red", linestyle="--")
    plt.title("$k={}$".format(k), fontsize=16)

save_fig("silhouette_analysis_diagram")
plt.show()


# ### Limits of K-Means

# In[137]:


X1, y1 = make_blobs(n_samples=1000, centers=((4, -4), (0, 0)), random_state=42)
X1 = X1.dot(np.array([[0.374, 0.95], [0.732, 0.598]]))
X2, y2 = make_blobs(n_samples=250, centers=1, random_state=42)
X2 = X2 + [6, -8]
X = np.r_[X1, X2]
y = np.r_[y1, y2]


# In[138]:


plot_clusters(X)


# In[139]:


kmeans_good = KMeans(n_clusters=3, init=np.array([[-1.5, 2.5], [0.5, 0], [4, 0]]), n_init=1, random_state=42)
kmeans_bad = KMeans(n_clusters=3, random_state=42)
kmeans_good.fit(X)
kmeans_bad.fit(X)


# In[140]:


plt.figure(figsize=(10, 3.2))

plt.subplot(121)
plot_decision_boundaries(kmeans_good, X)
plt.title("Inertia = {:.1f}".format(kmeans_good.inertia_), fontsize=14)

plt.subplot(122)
plot_decision_boundaries(kmeans_bad, X, show_ylabels=False)
plt.title("Inertia = {:.1f}".format(kmeans_bad.inertia_), fontsize=14)

save_fig("bad_kmeans_diagram")
plt.show()


# ### Using clustering for image segmentation

# In[141]:


# Download the ladybug image
images_path = os.path.join(PROJECT_ROOT_DIR, "images", "unsupervised_learning")
os.makedirs(images_path, exist_ok=True)
DOWNLOAD_ROOT = "https://raw.githubusercontent.com/ageron/handson-ml2/master/"
filename = "ladybug.png"
print("Downloading", filename)
url = DOWNLOAD_ROOT + "images/unsupervised_learning/" + filename
urllib.request.urlretrieve(url, os.path.join(images_path, filename))


# In[142]:


from matplotlib.image import imread
image = imread(os.path.join("images","unsupervised_learning","ladybug.png"))
image.shape


# In[143]:


X = image.reshape(-1, 3)
kmeans = KMeans(n_clusters=8, random_state=42).fit(X)
segmented_img = kmeans.cluster_centers_[kmeans.labels_]
segmented_img = segmented_img.reshape(image.shape)


# In[144]:


segmented_imgs = []
n_colors = (10, 8, 6, 4, 2)
for n_clusters in n_colors:
    kmeans = KMeans(n_clusters=n_clusters, random_state=42).fit(X)
    segmented_img = kmeans.cluster_centers_[kmeans.labels_]
    segmented_imgs.append(segmented_img.reshape(image.shape))


# In[145]:


plt.figure(figsize=(10,5))
plt.subplots_adjust(wspace=0.05, hspace=0.1)

plt.subplot(231)
plt.imshow(image)
plt.title("Original image")
plt.axis('off')

for idx, n_clusters in enumerate(n_colors):
    plt.subplot(232 + idx)
    plt.imshow(segmented_imgs[idx])
    plt.title("{} colors".format(n_clusters))
    plt.axis('off')

save_fig('image_segmentation_diagram', tight_layout=False)
plt.show()


# ### Using Clustering for Preprocessing

# Let's tackle the _digits dataset_ which is a simple MNIST-like dataset containing 1,797 grayscale 8×8 images representing digits 0 to 9.

# In[146]:


from sklearn.datasets import load_digits


# In[147]:


X_digits, y_digits = load_digits(return_X_y=True)


# Let's split it into a training set and a test set:

# In[148]:


from sklearn.model_selection import train_test_split


# In[149]:


X_train, X_test, y_train, y_test = train_test_split(X_digits, y_digits, random_state=42)


# Now let's fit a Logistic Regression model and evaluate it on the test set:

# In[150]:


from sklearn.linear_model import LogisticRegression


# In[151]:


log_reg = LogisticRegression(multi_class="ovr", solver="liblinear", random_state=42)
log_reg.fit(X_train, y_train)


# In[152]:


log_reg.score(X_test, y_test)


# Okay, that's our baseline: 96.7% accuracy. Let's see if we can do better by using K-Means as a preprocessing step. We will create a pipeline that will first cluster the training set into 50 clusters and replace the images with their distances to the 50 clusters, then apply a logistic regression model:

# In[153]:


from sklearn.pipeline import Pipeline


# In[154]:


pipeline = Pipeline([
    ("kmeans", KMeans(n_clusters=50, random_state=42)),
    ("log_reg", LogisticRegression(multi_class="ovr", solver="liblinear", random_state=42)),
])
pipeline.fit(X_train, y_train)


# In[155]:


pipeline.score(X_test, y_test)


# In[156]:


1 - (1 - 0.9822222) / (1 - 0.9666666)


# How about that? We almost divided the error rate by a factor of 2! But we chose the number of clusters $k$ completely arbitrarily, we can surely do better. Since K-Means is just a preprocessing step in a classification pipeline, finding a good value for $k$ is much simpler than earlier: there's no need to perform silhouette analysis or minimize the inertia, the best value of $k$ is simply the one that results in the best classification performance.

# In[157]:


from sklearn.model_selection import GridSearchCV


# In[158]:


param_grid = dict(kmeans__n_clusters=range(2, 100))
grid_clf = GridSearchCV(pipeline, param_grid, cv=3, verbose=2)
grid_clf.fit(X_train, y_train)


# In[159]:


grid_clf.best_params_


# In[160]:


grid_clf.score(X_test, y_test)


# The performance is slightly improved when $k=90$, so 90 it is.

# ### Clustering for Semi-supervised Learning

# Another use case for clustering is in semi-supervised learning, when we have plenty of unlabeled instances and very few labeled instances.

# Let's look at the performance of a logistic regression model when we only have 50 labeled instances:

# In[161]:


n_labeled = 50


# In[162]:


log_reg = LogisticRegression(multi_class="ovr", solver="liblinear", random_state=42)
log_reg.fit(X_train[:n_labeled], y_train[:n_labeled])
log_reg.score(X_test, y_test)


# It's much less than earlier of course. Let's see how we can do better. First, let's cluster the training set into 50 clusters, then for each cluster let's find the image closest to the centroid. We will call these images the representative images:

# In[163]:


k = 50


# In[164]:


kmeans = KMeans(n_clusters=k, random_state=42)
X_digits_dist = kmeans.fit_transform(X_train)
representative_digit_idx = np.argmin(X_digits_dist, axis=0)
X_representative_digits = X_train[representative_digit_idx]


# Now let's plot these representative images and label them manually:

# In[165]:


plt.figure(figsize=(8, 2))
for index, X_representative_digit in enumerate(X_representative_digits):
    plt.subplot(k // 10, 10, index + 1)
    plt.imshow(X_representative_digit.reshape(8, 8), cmap="binary", interpolation="bilinear")
    plt.axis('off')

save_fig("representative_images_diagram", tight_layout=False)
plt.show()


# In[166]:


y_representative_digits = np.array([
    4, 8, 0, 6, 8, 3, 7, 7, 9, 2,
    5, 5, 8, 5, 2, 1, 2, 9, 6, 1,
    1, 6, 9, 0, 8, 3, 0, 7, 4, 1,
    6, 5, 2, 4, 1, 8, 6, 3, 9, 2,
    4, 2, 9, 4, 7, 6, 2, 3, 1, 1])


# Now we have a dataset with just 50 labeled instances, but instead of being completely random instances, each of them is a representative image of its cluster. Let's see if the performance is any better:

# In[167]:


log_reg = LogisticRegression(multi_class="ovr", solver="liblinear", random_state=42)
log_reg.fit(X_representative_digits, y_representative_digits)
log_reg.score(X_test, y_test)


# Wow! We jumped from 82.7% accuracy to 92.4%, although we are still only training the model on 50 instances. Since it's often costly and painful to label instances, especially when it has to be done manually by experts, it's a good idea to make them label representative instances rather than just random instances.

# But perhaps we can go one step further: what if we propagated the labels to all the other instances in the same cluster?

# In[168]:


y_train_propagated = np.empty(len(X_train), dtype=np.int32)
for i in range(k):
    y_train_propagated[kmeans.labels_==i] = y_representative_digits[i]


# In[169]:


log_reg = LogisticRegression(multi_class="ovr", solver="liblinear", random_state=42)
log_reg.fit(X_train, y_train_propagated)


# In[170]:


log_reg.score(X_test, y_test)


# We got a tiny little accuracy boost. Better than nothing, but we should probably have propagated the labels only to the instances closest to the centroid, because by propagating to the full cluster, we have certainly included some outliers. Let's only propagate the labels to the 20th percentile closest to the centroid:

# In[171]:


percentile_closest = 20

X_cluster_dist = X_digits_dist[np.arange(len(X_train)), kmeans.labels_]
for i in range(k):
    in_cluster = (kmeans.labels_ == i)
    cluster_dist = X_cluster_dist[in_cluster]
    cutoff_distance = np.percentile(cluster_dist, percentile_closest)
    above_cutoff = (X_cluster_dist > cutoff_distance)
    X_cluster_dist[in_cluster & above_cutoff] = -1


# In[172]:


partially_propagated = (X_cluster_dist != -1)
X_train_partially_propagated = X_train[partially_propagated]
y_train_partially_propagated = y_train_propagated[partially_propagated]


# In[173]:


log_reg = LogisticRegression(multi_class="ovr", solver="liblinear", random_state=42)
log_reg.fit(X_train_partially_propagated, y_train_partially_propagated)


# In[174]:


log_reg.score(X_test, y_test)


# Nice! With just 50 labeled instances (just 5 examples per class on average!), we got 94.2% performance, which is pretty close to the performance of logistic regression on the fully labeled _digits_ dataset (which was 96.7%).

# This is because the propagated labels are actually pretty good: their accuracy is very close to 99%:

# In[175]:


np.mean(y_train_partially_propagated == y_train[partially_propagated])


# You could now do a few iterations of _active learning_:
# 1. Manually label the instances that the classifier is least sure about, if possible by picking them in distinct clusters.
# 2. Train a new model with these additional labels.

# ## DBSCAN

# In[176]:


from sklearn.datasets import make_moons


# In[177]:


X, y = make_moons(n_samples=1000, noise=0.05, random_state=42)


# In[178]:


from sklearn.cluster import DBSCAN


# In[179]:


dbscan = DBSCAN(eps=0.05, min_samples=5)
dbscan.fit(X)


# In[180]:


dbscan.labels_[:10]


# In[181]:


len(dbscan.core_sample_indices_)


# In[182]:


dbscan.core_sample_indices_[:10]


# In[183]:


dbscan.components_[:3]


# In[184]:


np.unique(dbscan.labels_)


# In[185]:


dbscan2 = DBSCAN(eps=0.2)
dbscan2.fit(X)


# In[186]:


def plot_dbscan(dbscan, X, size, show_xlabels=True, show_ylabels=True):
    core_mask = np.zeros_like(dbscan.labels_, dtype=bool)
    core_mask[dbscan.core_sample_indices_] = True
    anomalies_mask = dbscan.labels_ == -1
    non_core_mask = ~(core_mask | anomalies_mask)

    cores = dbscan.components_
    anomalies = X[anomalies_mask]
    non_cores = X[non_core_mask]
    
    plt.scatter(cores[:, 0], cores[:, 1],
                c=dbscan.labels_[core_mask], marker='o', s=size, cmap="Paired")
    plt.scatter(cores[:, 0], cores[:, 1], marker='*', s=20, c=dbscan.labels_[core_mask])
    plt.scatter(anomalies[:, 0], anomalies[:, 1],
                c="r", marker="x", s=100)
    plt.scatter(non_cores[:, 0], non_cores[:, 1], c=dbscan.labels_[non_core_mask], marker=".")
    if show_xlabels:
        plt.xlabel("$x_1$", fontsize=14)
    else:
        plt.tick_params(labelbottom=False)
    if show_ylabels:
        plt.ylabel("$x_2$", fontsize=14, rotation=0)
    else:
        plt.tick_params(labelleft=False)
    plt.title("eps={:.2f}, min_samples={}".format(dbscan.eps, dbscan.min_samples), fontsize=14)


# In[187]:


plt.figure(figsize=(9, 3.2))

plt.subplot(121)
plot_dbscan(dbscan, X, size=100)

plt.subplot(122)
plot_dbscan(dbscan2, X, size=600, show_ylabels=False)

save_fig("dbscan_diagram")
plt.show()


# In[188]:


dbscan = dbscan2


# In[189]:


from sklearn.neighbors import KNeighborsClassifier


# In[190]:


knn = KNeighborsClassifier(n_neighbors=50)
knn.fit(dbscan.components_, dbscan.labels_[dbscan.core_sample_indices_])


# In[191]:


X_new = np.array([[-0.5, 0], [0, 0.5], [1, -0.1], [2, 1]])
knn.predict(X_new)


# In[192]:


knn.predict_proba(X_new)


# In[193]:


plt.figure(figsize=(6, 3))
plot_decision_boundaries(knn, X, show_centroids=False)
plt.scatter(X_new[:, 0], X_new[:, 1], c="b", marker="+", s=200, zorder=10)
save_fig("cluster_classification_diagram")
plt.show()


# In[194]:


y_dist, y_pred_idx = knn.kneighbors(X_new, n_neighbors=1)
y_pred = dbscan.labels_[dbscan.core_sample_indices_][y_pred_idx]
y_pred[y_dist > 0.2] = -1
y_pred.ravel()


# ## Other Clustering Algorithms

# ### Spectral Clustering

# In[195]:


from sklearn.cluster import SpectralClustering


# In[196]:


sc1 = SpectralClustering(n_clusters=2, gamma=100, random_state=42)
sc1.fit(X)


# In[197]:


sc2 = SpectralClustering(n_clusters=2, gamma=1, random_state=42)
sc2.fit(X)


# In[198]:


np.percentile(sc1.affinity_matrix_, 95)


# In[199]:


def plot_spectral_clustering(sc, X, size, alpha, show_xlabels=True, show_ylabels=True):
    plt.scatter(X[:, 0], X[:, 1], marker='o', s=size, c='gray', cmap="Paired", alpha=alpha)
    plt.scatter(X[:, 0], X[:, 1], marker='o', s=30, c='w')
    plt.scatter(X[:, 0], X[:, 1], marker='.', s=10, c=sc.labels_, cmap="Paired")
    
    if show_xlabels:
        plt.xlabel("$x_1$", fontsize=14)
    else:
        plt.tick_params(labelbottom=False)
    if show_ylabels:
        plt.ylabel("$x_2$", fontsize=14, rotation=0)
    else:
        plt.tick_params(labelleft=False)
    plt.title("RBF gamma={}".format(sc.gamma), fontsize=14)


# In[200]:


plt.figure(figsize=(9, 3.2))

plt.subplot(121)
plot_spectral_clustering(sc1, X, size=500, alpha=0.1)

plt.subplot(122)
plot_spectral_clustering(sc2, X, size=4000, alpha=0.01, show_ylabels=False)

plt.show()


# ### Agglomerative Clustering

# In[201]:


from sklearn.cluster import AgglomerativeClustering


# In[202]:


X = np.array([0, 2, 5, 8.5]).reshape(-1, 1)
agg = AgglomerativeClustering(linkage="complete").fit(X)


# In[203]:


learned_parameters(agg)


# In[204]:


agg.children_


# # Gaussian Mixtures

# In[205]:


X1, y1 = make_blobs(n_samples=1000, centers=((4, -4), (0, 0)), random_state=42)
X1 = X1.dot(np.array([[0.374, 0.95], [0.732, 0.598]]))
X2, y2 = make_blobs(n_samples=250, centers=1, random_state=42)
X2 = X2 + [6, -8]
X = np.r_[X1, X2]
y = np.r_[y1, y2]


# Let's train a Gaussian mixture model on the previous dataset:

# In[206]:


from sklearn.mixture import GaussianMixture


# In[207]:


gm = GaussianMixture(n_components=3, n_init=10, random_state=42)
gm.fit(X)


# Let's look at the parameters that the EM algorithm estimated:

# In[208]:


gm.weights_


# In[209]:


gm.means_


# In[210]:


gm.covariances_


# Did the algorithm actually converge?

# In[211]:


gm.converged_


# Yes, good. How many iterations did it take?

# In[212]:


gm.n_iter_


# You can now use the model to predict which cluster each instance belongs to (hard clustering) or the probabilities that it came from each cluster. For this, just use `predict()` method or the `predict_proba()` method:

# In[213]:


gm.predict(X)


# In[214]:


gm.predict_proba(X)


# This is a generative model, so you can sample new instances from it (and get their labels):

# In[215]:


X_new, y_new = gm.sample(6)
X_new


# In[216]:


y_new


# Notice that they are sampled sequentially from each cluster.

# You can also estimate the log of the _probability density function_ (PDF) at any location using the `score_samples()` method:

# In[217]:


gm.score_samples(X)


# Let's check that the PDF integrates to 1 over the whole space. We just take a large square around the clusters, and chop it into a grid of tiny squares, then we compute the approximate probability that the instances will be generated in each tiny square (by multiplying the PDF at one corner of the tiny square by the area of the square), and finally summing all these probabilities). The result is very close to 1:

# In[218]:


resolution = 100
grid = np.arange(-10, 10, 1 / resolution)
xx, yy = np.meshgrid(grid, grid)
X_full = np.vstack([xx.ravel(), yy.ravel()]).T

pdf = np.exp(gm.score_samples(X_full))
pdf_probas = pdf * (1 / resolution) ** 2
pdf_probas.sum()


# Now let's plot the resulting decision boundaries (dashed lines) and density contours:

# In[219]:


from matplotlib.colors import LogNorm

def plot_gaussian_mixture(clusterer, X, resolution=1000, show_ylabels=True):
    mins = X.min(axis=0) - 0.1
    maxs = X.max(axis=0) + 0.1
    xx, yy = np.meshgrid(np.linspace(mins[0], maxs[0], resolution),
                         np.linspace(mins[1], maxs[1], resolution))
    Z = -clusterer.score_samples(np.c_[xx.ravel(), yy.ravel()])
    Z = Z.reshape(xx.shape)

    plt.contourf(xx, yy, Z,
                 norm=LogNorm(vmin=1.0, vmax=30.0),
                 levels=np.logspace(0, 2, 12))
    plt.contour(xx, yy, Z,
                norm=LogNorm(vmin=1.0, vmax=30.0),
                levels=np.logspace(0, 2, 12),
                linewidths=1, colors='k')

    Z = clusterer.predict(np.c_[xx.ravel(), yy.ravel()])
    Z = Z.reshape(xx.shape)
    plt.contour(xx, yy, Z,
                linewidths=2, colors='r', linestyles='dashed')
    
    plt.plot(X[:, 0], X[:, 1], 'k.', markersize=2)
    plot_centroids(clusterer.means_, clusterer.weights_)

    plt.xlabel("$x_1$", fontsize=14)
    if show_ylabels:
        plt.ylabel("$x_2$", fontsize=14, rotation=0)
    else:
        plt.tick_params(labelleft=False)


# In[220]:


plt.figure(figsize=(8, 4))

plot_gaussian_mixture(gm, X)

save_fig("gaussian_mixtures_diagram")
plt.show()


# You can impose constraints on the covariance matrices that the algorithm looks for by setting the `covariance_type` hyperparameter:
# * `"full"` (default): no constraint, all clusters can take on any ellipsoidal shape of any size.
# * `"tied"`: all clusters must have the same shape, which can be any ellipsoid (i.e., they all share the same covariance matrix).
# * `"spherical"`: all clusters must be spherical, but they can have different diameters (i.e., different variances).
# * `"diag"`: clusters can take on any ellipsoidal shape of any size, but the ellipsoid's axes must be parallel to the axes (i.e., the covariance matrices must be diagonal).

# In[221]:


gm_full = GaussianMixture(n_components=3, n_init=10, covariance_type="full", random_state=42)
gm_tied = GaussianMixture(n_components=3, n_init=10, covariance_type="tied", random_state=42)
gm_spherical = GaussianMixture(n_components=3, n_init=10, covariance_type="spherical", random_state=42)
gm_diag = GaussianMixture(n_components=3, n_init=10, covariance_type="diag", random_state=42)
gm_full.fit(X)
gm_tied.fit(X)
gm_spherical.fit(X)
gm_diag.fit(X)


# In[222]:


def compare_gaussian_mixtures(gm1, gm2, X):
    plt.figure(figsize=(9, 4))

    plt.subplot(121)
    plot_gaussian_mixture(gm1, X)
    plt.title('covariance_type="{}"'.format(gm1.covariance_type), fontsize=14)

    plt.subplot(122)
    plot_gaussian_mixture(gm2, X, show_ylabels=False)
    plt.title('covariance_type="{}"'.format(gm2.covariance_type), fontsize=14)


# In[223]:


compare_gaussian_mixtures(gm_tied, gm_spherical, X)

save_fig("covariance_type_diagram")
plt.show()


# In[224]:


compare_gaussian_mixtures(gm_full, gm_diag, X)
plt.tight_layout()
plt.show()


# ## Anomaly Detection using Gaussian Mixtures

# Gaussian Mixtures can be used for _anomaly detection_: instances located in low-density regions can be considered anomalies. You must define what density threshold you want to use. For example, in a manufacturing company that tries to detect defective products, the ratio of defective products is usually well-known. Say it is equal to 4%, then you can set the density threshold to be the value that results in having 4% of the instances located in areas below that threshold density:

# In[225]:


densities = gm.score_samples(X)
density_threshold = np.percentile(densities, 4)
anomalies = X[densities < density_threshold]


# In[226]:


plt.figure(figsize=(8, 4))

plot_gaussian_mixture(gm, X)
plt.scatter(anomalies[:, 0], anomalies[:, 1], color='r', marker='*')
plt.ylim(top=5.1)

save_fig("mixture_anomaly_detection_diagram")
plt.show()


# ## Model selection

# We cannot use the inertia or the silhouette score because they both assume that the clusters are spherical. Instead, we can try to find the model that minimizes a theoretical information criterion such as the Bayesian Information Criterion (BIC) or the Akaike Information Criterion (AIC):
# 
# ${BIC} = {\log(m)p - 2\log({\hat L})}$
# 
# ${AIC} = 2p - 2\log(\hat L)$
# 
# * $m$ is the number of instances.
# * $p$ is the number of parameters learned by the model.
# * $\hat L$ is the maximized value of the likelihood function of the model. This is the conditional probability of the observed data $\mathbf{X}$, given the model and its optimized parameters.
# 
# Both BIC and AIC penalize models that have more parameters to learn (e.g., more clusters), and reward models that fit the data well (i.e., models that give a high likelihood to the observed data).

# In[227]:


gm.bic(X)


# In[228]:


gm.aic(X)


# We could compute the BIC manually like this:

# In[229]:


n_clusters = 3
n_dims = 2
n_params_for_weights = n_clusters - 1
n_params_for_means = n_clusters * n_dims
n_params_for_covariance = n_clusters * n_dims * (n_dims + 1) // 2
n_params = n_params_for_weights + n_params_for_means + n_params_for_covariance
max_log_likelihood = gm.score(X) * len(X) # log(L^)
bic = np.log(len(X)) * n_params - 2 * max_log_likelihood
aic = 2 * n_params - 2 * max_log_likelihood


# In[230]:


bic, aic


# In[231]:


n_params


# There's one weight per cluster, but the sum must be equal to 1, so we have one degree of freedom less, hence the -1. Similarly, the degrees of freedom for an $n \times n$ covariance matrix is not $n^2$, but $1 + 2 + \dots + n = \dfrac{n (n+1)}{2}$.

# Let's train Gaussian Mixture models with various values of $k$ and measure their BIC:

# In[232]:


gms_per_k = [GaussianMixture(n_components=k, n_init=10, random_state=42).fit(X)
             for k in range(1, 11)]


# In[233]:


bics = [model.bic(X) for model in gms_per_k]
aics = [model.aic(X) for model in gms_per_k]


# In[234]:


plt.figure(figsize=(8, 3))
plt.plot(range(1, 11), bics, "bo-", label="BIC")
plt.plot(range(1, 11), aics, "go--", label="AIC")
plt.xlabel("$k$", fontsize=14)
plt.ylabel("Information Criterion", fontsize=14)
plt.axis([1, 9.5, np.min(aics) - 50, np.max(aics) + 50])
plt.annotate('Minimum',
             xy=(3, bics[2]),
             xytext=(0.35, 0.6),
             textcoords='figure fraction',
             fontsize=14,
             arrowprops=dict(facecolor='black', shrink=0.1)
            )
plt.legend()
save_fig("aic_bic_vs_k_diagram")
plt.show()


# Let's search for best combination of values for both the number of clusters and the `covariance_type` hyperparameter:

# In[235]:


min_bic = np.infty

for k in range(1, 11):
    for covariance_type in ("full", "tied", "spherical", "diag"):
        bic = GaussianMixture(n_components=k, n_init=10,
                              covariance_type=covariance_type,
                              random_state=42).fit(X).bic(X)
        if bic < min_bic:
            min_bic = bic
            best_k = k
            best_covariance_type = covariance_type


# In[236]:


best_k


# In[237]:


best_covariance_type


# ## Variational Bayesian Gaussian Mixtures

# Rather than manually searching for the optimal number of clusters, it is possible to use instead the `BayesianGaussianMixture` class which is capable of giving weights equal (or close) to zero to unnecessary clusters. Just set the number of components to a value that you believe is greater than the optimal number of clusters, and the algorithm will eliminate the unnecessary clusters automatically.

# In[238]:


from sklearn.mixture import BayesianGaussianMixture


# In[239]:


bgm = BayesianGaussianMixture(n_components=10, n_init=10, random_state=42)
bgm.fit(X)


# The algorithm automatically detected that only 3 components are needed:

# In[240]:


np.round(bgm.weights_, 2)


# In[241]:


plt.figure(figsize=(8, 5))
plot_gaussian_mixture(bgm, X)
plt.show()


# In[242]:


bgm_low = BayesianGaussianMixture(n_components=10, max_iter=1000, n_init=1,
                                  weight_concentration_prior=0.01, random_state=42)
bgm_high = BayesianGaussianMixture(n_components=10, max_iter=1000, n_init=1,
                                  weight_concentration_prior=10000, random_state=42)
nn = 73
bgm_low.fit(X[:nn])
bgm_high.fit(X[:nn])


# In[243]:


np.round(bgm_low.weights_, 2)


# In[244]:


np.round(bgm_high.weights_, 2)


# In[245]:


plt.figure(figsize=(9, 4))

plt.subplot(121)
plot_gaussian_mixture(bgm_low, X[:nn])
plt.title("weight_concentration_prior = 0.01", fontsize=14)

plt.subplot(122)
plot_gaussian_mixture(bgm_high, X[:nn], show_ylabels=False)
plt.title("weight_concentration_prior = 10000", fontsize=14)

save_fig("mixture_concentration_prior_diagram")
plt.show()


# Note: the fact that you see only 3 regions in the right plot although there are 4 centroids is not a bug. The weight of the top-right cluster is much larger than the weight of the lower-right cluster, so the probability that any given point in this region belongs to the top right cluster is greater than the probability that it belongs to the lower-right cluster.

# In[246]:


X_moons, y_moons = make_moons(n_samples=1000, noise=0.05, random_state=42)


# In[247]:


bgm = BayesianGaussianMixture(n_components=10, n_init=10, random_state=42)
bgm.fit(X_moons)


# In[248]:


plt.figure(figsize=(9, 3.2))

plt.subplot(121)
plot_data(X_moons)
plt.xlabel("$x_1$", fontsize=14)
plt.ylabel("$x_2$", fontsize=14, rotation=0)

plt.subplot(122)
plot_gaussian_mixture(bgm, X_moons, show_ylabels=False)

save_fig("moons_vs_bgm_diagram")
plt.show()


# Oops, not great... instead of detecting 2 moon-shaped clusters, the algorithm detected 8 ellipsoidal clusters. However, the density plot does not look too bad, so it might be usable for anomaly detection.

# ## Likelihood Function

# In[249]:


from scipy.stats import norm


# In[250]:


xx = np.linspace(-6, 4, 101)
ss = np.linspace(1, 2, 101)
XX, SS = np.meshgrid(xx, ss)
ZZ = 2 * norm.pdf(XX - 1.0, 0, SS) + norm.pdf(XX + 4.0, 0, SS)
ZZ = ZZ / ZZ.sum(axis=1) / (xx[1] - xx[0])


# In[251]:


from matplotlib.patches import Polygon

plt.figure(figsize=(8, 4.5))

x_idx = 85
s_idx = 30

plt.subplot(221)
plt.contourf(XX, SS, ZZ, cmap="GnBu")
plt.plot([-6, 4], [ss[s_idx], ss[s_idx]], "k-", linewidth=2)
plt.plot([xx[x_idx], xx[x_idx]], [1, 2], "b-", linewidth=2)
plt.xlabel(r"$x$")
plt.ylabel(r"$\theta$", fontsize=14, rotation=0)
plt.title(r"Model $f(x; \theta)$", fontsize=14)

plt.subplot(222)
plt.plot(ss, ZZ[:, x_idx], "b-")
max_idx = np.argmax(ZZ[:, x_idx])
max_val = np.max(ZZ[:, x_idx])
plt.plot(ss[max_idx], max_val, "r.")
plt.plot([ss[max_idx], ss[max_idx]], [0, max_val], "r:")
plt.plot([0, ss[max_idx]], [max_val, max_val], "r:")
plt.text(1.01, max_val + 0.005, r"$\hat{L}$", fontsize=14)
plt.text(ss[max_idx]+ 0.01, 0.055, r"$\hat{\theta}$", fontsize=14)
plt.text(ss[max_idx]+ 0.01, max_val - 0.012, r"$Max$", fontsize=12)
plt.axis([1, 2, 0.05, 0.15])
plt.xlabel(r"$\theta$", fontsize=14)
plt.grid(True)
plt.text(1.99, 0.135, r"$=f(x=2.5; \theta)$", fontsize=14, ha="right")
plt.title(r"Likelihood function $\mathcal{L}(\theta|x=2.5)$", fontsize=14)

plt.subplot(223)
plt.plot(xx, ZZ[s_idx], "k-")
plt.axis([-6, 4, 0, 0.25])
plt.xlabel(r"$x$", fontsize=14)
plt.grid(True)
plt.title(r"PDF $f(x; \theta=1.3)$", fontsize=14)
verts = [(xx[41], 0)] + list(zip(xx[41:81], ZZ[s_idx, 41:81])) + [(xx[80], 0)]
poly = Polygon(verts, facecolor='0.9', edgecolor='0.5')
plt.gca().add_patch(poly)

plt.subplot(224)
plt.plot(ss, np.log(ZZ[:, x_idx]), "b-")
max_idx = np.argmax(np.log(ZZ[:, x_idx]))
max_val = np.max(np.log(ZZ[:, x_idx]))
plt.plot(ss[max_idx], max_val, "r.")
plt.plot([ss[max_idx], ss[max_idx]], [-5, max_val], "r:")
plt.plot([0, ss[max_idx]], [max_val, max_val], "r:")
plt.axis([1, 2, -2.4, -2])
plt.xlabel(r"$\theta$", fontsize=14)
plt.text(ss[max_idx]+ 0.01, max_val - 0.05, r"$Max$", fontsize=12)
plt.text(ss[max_idx]+ 0.01, -2.39, r"$\hat{\theta}$", fontsize=14)
plt.text(1.01, max_val + 0.02, r"$\log \, \hat{L}$", fontsize=14)
plt.grid(True)
plt.title(r"$\log \, \mathcal{L}(\theta|x=2.5)$", fontsize=14)

save_fig("likelihood_function_diagram")
plt.show()


# # Exercise solutions

# ## 1. to 8.

# See appendix A.

# ## 9.

# *Exercise: Load the MNIST dataset (introduced in chapter 3) and split it into a training set and a test set (take the first 60,000 instances for training, and the remaining 10,000 for testing).*

# The MNIST dataset was loaded earlier.

# In[252]:


X_train = mnist['data'][:60000]
y_train = mnist['target'][:60000]

X_test = mnist['data'][60000:]
y_test = mnist['target'][60000:]


# *Exercise: Train a Random Forest classifier on the dataset and time how long it takes, then evaluate the resulting model on the test set.*

# In[253]:


from sklearn.ensemble import RandomForestClassifier

rnd_clf = RandomForestClassifier(n_estimators=10, random_state=42)


# In[254]:


import time

t0 = time.time()
rnd_clf.fit(X_train, y_train)
t1 = time.time()


# In[255]:


print("Training took {:.2f}s".format(t1 - t0))


# In[256]:


from sklearn.metrics import accuracy_score

y_pred = rnd_clf.predict(X_test)
accuracy_score(y_test, y_pred)


# *Exercise: Next, use PCA to reduce the dataset's dimensionality, with an explained variance ratio of 95%.*

# In[257]:


from sklearn.decomposition import PCA

pca = PCA(n_components=0.95)
X_train_reduced = pca.fit_transform(X_train)


# *Exercise: Train a new Random Forest classifier on the reduced dataset and see how long it takes. Was training much faster?*

# In[258]:


rnd_clf2 = RandomForestClassifier(n_estimators=10, random_state=42)
t0 = time.time()
rnd_clf2.fit(X_train_reduced, y_train)
t1 = time.time()


# In[259]:


print("Training took {:.2f}s".format(t1 - t0))


# Oh no! Training is actually more than twice slower now! How can that be? Well, as we saw in this chapter, dimensionality reduction does not always lead to faster training time: it depends on the dataset, the model and the training algorithm. See figure 8-6 (the `manifold_decision_boundary_plot*` plots above). If you try a softmax classifier instead of a random forest classifier, you will find that training time is reduced by a factor of 3 when using PCA. Actually, we will do this in a second, but first let's check the precision of the new random forest classifier.

# *Exercise: Next evaluate the classifier on the test set: how does it compare to the previous classifier?*

# In[260]:


X_test_reduced = pca.transform(X_test)

y_pred = rnd_clf2.predict(X_test_reduced)
accuracy_score(y_test, y_pred)


# It is common for performance to drop slightly when reducing dimensionality, because we do lose some useful signal in the process. However, the performance drop is rather severe in this case. So PCA really did not help: it slowed down training and reduced performance. :(
# 
# Let's see if it helps when using softmax regression:

# In[261]:


from sklearn.linear_model import LogisticRegression

log_clf = LogisticRegression(multi_class="multinomial", solver="lbfgs", random_state=42)
t0 = time.time()
log_clf.fit(X_train, y_train)
t1 = time.time()


# In[262]:


print("Training took {:.2f}s".format(t1 - t0))


# In[263]:


y_pred = log_clf.predict(X_test)
accuracy_score(y_test, y_pred)


# Okay, so softmax regression takes much longer to train on this dataset than the random forest classifier, plus it performs worse on the test set. But that's not what we are interested in right now, we want to see how much PCA can help softmax regression. Let's train the softmax regression model using the reduced dataset:

# In[264]:


log_clf2 = LogisticRegression(multi_class="multinomial", solver="lbfgs", random_state=42)
t0 = time.time()
log_clf2.fit(X_train_reduced, y_train)
t1 = time.time()


# In[265]:


print("Training took {:.2f}s".format(t1 - t0))


# Nice! Reducing dimensionality led to a 4× speedup. :)  Let's check the model's accuracy:

# In[266]:


y_pred = log_clf2.predict(X_test_reduced)
accuracy_score(y_test, y_pred)


# A very slight drop in performance, which might be a reasonable price to pay for a 4× speedup, depending on the application.

# So there you have it: PCA can give you a formidable speedup... but not always!

# ## 10.

# *Exercise: Use t-SNE to reduce the MNIST dataset down to two dimensions and plot the result using Matplotlib. You can use a scatterplot using 10 different colors to represent each image's target class.*

# The MNIST dataset was loaded above.

# Dimensionality reduction on the full 60,000 images takes a very long time, so let's only do this on a random subset of 10,000 images:

# In[267]:


np.random.seed(42)

m = 10000
idx = np.random.permutation(60000)[:m]

X = mnist['data'][idx]
y = mnist['target'][idx]


# Now let's use t-SNE to reduce dimensionality down to 2D so we can plot the dataset:

# In[268]:


from sklearn.manifold import TSNE

tsne = TSNE(n_components=2, random_state=42)
X_reduced = tsne.fit_transform(X)


# Now let's use Matplotlib's `scatter()` function to plot a scatterplot, using a different color for each digit:

# In[269]:


plt.figure(figsize=(13,10))
plt.scatter(X_reduced[:, 0], X_reduced[:, 1], c=y, cmap="jet")
plt.axis('off')
plt.colorbar()
plt.show()


# Isn't this just beautiful? :) This plot tells us which numbers are easily distinguishable from the others (e.g., 0s, 6s, and most 8s are rather well separated clusters), and it also tells us which numbers are often hard to distinguish (e.g., 4s and 9s, 5s and 3s, and so on).

# Let's focus on digits 3 and 5, which seem to overlap a lot.

# In[270]:


plt.figure(figsize=(9,9))
cmap = mpl.cm.get_cmap("jet")
for digit in (2, 3, 5):
    plt.scatter(X_reduced[y == digit, 0], X_reduced[y == digit, 1], c=[cmap(digit / 9)])
plt.axis('off')
plt.show()


# Let's see if we can produce a nicer image by running t-SNE on these 3 digits:

# In[271]:


idx = (y == 2) | (y == 3) | (y == 5) 
X_subset = X[idx]
y_subset = y[idx]

tsne_subset = TSNE(n_components=2, random_state=42)
X_subset_reduced = tsne_subset.fit_transform(X_subset)


# In[272]:


plt.figure(figsize=(9,9))
for digit in (2, 3, 5):
    plt.scatter(X_subset_reduced[y_subset == digit, 0], X_subset_reduced[y_subset == digit, 1], c=[cmap(digit / 9)])
plt.axis('off')
plt.show()


# Much better, now the clusters have far less overlap. But some 3s are all over the place. Plus, there are two distinct clusters of 2s, and also two distinct clusters of 5s. It would be nice if we could visualize a few digits from each cluster, to understand why this is the case. Let's do that now. 

# *Exercise: Alternatively, you can write colored digits at the location of each instance, or even plot scaled-down versions of the digit images themselves (if you plot all digits, the visualization will be too cluttered, so you should either draw a random sample or plot an instance only if no other instance has already been plotted at a close distance). You should get a nice visualization with well-separated clusters of digits.*

# Let's create a `plot_digits()` function that will draw a scatterplot (similar to the above scatterplots) plus write colored digits, with a minimum distance guaranteed between these digits. If the digit images are provided, they are plotted instead. This implementation was inspired from one of Scikit-Learn's excellent examples ([plot_lle_digits](http://scikit-learn.org/stable/auto_examples/manifold/plot_lle_digits.html), based on a different digit dataset).

# In[273]:


from sklearn.preprocessing import MinMaxScaler
from matplotlib.offsetbox import AnnotationBbox, OffsetImage

def plot_digits(X, y, min_distance=0.05, images=None, figsize=(13, 10)):
    # Let's scale the input features so that they range from 0 to 1
    X_normalized = MinMaxScaler().fit_transform(X)
    # Now we create the list of coordinates of the digits plotted so far.
    # We pretend that one is already plotted far away at the start, to
    # avoid `if` statements in the loop below
    neighbors = np.array([[10., 10.]])
    # The rest should be self-explanatory
    plt.figure(figsize=figsize)
    cmap = mpl.cm.get_cmap("jet")
    digits = np.unique(y)
    for digit in digits:
        plt.scatter(X_normalized[y == digit, 0], X_normalized[y == digit, 1], c=[cmap(digit / 9)])
    plt.axis("off")
    ax = plt.gcf().gca()  # get current axes in current figure
    for index, image_coord in enumerate(X_normalized):
        closest_distance = np.linalg.norm(np.array(neighbors) - image_coord, axis=1).min()
        if closest_distance > min_distance:
            neighbors = np.r_[neighbors, [image_coord]]
            if images is None:
                plt.text(image_coord[0], image_coord[1], str(int(y[index])),
                         color=cmap(y[index] / 9), fontdict={"weight": "bold", "size": 16})
            else:
                image = images[index].reshape(28, 28)
                imagebox = AnnotationBbox(OffsetImage(image, cmap="binary"), image_coord)
                ax.add_artist(imagebox)


# Let's try it! First let's just write colored digits:

# In[274]:


plot_digits(X_reduced, y)


# Well that's okay, but not that beautiful. Let's try with the digit images:

# In[275]:


plot_digits(X_reduced, y, images=X, figsize=(35, 25))


# In[276]:


plot_digits(X_subset_reduced, y_subset, images=X_subset, figsize=(22, 22))


# *Exercise: Try using other dimensionality reduction algorithms such as PCA, LLE, or MDS and compare the resulting visualizations.*

# Let's start with PCA. We will also time how long it takes:

# In[277]:


from sklearn.decomposition import PCA
import time

t0 = time.time()
X_pca_reduced = PCA(n_components=2, random_state=42).fit_transform(X)
t1 = time.time()
print("PCA took {:.1f}s.".format(t1 - t0))
plot_digits(X_pca_reduced, y)
plt.show()


# Wow, PCA is blazingly fast! But although we do see a few clusters, there's way too much overlap. Let's try LLE:

# In[278]:


from sklearn.manifold import LocallyLinearEmbedding

t0 = time.time()
X_lle_reduced = LocallyLinearEmbedding(n_components=2, random_state=42).fit_transform(X)
t1 = time.time()
print("LLE took {:.1f}s.".format(t1 - t0))
plot_digits(X_lle_reduced, y)
plt.show()


# That took a while, and the result does not look too good. Let's see what happens if we apply PCA first, preserving 95% of the variance:

# In[279]:


from sklearn.pipeline import Pipeline

pca_lle = Pipeline([
    ("pca", PCA(n_components=0.95, random_state=42)),
    ("lle", LocallyLinearEmbedding(n_components=2, random_state=42)),
])
t0 = time.time()
X_pca_lle_reduced = pca_lle.fit_transform(X)
t1 = time.time()
print("PCA+LLE took {:.1f}s.".format(t1 - t0))
plot_digits(X_pca_lle_reduced, y)
plt.show()


# The result is more or less the same, but this time it was almost 4× faster.

# Let's try MDS. It's much too long if we run it on 10,000 instances, so let's just try 2,000 for now:

# In[280]:


from sklearn.manifold import MDS

m = 2000
t0 = time.time()
X_mds_reduced = MDS(n_components=2, random_state=42).fit_transform(X[:m])
t1 = time.time()
print("MDS took {:.1f}s (on just 2,000 MNIST images instead of 10,000).".format(t1 - t0))
plot_digits(X_mds_reduced, y[:m])
plt.show()


# Meh. This does not look great, all clusters overlap too much. Let's try with PCA first, perhaps it will be faster?

# In[281]:


from sklearn.pipeline import Pipeline

pca_mds = Pipeline([
    ("pca", PCA(n_components=0.95, random_state=42)),
    ("mds", MDS(n_components=2, random_state=42)),
])
t0 = time.time()
X_pca_mds_reduced = pca_mds.fit_transform(X[:2000])
t1 = time.time()
print("PCA+MDS took {:.1f}s (on 2,000 MNIST images).".format(t1 - t0))
plot_digits(X_pca_mds_reduced, y[:2000])
plt.show()


# Same result, and no speedup: PCA did not help (or hurt).

# Let's try LDA:

# In[282]:


from sklearn.discriminant_analysis import LinearDiscriminantAnalysis

t0 = time.time()
X_lda_reduced = LinearDiscriminantAnalysis(n_components=2).fit_transform(X, y)
t1 = time.time()
print("LDA took {:.1f}s.".format(t1 - t0))
plot_digits(X_lda_reduced, y, figsize=(12,12))
plt.show()


# This one is very fast, and it looks nice at first, until you realize that several clusters overlap severely.

# Well, it's pretty clear that t-SNE won this little competition, wouldn't you agree? We did not time it, so let's do that now:

# In[283]:


from sklearn.manifold import TSNE

t0 = time.time()
X_tsne_reduced = TSNE(n_components=2, random_state=42).fit_transform(X)
t1 = time.time()
print("t-SNE took {:.1f}s.".format(t1 - t0))
plot_digits(X_tsne_reduced, y)
plt.show()


# It's twice slower than LLE, but still much faster than MDS, and the result looks great. Let's see if a bit of PCA can speed it up:

# In[284]:


pca_tsne = Pipeline([
    ("pca", PCA(n_components=0.95, random_state=42)),
    ("tsne", TSNE(n_components=2, random_state=42)),
])
t0 = time.time()
X_pca_tsne_reduced = pca_tsne.fit_transform(X)
t1 = time.time()
print("PCA+t-SNE took {:.1f}s.".format(t1 - t0))
plot_digits(X_pca_tsne_reduced, y)
plt.show()


# Yes, PCA roughly gave us a 25% speedup, without damaging the result. We have a winner!

# In[ ]: