def __init__(self, np_rng, theano_rng=None, n_ins=784,
hidden_layer_sizes=[500, 500],
n_outs=10,
corruption_rates=[0.1, 0.1]):
self.sigmoid_layers = []
self.dA_layers = []
self.params = []
self.n_layers = len(hidden_layer_sizes)
if theano_rng is None:
theano_rng = RandomStreams(np_rng.randint(2**30))
self.x = T.matrix('x')
self.y = T.ivector('y')
for i in range(self.n_layers):
if i == 0:
input_size = n_ins
layer_input = self.x
else:
input_size = hidden_layer_sizes[i-1]
layer_input = self.sigmoid_layers[-1].output
hidden_layer = HiddenLayer(
np_rng, layer_input, input_size, hidden_layer_sizes[i], sigmoid)
self.sigmoid_layers.append(hidden_layer)
self.params.extend(hidden_layer.params)
# construc the denoising autoencoder layer
da = dA(np_rng, theano_rng, layer_input,
n_visible=input_size, n_hidden=hidden_layer_sizes[i],
W=hidden_layer.W, bhid=hidden_layer.b)
self.dA_layers.append(da)
# LogisticRegression Layer for classification
self.logLayer = LogisticRegressionLayer(self.sigmoid_layers[-1].output, n_in=hidden_layer_sizes[-1],
n_out=n_outs)
self.params.extend(self.logLayer.params)
self.finetune_cost = self.logLayer.negative_log_likelihood(self.y)
self.errors = self.logLayer.errors(self.y)
def iterative_algorithm(self,
name,
pop_size=100,
genome_length=20,
lim_percentage=20,
corruption_level=0.2,
num_epochs=50,
lr=0.1,
max_evaluations=200000,
unique_training=False,
hiddens=40,
rtr=True):
self.mask = np.random.binomial(1, 0.5, genome_length)
trials = max_evaluations / pop_size
population_limit = int(pop_size * (lim_percentage / 100.0))
self.dA = dA(n_visible=genome_length, n_hidden=hiddens)
self.dA.build_dA(corruption_level)
self.build_sample_dA()
new_population = np.random.binomial(1, 0.5, (pop_size, genome_length))
self.population_fitnesses = self.fitness_many(new_population)
for iteration in range(0, trials):
print "iteration:", iteration
population = new_population
self.population = new_population
rw = self.tournament_selection_replacement(population)
good_strings, good_strings_fitnesses = self.get_good_strings(
population,
population_limit,
unique=unique_training,
fitnesses=self.population_fitnesses)
print "training A/E"
training_data = np.array(good_strings)
self.train_dA(training_data, num_epochs=num_epochs, lr=lr)
print "sampling..."
sampled_population = np.array(self.sample_dA(rw), "b")
self.sample_fitnesses = self.fitness_many(sampled_population)
if rtr:
new_population = self.RTR(
population,
sampled_population,
population_fitnesses=self.population_fitnesses,
sample_fitnesses=self.sample_fitnesses,
w=pop_size / 20)
else:
new_population = sampled_population
new_population[0:1] = good_strings[0:1]
self.population_fitnesses = self.sample_fitnesses
self.population_fitnesses[0:1] = good_strings_fitnesses[0:1]
print "{0},{1},{2}\n".format(np.mean(self.population_fitnesses),
np.min(self.population_fitnesses),
np.max(self.population_fitnesses))
print "best from previous:", (self.fitness(
new_population[np.argmax(self.population_fitnesses)]))
if np.max(self.population_fitnesses) == self.optimum:
pickle.dump(
{
"pop": self.population,
"fitnesses": self.population_fitnesses,
"iteration": iteration
}, open("final_shit_ae.pkl", "w"))
break
return new_population
def iterative_algorithm(
self,
name,
pop_size=100,
genome_length=20,
lim_percentage=20,
lim=20,
trials=10,
corruption_level=0.2,
num_epochs=50,
lr = 0.1,
online_training=False,
pickle_data=False,
max_evaluations=200000,
save_data=False):
results_path = "results/autoencoder/{0}/".format(name)
ensure_dir(results_path)
trials = max_evaluations/pop_size
population_limit = lim
if lim_percentage > 0:
population_limit = int(pop_size*(lim_percentage/100.0))
print "population_limit:",population_limit
print "{0}*({1}/100.0) = {2}".format(pop_size,lim_percentage,int(pop_size*(lim_percentage/100.0)))
fitfile = open("{0}fitnesses.dat".format(results_path),"w")
self.dA = dA(n_visible=genome_length,n_hidden=50)
self.dA.build_dA(corruption_level)
self.build_sample_dA()
all_strings,good_strings=self.generate_good_strings(pop_size,genome_length,population_limit)
self.train_dA(ar(good_strings),corruption_level=corruption_level,num_epochs=num_epochs,lr=lr,output_folder=results_path,iteration=0)
# sampled_population = [self.sample_dA([i]) for i in self.get_new_population_rw(all_strings)]
sampled_population = self.sample_dA(self.get_new_population_rw(all_strings))
print "s:",sampled_population
original_fitnesses,sample_fitnesses,differences,distances = self.get_statistics(all_strings,sampled_population)
data = {
"original":original_fitnesses,
"fitnesses_sampled":sample_fitnesses,
"differences_in_fitness":differences,
"distances":distances
}
if pickle_data:
pickle.dump(data,open("results/autoencoder/{0}_0.pkl".format(name),"wb"))
if save_data:
self.save_population(sampled_population,0)
self.save_training_data(good_strings,0)
random_pop = [self.generate_random_string(genome_length) for z in range(1000)]
print random_pop[0]
sampled_r_pop = self.sample_dA(random_pop)
self.save_sampled_random_population([r for r in sampled_r_pop],0)
self.save_random_population(random_pop,0)
self.save_pop_fitnesses(sample_fitnesses,"fitness_pop",0)
self.save_pop_fitnesses(self.fitness_many(good_strings),"fitness_training_data",0)
print "writing..."
fitfile.write("{0},{1},{2},{3}\n".format(np.mean(original_fitnesses),np.min(original_fitnesses),np.max(original_fitnesses),np.std(original_fitnesses)))
fitfile.write("{0},{1},{2},{3}\n".format(np.mean(sample_fitnesses),np.min(sample_fitnesses),np.max(sample_fitnesses),np.std(original_fitnesses)))
print "{0},{1},{2}\n".format(np.mean(original_fitnesses),np.min(original_fitnesses),np.max(original_fitnesses))
print "{0},{1},{2}\n".format(np.mean(sample_fitnesses),np.min(sample_fitnesses),np.max(sample_fitnesses))
print "writing over"
for iteration in range(0,trials):
print "choosing pop:"
population = self.get_new_population_rw([z for z in sampled_population])
print "choosing pop over"
if online_training == False:
print "building model..."
self.dA = dA(n_visible=genome_length,n_hidden=50)
self.dA.build_dA(corruption_level)
self.build_sample_dA()
good_strings,good_strings_fitnesses=self.get_good_strings(population,population_limit)
for f in good_strings_fitnesses:
print "good_strings_fitnesses:",f
self.train_dA(ar(good_strings),corruption_level=corruption_level,num_epochs=num_epochs,lr=lr,output_folder=results_path,iteration=iteration+1)
print "sampling..."
sampled_population = self.sample_dA(population)
print "sampling over"
original_fitnesses,sample_fitnesses,differences,distances = self.get_statistics(population,sampled_population)
data = {
"original":original_fitnesses,
"fitnesses_sampled":sample_fitnesses,
"differences_in_fitness":differences,
"distances":distances
}
if pickle_data:
pickle.dump(data,open("results/autoencoder/{0}_{1}.pkl".format(name,iteration+1),"wb"))
fitfile.write("{0},{1},{2},{3}\n".format(np.mean(sample_fitnesses),np.min(sample_fitnesses),np.max(sample_fitnesses),np.std(original_fitnesses)))
fitfile.flush()
print "{0},{1},{2}\n".format(np.mean(sample_fitnesses),np.min(sample_fitnesses),np.max(sample_fitnesses))
if save_data:
self.save_population(sampled_population,iteration+1)
self.save_training_data(good_strings,iteration+1)
random_pop = [self.generate_random_string(genome_length) for z in range(1000)]
print random_pop[0]
sampled_r_pop = self.sample_dA(random_pop)
self.save_sampled_random_population([r for r in sampled_r_pop],iteration+1)
self.save_random_population(random_pop,iteration+1)
self.save_pop_fitnesses(sample_fitnesses,"fitness_pop",iteration+1)
self.save_pop_fitnesses(good_strings_fitnesses,"fitness_training_data",iteration+1)
fitfile.close()
def __init__(
self,
numpy_rng,
theano_rng=None,
n_ins=784,
hidden_layers_sizes=[500, 500],
n_outs=10,
corruption_levels=[0.1, 0.1],
):
""" This class is made to support a variable number of layers.
:type numpy_rng: numpy.random.RandomState
:param numpy_rng: numpy random number generator used to draw initial
weights
:type theano_rng: theano.tensor.shared_randomstreams.RandomStreams
:param theano_rng: Theano random generator; if None is given one is
generated based on a seed drawn from `rng`
:type n_ins: int
:param n_ins: dimension of the input to the sdA
:type n_layers_sizes: list of ints
:param n_layers_sizes: intermediate layers size, must contain
at least one value
:type n_outs: int
:param n_outs: dimension of the output of the network
:type corruption_levels: list of float
:param corruption_levels: amount of corruption to use for each
layer
"""
self.sigmoid_layers = []
self.dA_layers = []
self.params = []
self.n_layers = len(hidden_layers_sizes)
assert self.n_layers > 0
if not theano_rng:
theano_rng = RandomStreams(numpy_rng.randint(2 ** 30))
# allocate symbolic variables for the data
self.x = T.matrix("x") # the data is presented as rasterized images
self.y = T.ivector("y") # the labels are presented as 1D vector of
# [int] labels
# end-snippet-1
# The SdA is an MLP, for which all weights of intermediate layers
# are shared with a different denoising autoencoders
# We will first construct the SdA as a deep multilayer perceptron,
# and when constructing each sigmoidal layer we also construct a
# denoising autoencoder that shares weights with that layer
# During pretraining we will train these autoencoders (which will
# lead to chainging the weights of the MLP as well)
# During finetunining we will finish training the SdA by doing
# stochastich gradient descent on the MLP
# start-snippet-2
for i in xrange(self.n_layers):
# construct the sigmoidal layer
# the size of the input is either the number of hidden units of
# the layer below or the input size if we are on the first layer
if i == 0:
input_size = n_ins
else:
input_size = hidden_layers_sizes[i - 1]
# the input to this layer is either the activation of the hidden
# layer below or the input of the SdA if you are on the first
# layer
if i == 0:
layer_input = self.x
else:
layer_input = self.sigmoid_layers[-1].output
sigmoid_layer = HiddenLayer(
rng=numpy_rng,
input=layer_input,
n_in=input_size,
n_out=hidden_layers_sizes[i],
activation=T.nnet.sigmoid,
)
# add the layer to our list of layers
self.sigmoid_layers.append(sigmoid_layer)
# its arguably a philosophical question...
# but we are going to only declare that the parameters of the
# sigmoid_layers are parameters of the StackedDAA
# the visible biases in the dA are parameters of those
# dA, but not the SdA
self.params.extend(sigmoid_layer.params)
# Construct a denoising autoencoder that shared weights with this
# layer
dA_layer = dA(
numpy_rng=numpy_rng,
theano_rng=theano_rng,
input=layer_input,
n_visible=input_size,
n_hidden=hidden_layers_sizes[i],
W=sigmoid_layer.W,
bhid=sigmoid_layer.b,
)
self.dA_layers.append(dA_layer)
# end-snippet-2
# We now need to add a logistic layer on top of the MLP
self.logLayer = LogisticRegression(
input=self.sigmoid_layers[-1].output, n_in=hidden_layers_sizes[-1], n_out=n_outs
)
self.params.extend(self.logLayer.params)
# construct a function that implements one step of finetunining
# compute the cost for second phase of training,
# defined as the negative log likelihood
self.finetune_cost = self.logLayer.negative_log_likelihood(self.y)
# compute the gradients with respect to the model parameters
# symbolic variable that points to the number of errors made on the
# minibatch given by self.x and self.y
self.errors = self.logLayer.errors(self.y)
self.example = T.dvector("ex")
self.use_fn = theano.function(inputs=[self.x], outputs=self.logLayer.p_y_given_x, name="use")
def __init__(
self,
numpy_rng,
theano_rng=None,
n_ins=784, # 28 * 28
hidden_layers_sizes=[500, 500], # 两个 denoising autoencoding，后面简称 dA
n_outs=10, # 最上层的 MLP 的输出
corruption_levels=[0.1, 0.1] # 每个 dA 的 Denoising Level
):
""" This class is made to support a variable number of layers.
:type numpy_rng: numpy.random.RandomState
:param numpy_rng: numpy random number generator used to draw initial
weights
:type theano_rng: theano.tensor.shared_randomstreams.RandomStreams
:param theano_rng: Theano random generator; if None is given one is
generated based on a seed drawn from `rng`
:type n_ins: int
:param n_ins: dimension of the input to the sdA
:type n_layers_sizes: list of ints
:param n_layers_sizes: intermediate layers size, must contain
at least one value
:type n_outs: int
:param n_outs: dimension of the output of the network
:type corruption_levels: list of float
:param corruption_levels: amount of corruption to use for each
layer
"""
self.sigmoid_layers = [] # 只是用于构建 dA，而不是指上层的 MLP
self.dA_layers = []
self.params = []
self.n_layers = len(hidden_layers_sizes)
assert self.n_layers > 0
if not theano_rng:
theano_rng = RandomStreams(numpy_rng.randint(2**30))
# allocate symbolic variables for the data
self.x = T.matrix('x') # the data is presented as rasterized images
self.y = T.ivector('y') # the labels are presented as 1D vector of
# [int] labels
# end-snippet-1
# The SdA is an MLP, for which all weights of intermediate layers
# are shared with a different denoising autoencoders
# We will first construct the SdA as a deep multilayer perceptron,
# and when constructing each sigmoidal layer we also construct a
# denoising autoencoder that shares weights with that layer
# During pretraining we will train these autoencoders (which will
# lead to chainging the weights of the MLP as well)
# During finetunining we will finish training the SdA by doing
# stochastich gradient descent on the MLP
# start-snippet-2
for i in xrange(self.n_layers):
# construct the sigmoidal layer
# the size of the input is either the number of hidden units of
# the layer below or the input size if we are on the first layer
if i == 0:
input_size = n_ins
else:
input_size = hidden_layers_sizes[i - 1]
# the input to this layer is either the activation of the hidden
# layer below or the input of the SdA if you are on the first
# layer
if i == 0:
layer_input = self.x
else:
layer_input = self.sigmoid_layers[
-1].output # 这就是借用 HiddenLayer 的原因
sigmoid_layer = HiddenLayer(rng=numpy_rng,
input=layer_input,
n_in=input_size,
n_out=hidden_layers_sizes[i],
activation=T.nnet.sigmoid)
# add the layer to our list of layers
self.sigmoid_layers.append(sigmoid_layer)
# its arguably a philosophical question...
# but we are going to only declare that the parameters of the
# sigmoid_layers are parameters of the StackedDAA
# the visible biases in the dA are parameters of those
# dA, but not the SdA
self.params.extend(sigmoid_layer.params)
# Construct a denoising autoencoder that shared weights with this
# layer
dA_layer = dA(
numpy_rng=numpy_rng,
theano_rng=theano_rng,
input=layer_input,
n_visible=input_size,
n_hidden=hidden_layers_sizes[i],
W=sigmoid_layer.W, # 借用 W 参数
bhid=sigmoid_layer.b) # 借用 b 参数
self.dA_layers.append(dA_layer)
# end-snippet-2
# We now need to add a logistic layer on top of the MLP, 直接接在第二个 dA 层的后面
self.logLayer = LogReg(input=self.sigmoid_layers[-1].output,
n_in=hidden_layers_sizes[-1],
n_out=n_outs)
self.params.extend(self.logLayer.params)
# construct a function that implements one step of finetunining
# compute the cost for second phase of training,
# defined as the negative log likelihood
self.finetune_cost = self.logLayer.negative_log_likelihood(self.y)
# compute the gradients with respect to the model parameters
# symbolic variable that points to the number of errors made on the
# minibatch given by self.x and self.y
self.errors = self.logLayer.errors(self.y)
def iterative_algorithm(
self,
name,
pop_size=100,
genome_length=20,
lim_percentage=20,
corruption_level=0.2,
num_epochs=50,
lr = 0.1,
max_evaluations=200000,
unique_training=False,
hiddens=40,
rtr = True
):
self.mask = np.random.binomial(1,0.5,genome_length)
trials = max_evaluations/pop_size
population_limit = int(pop_size*(lim_percentage/100.0))
self.dA = dA(n_visible=genome_length,n_hidden=hiddens)
self.dA.build_dA(corruption_level)
self.build_sample_dA()
new_population = np.random.binomial(1,0.5,(pop_size,genome_length))
self.population_fitnesses = self.fitness_many(new_population)
for iteration in range(0,trials):
print "iteration:",iteration
population = new_population
self.population = new_population
rw = self.tournament_selection_replacement(population)
good_strings,good_strings_fitnesses=self.get_good_strings(
population,
population_limit,
unique=unique_training,
fitnesses=self.population_fitnesses
)
print "training A/E"
training_data = np.array(good_strings)
self.train_dA(training_data,
num_epochs=num_epochs,
lr=lr)
print "sampling..."
sampled_population = np.array(self.sample_dA(rw),"b")
self.sample_fitnesses = self.fitness_many(sampled_population)
if rtr:
new_population = self.RTR(
population,
sampled_population,
population_fitnesses=self.population_fitnesses,
sample_fitnesses=self.sample_fitnesses,
w=pop_size/20
)
else:
new_population = sampled_population
new_population[0:1] = good_strings[0:1]
self.population_fitnesses = self.sample_fitnesses
self.population_fitnesses[0:1] = good_strings_fitnesses[0:1]
print "{0},{1},{2}\n".format(np.mean(self.population_fitnesses),
np.min(self.population_fitnesses),
np.max(self.population_fitnesses))
print "best from previous:",(
self.fitness(new_population[np.argmax(self.population_fitnesses)])
)
if np.max(self.population_fitnesses) == self.optimum:
pickle.dump({"pop":self.population,"fitnesses":self.population_fitnesses,"iteration":iteration},open("final_shit_ae.pkl","w"))
break
return new_population
its = 10
genome_l = 100
low = -50
high = 50
mask = np.array(np.loadtxt("mask.dat"),"b")
total_rows = 17
plt.figure(figsize=(7, 12))
gs = gridspec.GridSpec(total_rows, 2,hspace=0.5)
# plt.subplots_adjust( hspace=0.4 )
for iteration in range(1,20):
solution = np.array(np.loadtxt("rw_population_{0}.dat".format(iteration)),"b")
params = cPickle.load(open("best_params_{0}.pickle".format(iteration)))
n_visible = params[0].shape[0]
n_hiddens = params[0].shape[1]
ae = denoising_autoencoder.dA(n_visible=n_visible,n_hidden=n_hiddens,W=params[0],hbias=params[1],vbias=params[2])
ae.build_dA(corruption_level=0.1)
sample_func = theano.function([ae.input],ae.sample)
reconstruct_func = theano.function([ae.input],ae.z)
input_vector = solution[0]
# output_probabilites =
ax = plt.subplot(gs[0:2,:])
probabilities = []
for i in range(0,300):
input_vector=solution[i]
r=reconstruct_func([input_vector])[0]
r[np.where(mask==1)]=1-r[np.where(mask==1)]
probabilities.append(r)
