def test_reconstruction_error(self):
"""Tests the function `reconstruction_error`.
A curve is saved at
"tools/pseudo_visualization/reconstruction_error.png".
The test is successful if the curve
looks like the evolution of the cross
entropy between a visible unit fixed to
1.0 and a unit moving from 0.0 to 1.0.
"""
nb_points = 999
# The visible unit is modeled as
# the probability that a Bernoulli
# random variable turns on.
visible_unit = 1.
reconstruction = numpy.linspace(0.001, 0.999, num=nb_points)
rec_error = numpy.reshape(-visible_unit*numpy.log(reconstruction) -
(1. - visible_unit)*numpy.log(1. - reconstruction), (1, nb_points))
tls.plot_graphs(reconstruction,
rec_error,
'reconstruction',
'reconstruction error',
['visible unit = 1.'],
['b'],
'Evolution of the reconstruction error with the reconstruction',
'tools/pseudo_visualization/reconstruction_error.png')
def test_kl_divergence(self):
"""Tests the function `kl_divergence`.
A curve is saved at
"tools/pseudo_visualization/kl_divergence.png".
The test is successful if the curve
is convex and its minimum is 0.
"""
nb_points = 201
z_log_std_squared = numpy.linspace(-5., 5., num=nb_points)
kl_divergence = numpy.reshape(0.5*(-1. - z_log_std_squared +
numpy.exp(z_log_std_squared)), (1, nb_points))
tls.plot_graphs(z_log_std_squared,
kl_divergence,
'log of the std squared',
'KL divergence',
['zero mean'],
['b'],
'Evolution of the KL divergence with the log of the std squared',
'tools/pseudo_visualization/kl_divergence.png')
def test_plot_graphs(self):
"""Tests the function `plot_graphs`.
A plot is saved at
"tools/pseudo_visualization/plot_graphs.png".
The test is successful is the two
graphs in the plot are consistent
with the legend.
"""
x_values = numpy.linspace(-5., 5., num=101)
y_values = numpy.zeros((2, 101))
y_values[0, :] = 1.7159*numpy.tanh((2./3)*x_values)
y_values[1, :] = 1./(1. + numpy.exp(-x_values))
tls.plot_graphs(x_values,
y_values,
'input',
'neural activation',
['scaled tanh', 'sigmoid'],
['b', 'r'],
'Evolution of the neural activation with the input',
'tools/pseudo_visualization/plot_graphs.png')
def test_relu(self):
"""Tests the function `relu`.
A plot is saved at
"tools/pseudo_visualization/relu.png".
The test is successful if the curve
of ReLU is consistent with the curve
of its derivative.
"""
nb_x = 1001
x_values = numpy.linspace(-5., 5., num=nb_x)
y_values = numpy.zeros((2, nb_x))
y_values[0, :] = tls.relu(x_values)
y_values[1, :] = tls.relu_derivative(x_values)
tls.plot_graphs(x_values,
y_values,
'$x$',
'$y$',
['$f(x) = $ReLU$(x)$', r'$\partial f(x) / \partial x$'],
['r', 'b'],
'ReLU and its derivative',
'tools/pseudo_visualization/relu.png')
print('Training loss of density approximation: {}'.format(
loss_density_approx[0, counter]))
print('Quantization bin width: {}'.format(entropy_ae.bin_width))
counter += 1
entropy_ae.training_fct(training_float64)
entropy_ae.training_eae_bw(training_float64)
evenly_spaced = numpy.linspace(100,
args.nb_epochs_training,
num=nb_measures - 1,
dtype=numpy.int32)
x_values = numpy.concatenate(
(numpy.ones(1, dtype=numpy.int32), evenly_spaced))
tls.plot_graphs(
x_values, scaled_approx_entropy, 'epoch',
'scaled approximate entropy of the quantized latent variables',
['training'], ['r'],
'Evolution of the scaled approximate entropy over epochs',
os.path.join(path_to_checking_l, 'scaled_approximate_entropy.png'))
tls.plot_graphs(
x_values, rec_error, 'epoch', 'reconstruction error', ['training'],
['r'], 'Evolution of the reconstruction error over epochs',
os.path.join(path_to_checking_l, 'reconstruction_error.png'))
tls.plot_graphs(
x_values, loss_density_approx, 'epoch',
'loss of density approximation', ['training'], ['r'],
'Evolution of the loss of density approximation over epochs',
os.path.join(path_to_checking_l, 'loss_density_approximation.png'))
t_stop = time.time()
nb_minutes = int((t_stop - t_start) / 60)
print('\nTraining time: {} minutes.'.format(nb_minutes))
entropy_ae.checking_p_3(
'encoder', 1,
os.path.join(path_to_checking_p,
'image_gamma_1_' + str_epoch + '.png'))
entropy_ae.save(sess, path_to_model, path_to_nb_itvs_per_side_save)
# The optional argument `dtype` in
# the function `numpy.linspace` was
# introduced in Numpy 1.9.0.
x_values = numpy.linspace(1,
args.nb_epochs_training,
num=args.nb_epochs_training,
dtype=numpy.int32)
tls.plot_graphs(
x_values, mean_approx_entropy, 'epoch',
'mean approximate entropy of the quantized latent variables',
['training', 'validation'], ['r', 'b'],
'Evolution of the mean approximate entropy over epochs',
os.path.join(path_to_checking_l, 'mean_approximate_entropy.png'))
tls.plot_graphs(x_values, mean_disc_entropy, 'epoch',
'mean entropy of the quantized latent variables',
['training', 'validation'], ['r', 'b'],
'Evolution of the mean entropy over epochs',
os.path.join(path_to_checking_l, 'mean_entropy.png'))
tls.plot_graphs(
x_values, scaled_approx_entropy, 'epoch',
'scaled cumulated approximate entropy of the quantized latent variables',
['training', 'validation'], ['r', 'b'],
'Evolution of the scaled cumulated approximate entropy over epochs',
os.path.join(path_to_checking_l, 'scaled_approximate_entropy.png'))
tls.plot_graphs(
x_values, rec_error, 'epoch', 'reconstruction error',
'ar_mean_' + str_epoch + '.png'),
os.path.join(path_to_checking_a,
'ar_log_std_squared_' + str_epoch + '.png'),
os.path.join(path_to_checking_a,
'image_ar_l1_' + str_epoch + '.png'))
# The optional argument `dtype` in the
# function `numpy.linspace` was introduced
# in Numpy 1.9.0.
x_values = numpy.linspace(1,
args.nb_epochs_training,
num=args.nb_epochs_training,
dtype=numpy.int32)
tls.plot_graphs(
x_values, scaled_kld, 'epoch',
'scaled KL divergence of the approximate posterior from the prior',
['training', 'validation'], ['r', 'b'],
'Evolution of scaled KL divergence over epochs',
os.path.join(path_to_checking_l, 'scaled_kld.png'))
tls.plot_graphs(
x_values, rec_error, 'epoch', 'reconstruction error',
['training', 'validation'], ['r', 'b'],
'Evolution of the reconstruction error over epochs',
os.path.join(path_to_checking_l, 'reconstruction_error.png'))
tls.plot_graphs(x_values, w_decay, 'epoch', 'weight decay',
['l2-norm weight decay'], ['b'],
'Evolution of the weight decay over epochs',
os.path.join(path_to_checking_l, 'weight_decay.png'))
tls.plot_graphs(
x_values, mean_magnitude, 'epoch', 'mean magnitude ratio', [
'recognition l1', 'recognition mean',
'recognition log std squared', 'generation l1', 'generation mean'
'ae_latent_' + str_epoch + '.png'),
os.path.join(path_to_checking_a,
'ae_noisy_latent_' + str_epoch + '.png'),
os.path.join(path_to_checking_a,
'image_dead_zone_' + str_epoch + '.png'))
# The optional argument `dtype` in the
# function `numpy.linspace` was introduced
# in Numpy 1.9.0.
x_values = numpy.linspace(1,
args.nb_epochs_training,
num=args.nb_epochs_training,
dtype=numpy.int32)
tls.plot_graphs(
x_values, approx_entropy, 'epoch',
'approximate entropy of the quantized latent variables',
['training', 'validation'], ['r', 'b'],
'Evolution of the approximate entropy over epochs',
os.path.join(path_to_checking_l, 'approximate_entropy.png'))
tls.plot_graphs(x_values, disc_entropy, 'epoch',
'entropy of the quantized latent variables',
['training', 'validation'], ['r', 'b'],
'Evolution of the entropy over epochs',
os.path.join(path_to_checking_l, 'entropy.png'))
tls.plot_graphs(
x_values, scaled_approx_entropy, 'epoch',
'scaled approximate entropy of the quantized latent variables',
['training', 'validation'], ['r', 'b'],
'Evolution of the scaled approximate entropy over epochs',
os.path.join(path_to_checking_l, 'scaled_approximate_entropy.png'))
tls.plot_graphs(
x_values, rec_error, 'epoch', 'reconstruction error',