Skip to content

Instantly share code, notes, and snippets.

@sinhrks
Last active March 8, 2016 14:06
Show Gist options
  • Star 1 You must be signed in to star a gist
  • Fork 0 You must be signed in to fork a gist
  • Save sinhrks/f0a6185ef77b1ec4fa05 to your computer and use it in GitHub Desktop.
Save sinhrks/f0a6185ef77b1ec4fa05 to your computer and use it in GitHub Desktop.
Visualize Multilayer Perceptron Example in deeplearning.net
"""
This code is used to visualize
http://deeplearning.net/tutorial/mlp.html#tips-and-tricks-for-training-mlps
Usage: put the following on 389th line
title = "whatever you want"
plot_pca(classifier, x, train_set_x, train_set_y, index=epoch, title=title)
"""
def plot_pca(classifier, x_symbol, x_data, y_data, index=0,
title=None, sampling=True):
import itertools
import matplotlib.pyplot as plt
from sklearn.decomposition import PCA
fig, axes = plt.subplots(3, 3, figsize=(5, 5))
axes = axes.flatten()
apply_hidden = theano.function(inputs=[x_symbol], outputs=classifier.hiddenLayer.output)
z_data = apply_hidden(x_data.get_value())
labels = y_data.eval()
numbers = range(10)
colors = {0: '#263B1C', 1: '#263374', 2: '#3568B5', 3: '#8A5DDF', 4: '#DBB8EE',
5: '#46B1C9', 6: '#84C0C6', 7: '#9FB7B9', 8: '#BCC1BA', 9: '#F2E2D2'}
for ax, prod in zip(axes, zip(numbers[:-1], numbers[1:])):
# print(ax, prod)
pca = PCA(n_components=2)
indexer = numpy.arange(len(labels))[numpy.in1d(labels, prod)]
label = labels[indexer]
z = z_data[indexer]
pca.fit(z)
z_pca = pca.transform(z)
if sampling:
indexer = numpy.arange(len(label))
numpy.random.shuffle(indexer)
indexer = indexer[:300]
z_pca = z_pca[indexer]
label = label[indexer]
_c = [colors[l] for l in label]
ax.scatter(z_pca[:, 0], z_pca[:, 1], color=_c, alpha=0.3)
ax.xaxis.set_visible(False)
ax.yaxis.set_visible(False)
ax.set_title('{0}, {1}'.format(prod[0], prod[1]), size='small')
# plt.show()
if title is not None:
fig.suptitle(title)
plt.savefig('pca_{0:02d}.png'.format(index))
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment