import numpy as np
import matplotlib.cm as cm
import matplotlib.pyplot as plt
import matplotlib.cm as cm
import matplotlib.image as image
import matplotlib.patches as patches
from sklearn.decomposition import PCA
def plot_encoded_data(X_encoded, y_predicted, y_true, N_classes, title = "", output = "clustering", X_raw = None, show_images = False, verbosity = 1, subset = None, prototypes = None, devs = None):
plt.cla()
plt.clf()
plt.close()
show_mid_characters = X_raw is not None
#PCA on the vectors
pca = PCA(n_components=2)
pca.fit(X_encoded)
PCA(copy=True, iterated_power='auto', n_components=2, random_state=123,svd_solver='auto', tol=0.0, whiten=False)
if verbosity>0:
print("Variation explained by PCA:",pca.explained_variance_ratio_)
#reducing the data via the PCAed coordinates
X_encoded_reduced = pca.transform(X_encoded)
#calculating the sizes
min0,max0 = np.min(X_encoded_reduced[:,0]), np.max(X_encoded_reduced[:,0])
min1,max1 = np.min(X_encoded_reduced[:,1]), np.max(X_encoded_reduced[:,1])
size0 = max0-min0
size1 = max1-min1
p = 0.1
min0 -= size0*p
max0 += size0*p
min1 -= size1*p
max1 += size1*p
if subset is None:
colors = cm.nipy_spectral(np.linspace(0, 1, N_classes))
classes_used = np.unique(y_true)
else:
colors = cm.nipy_spectral(np.linspace(0, 1, np.min([N_classes,subset])))
classes_used = np.random.choice(np.unique(y_true), size = np.min([N_classes,subset]), replace = False)
if show_mid_characters:
#calculating centers of clusters
means = np.zeros((N_classes,2))
N = np.shape(X_encoded)[0]
for j,c in enumerate(classes_used):
inds = [i for i in range(N) if y_true[i] == c]
mean = np.mean(X_encoded_reduced[inds,:], axis = 0)
means[j,:] = np.reshape(mean,(2))
variance = np.sqrt(np.reshape(np.var(means, axis = 0),(2)))
fig, ax = plt.subplots()
plt.title(title + " true labels")
N = np.shape(X_encoded)[0]
handles_true = []
ax.set_ylim((min1,max1))
ax.set_xlim((min0,max0))
for j,c in enumerate(classes_used):
inds_true = [i for i in range(N) if y_true[i] == c]
if ((show_mid_characters) and (not show_images)):
image_to_show = X_raw[inds_true[0],:,:]
plt.set_cmap('binary')
img_size = 0.04
ax.imshow(image_to_show, aspect='auto', extent=(means[j,0]-size0*img_size, means[j,0]+size0*img_size, means[j,1]-size1*img_size, means[j,1]+size1*img_size), zorder=-1)
rect = patches.Rectangle((means[j,0]-size0*img_size,means[j,1]-size1*img_size),2*size0*img_size,2*size1*img_size,linewidth=1,edgecolor=colors[j],facecolor='none')
ax.add_patch(rect)
if prototypes is not None:
prototypes_reduced = pca.transform(prototypes)
plt.plot(prototypes_reduced[j,0],prototypes_reduced[j,1],color = colors[j], marker = "*", markersize=15)
if ((prototypes is not None) and (devs is not None)):
prototypes_reduced = pca.transform(prototypes)
dev = devs[j]
diag = np.diag(dev)
diag_reduced = np.array(pca.transform(diag))
lengths = np.linalg.norm(diag_reduced, axis = 1)
majori,minori = np.argsort(-1.0*lengths)[0:2]
major = np.linalg.norm(diag_reduced[majori])
minor = np.linalg.norm(diag_reduced[minori])
majora = np.arctan2(diag_reduced[majori,1],diag_reduced[majori,0])
minora = np.arctan2(diag_reduced[minori,1],diag_reduced[minori,0])
e = patches.Ellipse(xy = (prototypes_reduced[j,0],prototypes_reduced[j,1]), width = major, height = minor, angle = majora/2.0/np.pi*360.0, color = colors[j], alpha = 0.1)
ax.add_patch(e)
if not show_images:
handles_true.append(ax.scatter(X_encoded_reduced[inds_true,0],X_encoded_reduced[inds_true,1],facecolors = colors[j], edgecolors = colors[j], label = str(classes_used[j]),marker = "o",alpha = 1.0 ))
if show_images:
for i in inds_true:
plt.set_cmap('binary')
img_size = 0.03
ax.imshow(X_raw[i,:,:], aspect='auto', extent=(X_encoded_reduced[i,0]-size0*img_size, X_encoded_reduced[i,0]+size0*img_size,X_encoded_reduced[i,1]-size1*img_size, X_encoded_reduced[i,1]+size1*img_size), zorder=-1)
rect = patches.Rectangle((X_encoded_reduced[i,0]-size0*img_size,X_encoded_reduced[i,1]-size1*img_size),2*size0*img_size,2*size1*img_size,linewidth=1,edgecolor=colors[j],facecolor='none')
ax.add_patch(rect)
plt.savefig(output+"_true.png", bbox_inches='tight', format='png', dpi=300)
plt.cla()
plt.clf()
plt.close()
fig, ax = plt.subplots()
plt.title(title + " pred labels")
N = np.shape(X_encoded)[0]
handles_pred = []
ax.set_ylim((min1,max1))
ax.set_xlim((min0,max0))
for j,c in enumerate(classes_used):
inds_pred = [i for i in range(N) if y_predicted[i] == c]
if ((show_mid_characters) and (not show_images)):
inds_true = [i for i in range(N) if y_true[i] == c]
image_to_show = X_raw[inds_true[0],:,:]
plt.set_cmap('binary')
img_size = 0.04
ax.imshow(image_to_show, aspect='auto', extent=(means[j,0]-size0*img_size, means[j,0]+size0*img_size, means[j,1]-size1*img_size, means[j,1]+size1*img_size), zorder=-1)
rect = patches.Rectangle((means[j,0]-size0*img_size,means[j,1]-size1*img_size),2*size0*img_size,2*size1*img_size,linewidth=1,edgecolor=colors[j],facecolor='none')
ax.add_patch(rect)
if prototypes is not None:
prototypes_reduced = pca.transform(prototypes)
plt.plot(prototypes_reduced[j,0],prototypes_reduced[j,1],color = colors[j], marker = "*", markersize=15)
if not show_images:
handles_pred.append(ax.scatter(X_encoded_reduced[inds_pred,0],X_encoded_reduced[inds_pred,1],facecolors = colors[j], edgecolors = colors[j], label = str(classes_used[j]),marker = "o",alpha = 1.0))
if show_images:
for i in inds_pred:
plt.set_cmap('binary')
img_size = 0.03
ax.imshow(X_raw[i,:,:], aspect='auto', extent=(X_encoded_reduced[i,0]-size0*img_size, X_encoded_reduced[i,0]+size0*img_size,X_encoded_reduced[i,1]-size1*img_size, X_encoded_reduced[i,1]+size1*img_size), zorder=-1)
rect = patches.Rectangle((X_encoded_reduced[i,0]-size0*img_size,X_encoded_reduced[i,1]-size1*img_size),2*size0*img_size,2*size1*img_size,linewidth=1,edgecolor=colors[j],facecolor='none')
ax.add_patch(rect)
plt.savefig(output+"_pred.png", bbox_inches='tight', format='png', dpi=300)
return None
def visualize(images, output = "characters.png", width = 20):
images = 1.0 - images
N,h,w = np.shape(images)
Wi = int(np.floor(np.sqrt(N)))
Wi = width
Hi = int(np.ceil((1.0*N) / (1.0*Wi)))
big = np.zeros((Hi*h,Wi*w))
for yi in range(Hi):
for xi in range(Wi):
i = yi*Wi + xi
if i < N:
big[yi*h:(yi+1)*h,xi*w:(xi+1)*w] = images[i,:,:]
print(big.shape)
plt.axis('off')
plt.set_cmap('Greys')
plt.imshow(big)
plt.savefig(output, bbox_inches='tight', format='png', dpi=1200)
plt.show()
