t-SNE visualization code

tsne visualization

# Dinesh Jayaraman

# Based on code by
# Authors: Fabian Pedregosa <fabian.pedregosa@inria.fr>
#          Olivier Grisel <olivier.grisel@ensta.org>
#          Mathieu Blondel <mathieu@mblondel.org>
#          Gael Varoquaux
# License: BSD 3 clause (C) INRIA 2011

print(__doc__)
from time import time

import numpy as np
import matplotlib.pyplot as plt
from matplotlib import offsetbox
from matplotlib.patches import Rectangle
from sklearn import (manifold, datasets, decomposition, ensemble,
                     discriminant_analysis, random_projection)
import scipy.misc
import matplotlib.patches as mpatches
imresize = scipy.misc.imresize

from collections import namedtuple
import h5py
import ipdb

num_samples = 10000
n_neighbors = 30
dsname = 'ModelNet10'
datafield = 'view'
labelfield = 'labs'
#dsname = 'ModelNet30'; datafield = 'data'; labelfield = 'label';

data_h5filename = '%s_1view_train_torchfeed.h5' % dsname
feat_h5filename = '3000112_-%s_trainfeatures.h5' % dsname

if dsname == 'ModelNet10':
    classnames = [
          'bathtub',
          'bed',
          'chair',
          'desk',
          'dresser',
          'monitor',
          'night_stand',
          'sofa',
          'table',
          'toilet'
      ]
else:
    classnames = []

#ds = datasets.load_digits(n_class=6)

ds = namedtuple('dataset', ['data', 'target', 'images'])
h5file = h5py.File(data_h5filename, 'r')
ds.images = np.squeeze(h5file[datafield][:]).transpose(0, 2, 1)  # Nx32x32
ds.target = h5file[labelfield][:].reshape(-1)  # N
h5file.close()
h5file = h5py.File(feat_h5filename, 'r')
ds.data = h5file['feat_ip1'][:]  # D=Nx256
lab = h5file['cls_labs'][:].reshape(-1)  # N
assert(np.array_equal(lab, ds.target))
h5file.close()


# subsampling the data
ds.data = ds.data[:num_samples]
ds.target = ds.target[:num_samples].astype(int)
num_cls = np.max(ds.target)+1
ds.images = ds.images[:num_samples]

X = ds.data
y = ds.target
n_samples, n_features = X.shape
# ipdb.set_trace()

#----------------------------------------------------------------------
# Scale and visualize the embedding vectors


def plot_embedding(xy, title=None, show_images=True, show_points=True):
    assert(show_images or show_points)
    xy_min, xy_max = np.min(xy, 0), np.max(xy, 0)
    xy = (xy - xy_min) / (xy_max - xy_min)
    if show_points:
        fig = plt.figure(figsize=(11, 6))
    else:
        fig = plt.figure(figsize=(12, 6))
    ax = plt.subplot(111)
    if show_points:
        seen_cls = np.zeros(num_cls, dtype='bool')
        for i in range(xy.shape[0]):
            cls = ds.target[i]
            plt.text(xy[i, 0], xy[i, 1], str(cls+1),
                     color=plt.cm.tab10((cls) / 10.),
                     fontdict={'weight': 'bold', 'size': 9})
        patches = []
        for clsno in range(num_cls):
            patches.append(
                mpatches.Patch(
                    color=plt.cm.tab10(
                        (clsno) /
                        num_cls),
                    label=(
                        '%d: ' %
                        (clsno +
                         1)) +
                    classnames[clsno]))
        lgd = ax.legend(handles=patches, fontsize=16, bbox_to_anchor=(1.3, 1))

    if show_images and hasattr(offsetbox, 'AnnotationBbox'):
        # only print thumbnails with matplotlib > 1.0
        shown_images = np.array([[1., 1.]])  # just something big
        for i in range(ds.data.shape[0]):
            dist = np.sum((xy[i] - shown_images) ** 2, 1)
            if np.min(dist) < 2e-3:
                # don't show points that are too close
                continue
            shown_images = np.r_[shown_images, [xy[i]]]
            imagebox = offsetbox.AnnotationBbox(
                offsetbox.OffsetImage(
                    np.pad(
                        ds.images[i],
                        1,
                        'constant'),
                    cmap=plt.cm.gray),
                xy[i],
                frameon=False)
            ax.add_artist(imagebox)
        lgd = None
    plt.xticks([]), plt.yticks([])
    if title is not None:
        plt.title(title)
    return [fig, lgd]
    # print(fig.get_size_inches())


#----------------------------------------------------------------------
# Plot images of the ds
img_h = ds.images.shape[2]
pad = 4
n_img_per_row = min(20, int(np.floor(np.sqrt(n_samples))))
img = np.zeros(((img_h+pad) * n_img_per_row, (img_h+pad) * n_img_per_row))
for i in range(n_img_per_row):
    ix = (img_h+pad) * i + pad/2
    for j in range(n_img_per_row):
        iy = (img_h+pad) * j + pad/2
        img[int(ix):int(ix + img_h), int(iy):int(iy + img_h)
            ] = ds.images[int(i * n_img_per_row + j)]

plt.imshow(img, cmap=plt.cm.gray)
plt.xticks([])
plt.yticks([])
plt.title('A selection from the dataset')


#----------------------------------------------------------------------
# t-SNE embedding of the ds dataset
print("Computing t-SNE embedding")
tsne = manifold.TSNE(n_components=2, init='pca', random_state=0)
t0 = time()
X_tsne = tsne.fit_transform(X)
print("plotting")
fig1, lgd1 = plot_embedding(X_tsne,
                            title=None,
                            show_images=False,
                            show_points=True)
fig2, _ = plot_embedding(X_tsne,
                         title=None,
                         show_images=True,
                         show_points=False)

if n_samples > 0:
    print("saving figures")
    plt.figure(fig1.number)
    plt.savefig(
        '/home/dineshj/Documents/Drafts/oneshot/figs/%s_tsne_noims-%dsamples-%dnbrs-wide.pdf' %
        (dsname, num_samples, n_neighbors), bbox_extra_artists=(
            lgd1,), bbox_inches='tight')
    plt.figure(fig2.number)
    plt.savefig(
        '/home/dineshj/Documents/Drafts/oneshot/figs/%s_tsne-%dsamples-%dnbrs-wide.pdf' %
        (dsname, num_samples, n_neighbors), bbox_inches='tight')
print("%.2f s" % (time()-t0))
plt.show()