def lossless_decoding(population):
# vocabulary of target neural nets
nets = {}
for i in range(len(population)):
# get each element of the net by splitting the string that represent it
# in its components
el = population[i].split('[')
input_size = int(el[0].replace(',', ''))
el = el[1].split(']')
net_layers = el[0].split(',')
el = el[1].split(',')[1:]
net_loss = el[0].replace(',', '')
net_l_rate = el[1].replace(',', '')
# create the net
layers = np.array([[int(net_layers[l]), net_layers[l + 1]]
for l in range(0,
len(net_layers) - 1, 2)])
net = dn.DeepNet(input_size, layers, net_loss)
net.learning_rate = net_l_rate
# add the net to the vocabulary
nets[i] = net
return nets
def demo_autoencoder():
#load and norm the data
data = np.load('scaled_images.npy')
data = np.asarray(data, dtype='float32')
data /= 255.0
#set up and train the initial deepnet
dnn = deepnet.DeepNet(
[data.shape[1], data.shape[1], data.shape[1], data.shape[1] * 2],
['gaussian', 'sigmoid', 'sigmoid', 'sigmoid'])
dnn.train(data, [5, 5, 5], 0.0025)
#save the trained deepnet
pickle.dump(dnn, file('pretrained.pkl', 'wb'))
#unroll the deepnet into an autoencoder
autoenc = unroll_network(dnn.network)
#print out the sizes of the autoenc layers
for i in range(len(autoenc)):
print autoenc[i].W.shape
print autoenc[i].hbias.shape
#fine-tune with backprop
mlp = backprop.NeuralNet(network=autoenc)
trained = mlp.train(mlp.network,
data,
data,
max_iter=10,
validErrFunc='reconstruction',
targetCost='linSquaredErr')
#save
pickle.dump(trained, file('network.pkl', 'wb'))
def demo_autoencoder():
#load and norm the xs
import GwData
xs = GwData.GwData.as_binary()
#OLD - MNIST training
#xs = np.load('scaled_images.npy')
#xs = np.asarray(xs, dtype='float32')
#xs /= 255.0
#set up and train the initial deepnet
dnn = deepnet.DeepNet([xs.shape[1], 300, 150],
['sigmoid', 'sigmoid', 'sigmoid'])
dnn.train(xs, [500, 500], [0.25, 0.1])
#save the trained deepnet
pickle.dump(dnn, file('pretrained.pkl', 'wb'))
#unroll the deepnet into an autoencoder
autoenc = unroll_network(dnn.network)
##fine-tune with backprop
mlp = backprop.NeuralNet(network=autoenc)
trained = mlp.train(mlp.network,
xs,
xs,
max_iter=30,
validErrFunc='reconstruction',
targetCost='linSquaredErr')
##save
pickle.dump(trained, file('network.pkl', 'wb'))
def train(self):
# this will be replaced by calls to loadData.py
#data = np.load('scaled_images.npy')
#data = np.asarray(data, dtype='float32')
#data /= 255.0
l = loadData.Loader(str(self.dataDir),stream=self.stream)
if self.layer_types[0] != 'sigmoid':
layer1_sigmoid = False
else:
layer1_sigmoid = True
l.loadData(layer1_sigmoid)
data = l.XC
if self.limit:
inds = np.arange(data.shape[0])
np.random.shuffle(inds)
data = data[inds[:self.limit_num],:]
self.stream.write(data.shape)
# parse the layer sizes
sizes = []
for i in self.layer_sizes:
if i == -1:
sizes.append(data.shape[1])
else:
sizes.append(i)
#set up and train the initial deepnet
dnn = deepnet.DeepNet(sizes, self.layer_types, stream=self.stream)
dnn.train(data, self.pretrain_iter, self.pretrain_lr)
#save the trained deepnet
#pickle.dump(dnn, file('pretrained.pkl','wb')) # Looks like pickle won't work with Qt :(
self.save(dnn.network, 'pretrained.mat')
#unroll the deepnet into an autoencoder
autoenc = autoencoder.unroll_network(dnn.network)
#fine-tune with backprop
mlp = backprop.NeuralNet(network=autoenc, stream=self.stream)
trained = mlp.train(mlp.network, data, data, max_iter=self.backprop_iter,
validErrFunc='reconstruction', targetCost='linSquaredErr')
#save
#pickle.dump(trained, file('network.pkl','wb'))
self.save(trained, 'network.mat')
def rand_population(i_size, o_size, pop_size, connection_percentage=1.):
# list that contains the final population of nets
net_population = list()
# size of activations' dictionary in deepnet module
dict_act_size = len(dn.activations_dict)
# size of the losses' dictionary in deepnet module
loss_dict_size = len(dn.loss_dict)
for n in range(pop_size):
n_layers = np.random.randint(MIN_NUM_LAYERS, MAX_NUM_LAYERS)
layer = np.array([])
# append a random layer
for i in range(n_layers - 1):
neurons = np.random.randint(MIN_NUM_NEURONS, MAX_NUM_NEURONS)
activation = list(dn.activations_dict.keys())[np.random.randint(
0, dict_act_size)]
layer = np.append(layer, np.array([neurons, activation]))
# append the last layer
activation = list(dn.activations_dict.keys())[np.random.randint(
0, dict_act_size)]
layer = np.append(layer, np.array([o_size, activation]))
# generate the parameters for the net
loss = list(dn.loss_dict.keys())[np.random.randint(0, loss_dict_size)]
if connection_percentage == 1.:
# tuple's form (verbose=, fully_connected=, connection_percentage=)
params = (True, True, 1.)
else:
# tuple's form (verbose=, fully_connected=, connection_percentage=)
params = (True, False, connection_percentage)
# create the net and append it to the population
net = dn.DeepNet(i_size, layer.reshape(n_layers, 2), loss, *params)
net_population.append(net)
return net_population
def demo_autoencoder():
#load and norm the data
data = np.load('scaled_images.npy')
data = np.asarray(data, dtype='float32')
data /= 255.0
#set up and train the initial deepnet
dnn = deepnet.DeepNet([data.shape[1], data.shape[1], data.shape[1], 42],
['sigmoid', 'sigmoid', 'sigmoid', 'sigmoid'])
dnn.train(data, [225, 75, 75], 0.0025)
#save the trained deepnet
pickle.dump(dnn, file('pretrained.pkl', 'wb'))
#unroll the deepnet into an autoencoder
autoenc = unroll_network(dnn.network)
##fine-tune with backprop
mlp = backprop.NeuralNet(network=autoenc)
trained = mlp.train(mlp.network,
data,
data,
max_iter=30,
validErrFunc='reconstruction',
targetCost='linSquaredErr')
##save
pickle.dump(trained, file('network.pkl', 'wb'))
MNIST test script
"""
# solve the relative import ussues
import sys
import os
sys.path.append(os.path.dirname(os.path.realpath(__file__)) + "/../deepnet/")
import deepnet as dn
import utils.utils_digit_recognition as drec
import numpy as np
if __name__ == "__main__":
# create the net and use a 50% connectivity (pre defined dropout)
net = dn.DeepNet(input_size=784,
layers=np.array([[35, "relu"], [10, "sigmoid"]]),
loss="CrossEntropy",
verbose=True)
net.learning_rate = 1e-4
# initialize train, test, validation
train_percentage = 90
train_digits = np.loadtxt('data/mnist/mnist_train.csv',
delimiter=',',
skiprows=40000)
test_digits = np.loadtxt('data/mnist/mnist_test.csv',
delimiter=',',
skiprows=0)
train_size = len(train_digits)
# solve the relative import ussues
import sys
import os
sys.path.append(os.path.dirname(os.path.realpath(__file__)) + "/../deepnet/");
import deepnet as dn
import numpy as np
import utils.utils_digit_recognition as drec
from sklearn import datasets
if __name__ == "__main__":
# create the net
net = dn.DeepNet(4, np.array([[21, "relu"], [3, "sigmoid"]]), "CrossEntropy", verbose=True);
net.learning_rate = 1e-4; # set the learning rate
# import data
iris = datasets.load_iris();
X = iris.data;
y = drec.binarization(iris.target)[:,:3];
# split train and test (70% train, 30% test)
train, test = drec.data_split(X, 70);
train_Y, test_Y = drec.data_split(y, 30);
# initialize the weights
for i in range(len(net.weights)):
net.weights[i] = dn.weights_dict['lecun'](net.weights[i]);
# solve the relative import ussues
import sys
import os
sys.path.append(os.path.dirname(os.path.realpath(__file__)) + "/../deepnet/")
import deepnet as dn
import parallelization.dispatcher as disp
import parallelization.dprediction as dpred
import numpy as np
if __name__ == '__main__':
# without this instruction it does not work ok the IPyhton cosole
__spec__ = "ModuleSpec(name='builtins', loader=<class '_frozen_importlib.BuiltinImporter'>)"
# create the nets
net1 = dn.DeepNet(2,
np.array([[3, "linear"], [2, "exp"]]),
"L2",
verbose=True)
net2 = dn.DeepNet(2,
np.array([[3, "linear"], [2, "exp"]]),
"L2",
verbose=True)
net3 = dn.DeepNet(2,
np.array([[3, "linear"], [2, "exp"]]),
"L2",
verbose=True)
net4 = dn.DeepNet(2,
np.array([[3, "linear"], [2, "exp"]]),
"L2",
verbose=True)
# put the nets in a list
nets = list([net1, net2, net3, net4])
elif re.match(':', chunk):
r1 = 1;
r2 = layer_shape;
else:
print('error while parsing string', chunk);
return;
return '['+str(r1) + ',' + str(r2) +']';
"""
In place test: take this code as a small demonstration of the code
"""
import deepnet as dn
import deepplot.netplot as nep
verbose = False;
if verbose:
example_str = 'layer(1): 1|1, 2|2:3, 3|4:5 layer(2): :|:';
net = dn.DeepNet(3, np.array([[5, "tanh"], [6, "linear"]]), "L2", True);
mask = Mask(net, example_str);
net.fully_connected = False;
net.set_mask(mask.weights);
nep.NetPlot(net);
sample_size = 3
# input/output size
i_size = 100
o_size = 1
X = np.zeros((sample_size, 100, 1))
# turn on 3 bits on each sample, randomly
for x in X:
for i in np.random.choice(i_size, sample_size):
x[i] = 1
# target
Y = 3 * np.ones(X.shape[:1]).reshape(sample_size, 1, 1)
# create the perceptron
net = dn.DeepNet(i_size, np.array([[1, "linear"]]), "L2")
net.learning_rate = 3.75e-1
# initialize the parameters
for i in range(len(net.weights)):
net.weights[i] = dn.weights_dict['lecun'](net.weights[i])
net.bias[i] = dn.weights_dict['lecun'](net.bias[i])
# train
for i in range(sample_size):
net.batch_backpropagation(X[i], Y[i], batch_size=1)
test_size = 10
X_test = np.zeros((test_size, 100, 1))
# test on 10 samples
"""
# solve the relative import ussues
import sys
import os
sys.path.append(os.path.dirname(os.path.realpath(__file__)) + "/../deepnet/")
import matplotlib.pyplot as plt
import numpy as np
import utils.utils_digit_recognition as drec
import deepnet as dn
if __name__ == "__main__":
# initialize the net
net = dn.DeepNet(64, np.array([[35, "leakyrelu"], [10, "leakyrelu"]]),
"L2")
net.learning_rate = 2e-3
# uncomment the following two lines if you want to try a non-fully connected topology
# net = dn.DeepNet(64, np.array([[35, "sigmoid"]]), "CrossEntropy");
# net.add_block(10, "sigmoid", fully_connected=False, connection_percentage=.5);
# initialize train, test, validation
train_percentage = 70
# this percentage must be lower than the test set, since it's taken directly
# from it (for the sake of simplicity)
validation_percentage = 50
# import the dataset
digits = drec.load_digits()
# shuffle together inputs and supervised outputs
