swiss_roll_distribution.ipynb (Source)

In [1]:
%matplotlib notebook
In [2]:
import numpy as np
import matplotlib.pyplot as plt

import keras.backend as K

from keras.layers import Lambda, Input
from keras.models import Sequential, Model

Using TensorFlow backend.
In [3]:
np.set_printoptions(precision=2,
                    edgeitems=3,
                    linewidth=80,
                    suppress=True)
In [4]:
'TensorFlow version: ' + K.tf.__version__
Out[4]:
'TensorFlow version: 1.4.0'
In [5]:
golden_figsize = lambda width: (width, 2. * width / (1 + np.sqrt(5)))
In [6]:
def swiss_roll(theta, a=0., b=.25):
    """ Archimedean spiral. """

    theta = 4. * np.pi * K.sqrt(theta)
    r = a + b * theta

    return K.stack([r * K.cos(theta), r * K.sin(theta)], axis=-1)
In [7]:
fig, ax = plt.subplots(figsize=(5, 5))

ax.scatter(*K.eval(swiss_roll(K.constant(np.linspace(0, 1, 100)))).T)

ax.set_xlim(-5, 5)
ax.set_ylim(-5, 5)

plt.show()
In [8]:
fig, ax = plt.subplots(figsize=(5, 5))

ax.scatter(*K.eval(swiss_roll(K.random_uniform(shape=(100,)))).T)

ax.set_xlim(-5, 5)
ax.set_ylim(-5, 5)

plt.show()

Within Keras layers

In [9]:
x = Input(shape=(16,))
eps = Input(tensor=K.random_uniform(shape=(K.shape(x)[0],)))
swiss_roll_sample = Lambda(swiss_roll)(eps)
In [10]:
m = Model([x, eps], swiss_roll_sample)
In [11]:
m.predict(np.ones((10, 16)))
Out[11]:
array([[ 1.47, -0.35],
       [ 1.25, -2.57],
       [ 1.59,  0.31],
       [ 1.48, -0.32],
       [ 1.09,  1.43],
       [-0.37, -2.69],
       [ 1.32,  1.14],
       [ 0.13,  0.22],
       [-2.35,  0.09],
       [ 0.85, -2.69]], dtype=float32)
In [12]:
fig, ax = plt.subplots(figsize=(5, 5))

ax.scatter(*m.predict(np.ones((100, 16))).T)

ax.set_xlim(-5, 5)
ax.set_ylim(-5, 5)

plt.show()