def variance_verify_deformation(img, landmarks, filename):
sys.path.append('../nn')
import nn_lib
net = nn_lib.Net()
# net.Load("../nn/deformation_model_10-12-2014_12-54-16.bin")
net.Load(filename)
N = 50000
sel_label = range(0, 25, 4) + range(25, 50, 4)
sel_label_other = range(2, 25, 4) + range(27, 50, 4)
all_network_dims = eval(net.GetAllLayerOutputDims())
print all_network_dims
layer_outputs = []
for dims in all_network_dims:
mat_size = dims[0] * dims[1] * dims[2]
layer_outputs.append(numpy.zeros((N, mat_size), dtype='f4'))
# Dump all responses.
ps, deforms = dump_deformed_images(img, landmarks, 'backward', N)
net.DumpAllResponses(deforms, layer_outputs)
# net.EvaluateRegression(deforms, labels_est);
threshold = 2.0
sel_data_label = selector(ps, sel_label, 0, threshold)
sel_data_other = selector(ps, sel_label_other, 0, threshold)
for i in range(len(all_network_dims)):
var_label = verify_variance(sel_data_label, layer_outputs[i])
var_other = verify_variance(sel_data_other, layer_outputs[i])
print "Variance_label in layer %d = %f" % (i, var_label)
print "Variance_other in layer %d = %f" % (i, var_other)
def lipschitz_verify_deformation(img, landmarks, filename):
sys.path.append('../nn')
import nn_lib
net = nn_lib.Net()
# net.Load("../nn/deformation_model_10-12-2014_12-54-16.bin")
net.Load(filename)
sel_label = range(0, 25, 4) + range(25, 50, 4)
# Dump all responses.
all_network_dims = eval(net.GetAllLayerOutputDims())
print all_network_dims
# build a list of numpy arrays.
N = 2000
layer_outputs = []
for dims in all_network_dims:
mat_size = dims[0] * dims[1] * dims[2]
layer_outputs.append(numpy.zeros((N, mat_size), dtype='f4'))
ps, deforms = dump_deformed_images(img, landmarks, 'backward', N)
net.DumpAllResponses(deforms, layer_outputs)
sel_label = range(0, 25, 4) + range(25, 50, 4)
#labels = (ps[:,sel_label] > 0).astype('f4');
labels = 1.0 / (1.0 + numpy.exp(-ps[:, sel_label]))
# For each layer, find pairwise distances.
pw_images = pairwise_l2_dist(deforms)
nlayer = len(all_network_dims)
pw_labels = pairwise_l2_dist(labels)
colors = ['r', 'g', 'b', 'c']
styles = ['-', '--']
cs = list(itertools.product(styles, colors))
for i in range(len(all_network_dims)):
# print "Size at layer " + str(i) + ":" + str(layer_outputs[i].shape);
pw_layer_i = pairwise_l2_dist(layer_outputs[i])
Gs, gs, max_G = compute_lipschitz(pw_layer_i, pw_labels)
print "layer %d: max_G = %f" % (i, max_G)
plt.plot(Gs, gs, cs[i][0] + cs[i][1], label="Layer %d" % i)
Gs, gs, max_G = compute_lipschitz(pw_labels, pw_labels)
print "GroundTruth layer: max_G = %f" % max_G
plt.plot(Gs, gs, 'b-', label="GroundTruth")
Gs, gs, max_G = compute_lipschitz(pw_images, pw_labels)
print "Input image: max_G = %f" % max_G
plt.plot(Gs, gs, 'k-', label="InputImage")
plt.legend()
plt.show()
import pdb
pdb.set_trace()
output_num = nn_letters_train.get_outputs_num()
layer_depths = [input_num, 3, 3, output_num]
# depth of the network is the greatest layer depth
net_depth = max(layer_depths)
# number of layers in network is the length of the depths list
net_layer = len(layer_depths)
# each weight initialized to same value
initial_weight = 0.5
# each neuron's bias initialized to same value
initial_bias = 0.1
# learning rate
N = 0.5
# create and display training dataset from nn_train_data.py
train_data = nn_letters_train.create_train_set()
print 'Training Data: ASCII art of a letter'
print 'Output 1 if the image is an A, and 0 if it is not', '\n\n'
for elem in train_data:
print elem
print ''
# initialize a network
net0 = nn_lib.Net(net_depth, net_layer, layer_depths, initial_weight,
initial_bias, N)
# one time only before running, set weights and biases to unique values of neuron's location
net0.set_wb()
# train the network
net0.train(train_data)