def test_density_gauss_1d(self):
# Arrange
mu = 0
sigma = 1
n_bins = 180
N_points = int(10e5)
samples = np.random.randn(N_points, 1) * sigma + mu
p_min = np.min(samples)
p_max = np.max(samples)
delta = 3 * (p_max - p_min) / n_bins
points_linsp = np.linspace(p_min - delta, p_max + delta, n_bins)
net = nn.RMIOptimizer(H_nbins=n_bins)
# Act
dx_expected = points_linsp[1] - points_linsp[0]
px_expected = 1 / np.sqrt(2 * np.pi * sigma**2) * np.exp(
-0.5 * (points_linsp - mu)**2 / sigma**2)
tf_y = nn.tf.convert_to_tensor(samples.T, nn.K.backend.floatx())
px, dx = net.tf_calcProbabilityDistribution(tf_y)
px = px.numpy()
dx = dx.numpy()
# Assert
np.testing.assert_almost_equal(px, px_expected, decimal=2)
def test_density_gauss_2d(self):
# Arrange
mu = np.array([0, 1])
sigma = np.array([[1, 0.5], [0.5, 2]])
det_sigma = np.linalg.det(sigma)
n_bins = 100
N_points = int(10e4)
samples = np.random.multivariate_normal(mu, sigma, N_points)
p_min = np.min(samples, 0)
p_max = np.max(samples, 0)
delta = 3 * (p_max - p_min) / n_bins
points_linspX = np.linspace(p_min[0] - delta[0], p_max[0] + delta[0],
n_bins)
points_linspY = np.linspace(p_min[1] - delta[0], p_max[1] + delta[0],
n_bins)
mesh_x, mesh_y = np.meshgrid(points_linspX, points_linspY)
point_x = mesh_x.flatten()
point_y = mesh_y.flatten()
points = np.array([point_x, point_y]).T
net = nn.RMIOptimizer(H_nbins=n_bins)
# Act
dx_expected = (p_max - p_min) / n_bins
px_expected = 1/np.sqrt((2*np.pi)**2*det_sigma) * \
np.exp(-0.5*np.diag(np.dot((points-mu), np.dot((points-mu), np.linalg.inv(sigma)).T)))
px_expected = px_expected.reshape([n_bins, n_bins]).T
tf_y = nn.tf.convert_to_tensor(samples.T, nn.K.backend.floatx())
px, dx = net.tf_calcProbabilityDistribution(tf_y)
px = px.numpy()
dx = dx.numpy()
# Assert
np.testing.assert_almost_equal(dx, dx_expected, decimal=2)
np.testing.assert_almost_equal(px, px_expected, decimal=2)
def get_batch(N_batch):
return wp.produce_Wave_Packet(n_pixels=args.N_pixels,
n_samples=N_batch,
width=args.width,
noise=args.noise,
pos_range=args.pos_range)
rmi_optimizer = nn.RMIOptimizer(H_binsize,
H_kernel_size,
coeff_gauss,
reg_amplitude=reg_amplitude,
reg_decay=reg_decay,
layers=[
nn.K.layers.Dense(
args.neurons,
activation="tanh",
input_shape=(args.N_pixels, )),
nn.K.layers.Dense(args.neurons,
activation="tanh"),
nn.K.layers.Dense(n_out)
])
rmi_optimizer.compile(optimizer=nn.tf.optimizers.Adam(args.eta))
rmi_net = nn.Net(rmi_optimizer, mode="w", path=path)
rmi_net.fit_generator(lambda: get_batch(args.batchsize)[0],
args.N_train,
force_training=True)
rmi_net.plot_rmi_cost(save_file="../logs/wavepacket_" + args.output + ".pdf")
path = None
def get_batch(N_batch):
return wp.produce_Wave_Packet(n_pixels=args.N_pixels,
n_samples=N_batch,
width=args.width,
noise=args.noise,
pos_range=args.pos_range)
encoder = nn.RMIOptimizer(H_binsize,
H_kernel_size,
layers=[
nn.K.layers.Dense(args.neurons,
activation="tanh",
input_shape=(args.N_pixels, )),
nn.K.layers.Dense(args.neurons,
activation="tanh"),
nn.K.layers.Dense(n_out)
])
decoder = nn.K.Sequential(layers=[
nn.K.layers.Dense(args.neurons, activation="relu", input_shape=(n_out, )),
nn.K.layers.Dense(args.neurons, activation="relu"),
nn.K.layers.Dense(args.N_pixels)
])
autoencoder = nn.Supervised(cost="mse_contract", layers=[encoder, decoder])
autoencoder.compile(optimizer=nn.tf.optimizers.Adam(args.eta))
ae_net = nn.Net(autoencoder, mode="w", path=path)
path = None
H_binsize = 100
H_kernel_size = 2
n_out = 2
def get_batch(batchsize):
random_indices = np.random.choice(np.arange(N_samples), size=batchsize)
return samples[random_indices], None, random_indices
encoder = nn.RMIOptimizer(H_binsize,
H_kernel_size,
layers=[
nn.K.layers.Dense(args.n_neurons,
activation="relu",
input_shape=(n_in, )),
nn.K.layers.Dense(n_out)
])
decoder = nn.K.Sequential(layers=[
nn.K.layers.Dense(args.n_neurons, activation="relu", input_shape=(
n_out, )),
nn.K.layers.Dense(n_in)
])
autoencoder = nn.Supervised(cost="mse_contract", layers=[encoder, decoder])
autoencoder.compile(optimizer=nn.tf.optimizers.RMSprop(eta))
ae_net = nn.Net(autoencoder, mode="w", path=path)
enc_net = nn.Net(encoder)