numx.random.seed(42)
# Load data (download is not existing)
data = io.load_natural_image_patches('NaturalImage.mat')
# Remove the mean of ech image patch separately (also works without)
data = pre.remove_rows_means(data)
# Set input/output dimensions
v1 = 14
v2 = 14
h1 = 14
h2 = 14
# Whiten data using ZCA
zca = pre.ZCA(v1 * v2)
zca.train(data)
data = zca.project(data)
# Split into training/test data
train_data = data[0:40000]
test_data = data[40000:70000]
# Set restriction factor, learning rate, batch size and maximal number of epochs
restrict = 0.01 * numx.max(numxext.get_norms(train_data, axis=1))
eps = 0.1
batch_size = 100
max_epochs = 200
# Create model, initial weights=Glorot init., initial sigma=1.0, initial bias=0,
# no centering (Usually pass the data=training_data for a automatic init. that is
import pydeep.rbm.model as model
import pydeep.rbm.trainer as trainer
import pydeep.rbm.estimator as estimator
# Import linear mixture, preprocessing, and visualization
from pydeep.misc.toyproblems import generate_2d_mixtures
import pydeep.preprocessing as pre
import pydeep.misc.visualization as vis
numx.random.seed(42)
# Create a 2D mxiture
data, mixing_matrix = generate_2d_mixtures(100000, 1, 1.0)
# Whiten data
zca = pre.ZCA(data.shape[1])
zca.train(data)
whitened_data = zca.project(data)
# split training test data
train_data = whitened_data[0:numx.int32(whitened_data.shape[0] / 2.0), :]
test_data = whitened_data[numx.int32(whitened_data.shape[0] /
2.0):whitened_data.shape[0], :]
# Input output dims
h1 = 2
h2 = 2
v1 = whitened_data.shape[1]
v2 = 1
# Create model