def test_sampling(self):
init = initilaze_topic_model()
init.initilize()
sampleman = Sampling(init.xcorpus, init.ycorpus)
sampleman.sampling(init.TOPICS, init.xcounts, init.ycounts, init.docid, init.different_word)
print sampleman.xcorpus
print sampleman.ycorpus
def fisher_sample(self):
if self.sampler is None:
print("preprocessing")
if self.dico_fisher is None:
print('you need to compute the fisher information first')
return
self.sampler = Sampling(self.build_mean(), self.dico_fisher)
print('sampling ok')
config = self.network.get_config()
if self.network.__class__.__name__=='Sequential':
new_model = Sequential.from_config(config)
else:
new_model = Model.from_config(config)
new_params = self.sampler.sample()
"""
means = self.sampler.mean
for key in means:
if np.max(np.abs(means[key] - new_params[key]))==0:
print key
print('kikou')
import pdb; pdb.set_trace()
"""
#tmp_prob = self.sampler.prob(new_params)
new_model.compile(loss=self.network.loss,
optimizer=str.lower(self.network.optimizer.__class__.__name__),
metrics = self.network.metrics)
new_model.set_weights(self.network.get_weights())
self.copy_weights(new_model, new_params)
return new_model
def __init__(self, ):
self.sampling_mpu_0x68 = Sampling(AK8963_ADDRESS, MPU9050_ADDRESS_68,
MPU9050_ADDRESS_68, 1, GFS_1000,
AFS_8G, AK8963_BIT_16,
AK8963_MODE_C100HZ)
self.sampling_mpu_0x69 = Sampling(AK8963_ADDRESS, MPU9050_ADDRESS_69,
None, 1, GFS_1000, AFS_8G,
AK8963_BIT_16, AK8963_MODE_C100HZ)
def test_sampling(self):
init = initilaze_topic_model()
init.initilize()
sampleman = Sampling(init.xcorpus, init.ycorpus)
sampleman.sampling(init.TOPICS, init.xcounts, init.ycounts, init.docid,
init.different_word)
print sampleman.xcorpus
print sampleman.ycorpus
def test_uniformly_sampling(self, test_input, expected):
batch_size = 4
order_provider = Sampling.get_uniformly_sampled_order(
test_input, 4, batch_size)
order = next(order_provider)
result = chain.from_iterable(
[[test_input[ind] for ind in order[i:i + batch_size]]
for i in range(0, len(test_input), batch_size)])
assert all(a == b for a, b in zip(result, expected))
def test_order_iterator(self, test_input):
order_provider = OrderProvider(
Sampling.get_random_order(len(test_input)))
order1 = next(order_provider)
order2 = next(order_provider)
assert all(a == b for a, b in zip(order1, order2))
order_provider.update()
order3 = next(order_provider)
assert any(a != b for a, b in zip(order1, order3))
def __create_sequences(self, X, labels, sampling=True):
sequences = self.tokenizer.texts_to_sequences(X)
data = pad_sequences(sequences, padding='post',
maxlen=self.params['max_length'])
indices = np.arange(data.shape[0])
np.random.shuffle(indices)
data = data[indices]
labels = labels[indices]
if sampling:
sample = Sampling(2., .5)
x_train, y_train = sample.perform_sampling(data, labels, [0, 1])
else:
x_train, y_train = data, labels
return x_train, y_train
def __init__(self, data_infile, fasttext_model_path, triplet_margin=0.1):
self.sampling = Sampling(data_infile, fasttext_model_path)
self.amount_negative_names = 1
self.triplet_margin = triplet_margin
self.anchor_margin = 0
self.loss_weights = {'synonym': 1, 'proto': 1}
torch.autograd.set_detect_anomaly(True)
def __init__(self, data_infile, fasttext_model_path, triplet_margin=0.1):
self.sampling = Sampling(data_infile, fasttext_model_path)
self.amount_negative_names = 1
self.triplet_margin = triplet_margin
self.anchor_margin = 0
self.loss_weights = {
'semantic_similarity': 1,
'contextual': 1,
'grounding': 1
}
torch.autograd.set_detect_anomaly(True)
def __init__(self, latent_dim, seed):
super(encoder, self).__init__()
np.random.seed(seed)
self.layer_1 = Conv2D(
filters=32,
kernel_size=(4, 4),
activation="relu",
strides=2,
padding="same",
kernel_initializer=tf.keras.initializers.HeNormal(seed))
self.layer_2 = Conv2D(
filters=32,
kernel_size=(4, 4),
activation="relu",
strides=2,
padding="same",
kernel_initializer=tf.keras.initializers.HeNormal(seed))
self.layer_3 = Conv2D(
filters=64,
kernel_size=(4, 4),
activation="relu",
strides=2,
padding="same",
kernel_initializer=tf.keras.initializers.HeNormal(seed))
self.layer_4 = Conv2D(
filters=64,
kernel_size=(4, 4),
activation="relu",
strides=2,
padding="same",
kernel_initializer=tf.keras.initializers.HeNormal(seed))
self.layer_5 = Dense(
units=128,
activation='relu',
kernel_initializer=tf.keras.initializers.HeNormal(seed))
self.dense_log_var = Dense(
units=latent_dim,
kernel_initializer=tf.keras.initializers.HeNormal(seed))
self.dense_mean = Dense(
units=latent_dim,
kernel_initializer=tf.keras.initializers.HeNormal(seed))
self.sampling = Sampling()
self.batch_norm_1 = BatchNormalization()
self.batch_norm_2 = BatchNormalization()
self.batch_norm_3 = BatchNormalization()
self.batch_norm_4 = BatchNormalization()
self.flatten = Flatten()
tokenizer = Tokenizer(num_words=MAX_NB_WORDS)
tokenizer.fit_on_texts(X)
sequences = tokenizer.texts_to_sequences(X)
word_index = tokenizer.word_index
data = pad_sequences(sequences, padding='post', maxlen=MAX_SEQUENCE_LENGTH)
print('Shape of data tensor:', data.shape)
print('Shape of label tensor:', labels.shape)
indices = np.arange(data.shape[0])
np.random.shuffle(indices)
data = data[indices]
labels = labels[indices]
num_validation_samples = int(VALIDATION_SPLIT * data.shape[0])
sample = Sampling(2., .5)
x_train, y_train = sample.perform_sampling(data[:-num_validation_samples],
labels[:-num_validation_samples],
[0, 1])
x_val = data[-num_validation_samples:]
y_val = labels[-num_validation_samples:]
print('Number of entries in each category:')
print('training: ', y_train.sum(axis=0))
print('validation: ', y_val.sum(axis=0))
model = Word2Vec.load('1ft.modelFile')
embeddings_index = {}
embedding_matrix = np.random.random((len(word_index) + 1, EMBEDDING_DIM))
for word, i in word_index.items():
embedding_vector = model.wv[word]
#!/usr/bin/env python
import numpy as np
import fitsio
import scipy.optimize as optimize
from sampling import Sampling
sampling = Sampling(nsamples=1000)
sampling.set_flux(total_flux=1000., noise=0.001)
def mem_function(u, A, f, llambda):
Ar = (A.dot(u) - f)
As = (Ar**2).sum()
Bs = (u * np.log(u)).sum()
val = As + llambda * Bs
grad = 2. * A.T.dot(Ar) + llambda * (1. + np.log(u))
return (val, grad)
def mem_fit(sampling, llambda=1.e-2):
S_M0 = np.ones(sampling.nx * sampling.ny)
bounds = zip([1.e-5] * len(S_M0), [None] * len(S_M0))
bounds = [x for x in bounds]
flux = sampling.flux
results = optimize.minimize(mem_function,
S_M0,
args=(sampling.A, flux, llambda),
method='L-BFGS-B',
jac=True,
bounds=bounds)
def test_random_sampling(self, test_input, expected):
order_provider = Sampling.get_random_order(test_input)
order = next(order_provider)
result = set(order)
assert result == expected
def comp_sketch(matrix, objective, load_N=False, save_N=False, N_dir='../N_file/', **kwargs):
"""
Given matrix A, the function comp_sketch computes a sketch for A and performs further operations on PA.
It returns the total running time and the desired quantity.
parameter:
matrix: a RowMatrix object storing the matrix [A b]
objective: either 'x' or 'N'
'x': the function returns the solution to the problem min_x || PA[:,:-1]x - PA[:,-1] ||_2
'N': the function returns a square matrix N such that PA[:,:-1]*inv(N) is a matrix with orthonormal columns
load_N: load the precomputed N matrices if possible (it reduces the actual running time for sampling sketches)
save_N: save the computed N matrices for future use
sketch_type: either 'projection' or 'sampling'
projection_type: cw, gaussian, rademacher or srdht
c: projection size
s: sampling size (for sampling sketch only)
k: number of independent trials to run
"""
sketch_type = kwargs.get('sketch_type')
if not os.path.exists(N_dir):
os.makedirs(N_dir)
if objective == 'x':
if sketch_type == 'projection':
projection = Projections(**kwargs)
t = time.time()
x = projection.execute(matrix, 'x', save_N)
t = time.time() - t
if save_N:
logger.info('Saving N matrices from projections!')
N = [a[0] for a in x]
x = [a[1] for a in x]
# saving N
filename = N_dir + 'N_' + matrix.name + '_projection_' + kwargs.get('projection_type') + '_c' + str(int(kwargs.get('c'))) + '_k' + str(int(kwargs.get('k')))+ '.dat'
data = {'N': N, 'time': t}
pickle_write(filename,data)
elif sketch_type == 'sampling':
s = kwargs.get('s')
new_N_proj = 0
N_proj_filename = N_dir + 'N_' + matrix.name + '_projection_' + kwargs.get('projection_type') + '_c' + str(int(kwargs.get('c'))) + '_k' + str(int(kwargs.get('k'))) +'.dat'
if load_N and os.path.isfile(N_proj_filename):
logger.info('Found N matrices from projections, loading them!')
N_proj_filename = N_dir + 'N_' + matrix.name + '_projection_' + kwargs.get('projection_type') + '_c' + str(int(kwargs.get('c'))) + '_k' + str(int(kwargs.get('k'))) +'.dat'
result = pickle_load(N_proj_filename)
N_proj = result['N']
t_proj = result['time']
else: # otherwise, compute it
t = time.time()
projection = Projections(**kwargs)
N_proj = projection.execute(matrix, 'N')
t_proj = time.time() - t
new_N_proj = 1
sampling = Sampling(N=N_proj)
t = time.time()
x = sampling.execute(matrix, 'x', s, save_N )
t = time.time() - t + t_proj
if save_N and new_N_proj:
logger.info('Saving N matrices from projections!')
#filename = N_dir + 'N_' + matrix.name + '_projection_' + kwargs.get('projection_type') + '_c' + str(int(kwargs.get('c'))) + '_k' + str(int(kwargs.get('k'))) + '.dat'
data = {'N': N_proj, 'time': t_proj}
pickle_write(N_proj_filename,data)
if save_N:
logger.info('Saving N matrices from sampling!')
N = [a[0] for a in x]
x = [a[1] for a in x]
filename = N_dir + 'N_' + matrix.name + '_sampling_s' + str(int(kwargs.get('s'))) + '_' + kwargs.get('projection_type') + '_c' + str(int(kwargs.get('c'))) + '_k' + str(int(kwargs.get('k'))) + '.dat'
data = {'N': N, 'time': t}
pickle_write(filename,data)
else:
raise ValueError('Please enter a valid sketch type!')
return x, t
elif objective == 'N':
if sketch_type == 'projection':
N_proj_filename = N_dir + 'N_' + matrix.name + '_projection_' + kwargs.get('projection_type') + '_c' + str(int(kwargs.get('c'))) + '_k' + str(int(kwargs.get('k'))) + '.dat'
if load_N and os.path.isfile(N_proj_filename):
logger.info('Found N matrices from projections, loading them!')
result = pickle_load(N_proj_filename)
N = result['N']
t = result['time']
else:
t = time.time()
projection = Projections(**kwargs)
N = projection.execute(matrix, 'N')
t = time.time() - t
if save_N:
logger.info('Saving N matrices from projections!')
data = {'N': N, 'time': t}
pickle_write(N_proj_filename,data)
elif sketch_type == 'sampling':
s = kwargs.get('s')
new_N_proj = 0
new_N_samp = 0
N_samp_filename = N_dir + 'N_' + matrix.name + '_sampling_s' + str(int(kwargs.get('s'))) + '_' + kwargs.get('projection_type') + '_c' + str(int(kwargs.get('c'))) + '_k' + str(int(kwargs.get('k'))) + '.dat'
N_proj_filename = N_dir + 'N_' + matrix.name + '_projection_' + kwargs.get('projection_type') + '_c' + str(int(kwargs.get('c'))) + '_k' + str(int(kwargs.get('k'))) + '.dat'
if load_N and os.path.isfile(N_samp_filename):
logger.info('Found N matrices from sampling, loading them!')
result = pickle_load(N_samp_filename)
N = result['N']
t = result['time']
elif load_N and os.path.isfile(N_proj_filename):
logger.info('Found N matrices from projections, loading them!')
result = pickle_load(N_proj_filename)
N_proj = result['N']
t_proj = result['time']
sampling = Sampling(N=N_proj)
t = time.time()
N = sampling.execute(matrix, 'N', s)
t = time.time() - t + t_proj
new_N_samp = 1
else:
t = time.time()
projection = Projections(**kwargs)
N_proj = projection.execute(matrix, 'N')
t_proj = time.time() - t
new_N_proj = 1
t = time.time()
sampling = Sampling(N=N_proj)
N = sampling.execute(matrix, 'N', s)
t = time.time() - t + t_proj
new_N_samp = 1
if save_N and new_N_proj:
logger.info('Saving N matrices from projections!')
data = {'N': N_proj, 'time': t_proj}
pickle_write(N_proj_filename,data)
if save_N and new_N_samp:
logger.info('Saving N matrices from sampling!')
data = {'N': N, 'time': t}
pickle_write(N_samp_filename,data)
else:
raise ValueError('Please enter a valid sketch type!')
return N, t
else:
raise ValueError('Please enter a valid objective!')
path1 = "/Users/liumeiyu/Downloads/IMG_7575.JPG" path2 = "/Users/liumeiyu/Downloads/test1.jpg" path3 = "/Users/liumeiyu/Downloads/test2.jpg" A = Histogram(path1) B = Smooth(path1) C = Change(path1) D = Base(path1) E = Binary(path1) F = D_E(path1) G = Warp(path1) H = Cvt(path1) K = Edge_detection(path1) L = Segmentation(path1) M = Mosaic(path3) N = Sampling(path1) P = Fusion(path1) # A.img_histogram_trans() # A.img_histogram() # B.linear_smooth_np() # B.linear_smooth() # B.box_smooth() # B.gaussian_smooth() # B.median_smooth() # B.median_smooth_x(5) # C.fft_high_change(60) # C.change_cv()
def s_res_width(img, width, smooth_type):
ratio_scale = (width * 1.0) / img.shape[1]
height = int(ratio_scale * img.shape[0])
nimg = s.smooth_resize(img, height, width, smooth_type)
return nimg
def test_sentence_size_sorted_sampling(self, test_input, expected):
order_provider = Sampling.get_sentence_size_sorted_order(test_input, 2)
order = next(order_provider)
result = set(order)
assert result == expected
class Sensors:
sampling_mpu_0x68 = None
sampling_mpu_0x69 = None
running = False
def __init__(self, ):
self.sampling_mpu_0x68 = Sampling(AK8963_ADDRESS, MPU9050_ADDRESS_68,
MPU9050_ADDRESS_68, 1, GFS_1000,
AFS_8G, AK8963_BIT_16,
AK8963_MODE_C100HZ)
self.sampling_mpu_0x69 = Sampling(AK8963_ADDRESS, MPU9050_ADDRESS_69,
None, 1, GFS_1000, AFS_8G,
AK8963_BIT_16, AK8963_MODE_C100HZ)
# self.calibrate()
def configure(self):
self.sampling_mpu_0x68.configure()
self.sampling_mpu_0x69.configure()
def reset(self):
self.sampling_mpu_0x68.reset()
self.sampling_mpu_0x69.reset()
# self.configure()
def calibrate(self):
self.sampling_mpu_0x68.calibrate()
self.sampling_mpu_0x69.calibrate()
# self.configure()
def start(self):
timeSync = time.time()
self.sampling_mpu_0x68.startSampling(timeSync)
self.sampling_mpu_0x69.startSampling(timeSync)
self.running = True
def stop(self):
if self.running:
self.sampling_mpu_0x68.stopSampling()
self.sampling_mpu_0x69.stopSampling()
self.running = False
def showCurrent(self):
def formatValue(array, index):
format = "{: 4.17f}"
if len(array) > index:
return format.format(array[index])
else:
" null "
def formatLabel(value):
return value.center(20)
data_0x68 = self.sampling_mpu_0x68.getAllData()
data_0x69 = self.sampling_mpu_0x69.getAllData()
print(
"----------------------------------------------------------------------------------------------------------",
"\n",
"Time: ",
time.time(),
"\n",
"----------------------------------------------------------------------------------------------------------",
"\n",
"MPU = ",
formatLabel("0x68_master"),
" | ",
formatLabel("0x68_slave_of_0x68"),
" | ",
formatLabel("0x69_master"),
" | ",
formatLabel("0x68_slave_of_0x69"),
"\n",
"----------------------------------------------------------------------------------------------------------",
"\n",
"A_X = ",
formatValue(data_0x68, 1),
" | ",
formatValue(data_0x68, 7),
" | ",
formatValue(data_0x69, 1),
" | ",
formatValue(data_0x69, 7),
"\n",
"A_Y = ",
formatValue(data_0x68, 2),
" | ",
formatValue(data_0x68, 8),
" | ",
formatValue(data_0x69, 2),
" | ",
formatValue(data_0x69, 8),
"\n",
"A_Z = ",
formatValue(data_0x68, 3),
" | ",
formatValue(data_0x68, 9),
" | ",
formatValue(data_0x69, 3),
" | ",
formatValue(data_0x69, 9),
"\n",
"----------------------------------------------------------------------------------------------------------",
"\n",
"G_X = ",
formatValue(data_0x68, 4),
" | ",
formatValue(data_0x68, 10),
" | ",
formatValue(data_0x69, 4),
" | ",
formatValue(data_0x69, 10),
"\n",
"G_Y = ",
formatValue(data_0x68, 5),
" | ",
formatValue(data_0x68, 11),
" | ",
formatValue(data_0x69, 5),
" | ",
formatValue(data_0x69, 11),
"\n",
"G_Z = ",
formatValue(data_0x68, 6),
" | ",
formatValue(data_0x68, 12),
" | ",
formatValue(data_0x69, 6),
" | ",
formatValue(data_0x69, 12),
"\n",
"----------------------------------------------------------------------------------------------------------",
"\n",
"M_X = ",
formatValue(data_0x68, 13),
" | ",
" null ",
" | ",
formatValue(data_0x69, 13),
" | ",
" null ",
"\n",
"M_Y = ",
formatValue(data_0x68, 14),
" | ",
" null ",
" | ",
formatValue(data_0x69, 14),
" | ",
" null ",
"\n",
"M_Z = ",
formatValue(data_0x68, 15),
" | ",
" null ",
" | ",
formatValue(data_0x69, 15),
" | ",
" null ",
"\n",
"----------------------------------------------------------------------------------------------------------",
"\n",
)
with open(path, "r", encoding=encoding) as f:
for line in f:
words = tokenize(line.strip())
if len(words) < window_size + 1:
continue
for i in range(len(words)):
example = (
words[max(0, i - window_size):i] +
words[min(i + 1, len(words)
):min(len(words), i + window_size) + 1],
words[i])
examples.append(Example.fromlist(example, fields))
super(CBOWDataset, self).__init__(examples, fields, **kwargs)
if __name__ == '__main__':
test_path = '/home/lightsmile/NLP/corpus/novel/test.txt'
dataset = CBOWDataset(test_path, Fields)
print(len(dataset))
print(dataset[0])
print(dataset[0].context)
print(dataset[0].target)
TARGET.build_vocab(dataset)
from sampling import Sampling
samp = Sampling(TARGET.vocab)
print(samp.sampling(3))
class Fisher(object):
def __init__(self, network):
"""
proper documentation
"""
# remove dropout from the network if needed
self.network = network
self.f = None
self.stochastic_layers = {}
self.filter_layer()
layers = self.network.layers
for layer in layers:
if layer.name in self.stochastic_layers:
tmp = self.stochastic_layers[layer.name]
setattr(layer, tmp[0], tmp[2])
layers = self.network.layers
intermediate_layers_input = []
intermediate_layers_output = []
for layer in layers:
if re.match('merge_(.*)', layer.name):
intermediate_layers_input.append(layer.input[0])
intermediate_layers_output.append(layer.output)
else:
intermediate_layers_input.append(layer.input)
intermediate_layers_output.append(layer.output)
self.intermediate_input = Model(self.network.input, [input_ for input_ in intermediate_layers_input])
self.intermediate_output = Model(self.network.input, [output_ for output_ in intermediate_layers_output])
self.f = None
layers = self.network.layers
for layer in layers:
if layer.name in self.stochastic_layers:
tmp = self.stochastic_layers[layer.name]
setattr(layer, tmp[0], tmp[1])
self.dico_fisher = None
self.sampler = None
def filter_layer(self):
layers = self.network.layers
for layer in layers:
if re.match('dropout_(.*)', layer.name):
self.stochastic_layers[layer.name]=['p', layer.p, 0]
if re.match('batchnormalization_(.*)', layer.name):
self.stochastic_layers[layer.name]=['mode',layer.mode, 1]
def fisher_information(self, X, Y, recompute=False):
if not(recompute) and not(self.dico_fisher is None):
return self.dico_fisher
assert X.ndim==4, ("X must be a 4d tensor of shape (num_elem, num_channels, width, height) but has %d dimensions", X.ndim)
assert Y.ndim<=2, ("Y contains the class label and should be of size (num_elem, 1) but has %d dimensions", Y.ndim)
if Y.ndim==1:
#Y = Y.dimshuffle((0, 'x'))
Y = Y[:,None]
if Y.ndim==2 and Y.shape[1]>1:
Y = np.argmax(Y, axis=1)[:,None]
layers = self.network.layers
for layer in layers:
if layer.name in self.stochastic_layers:
tmp = self.stochastic_layers[layer.name]
setattr(layer, tmp[0], tmp[2])
dico_fisher, f = build_fisher(X, Y, self.network, self.intermediate_input, self.intermediate_output, f =self.f) # remove break
self.f = f
dico_conv_biases = build_fisher_biases(X, Y, self.network)
for key in dico_conv_biases:
dico_fisher[key] = dico_conv_biases[key]
for layer in layers:
if layer.name in self.stochastic_layers:
tmp = self.stochastic_layers[layer.name]
setattr(layer, tmp[0], tmp[1])
# TO DO INVERSE LAYERS !!!!!!!
self.dico_fisher = dico_fisher
return self.dico_fisher
def fisher_queries(self, X, Y):
assert X.ndim==4, ("X must be a 4d tensor of shape (num_elem, num_channels, width, height) but has %d dimensions", X.ndim)
assert Y.ndim<=2, ("Y contains the class label and should be of size (num_elem, 1) but has %d dimensions", Y.ndim)
if Y.ndim==1:
#Y = Y.dimshuffle((0, 'x'))
Y = Y[:,None]
if Y.ndim==2 and Y.shape[1]>1:
Y = np.argmax(Y, axis=1)[:,None]
layers = self.network.layers
for layer in layers:
if layer.name in self.stochastic_layers:
tmp = self.stochastic_layers[layer.name]
setattr(layer, tmp[0], tmp[2])
dico_fisher = build_queries(X, Y, self.network, self.intermediate_input, self.intermediate_output) # remove break
"""
dico_conv_biases = build_fisher_biases(X, Y, self.network)
for key in dico_conv_biases:
dico_fisher[key] = dico_conv_biases[key]
for layer in layers:
if layer.name in self.stochastic_layers:
tmp = self.stochastic_layers[layer.name]
setattr(layer, tmp[0], tmp[1])
# TO DO INVERSE LAYERS !!!!!!!
"""
return dico_fisher
def save(self, repo, filename):
with closing(open(os.path.join(repo, filename), 'wb')) as f:
pkl.dump(self.dico_fisher, f, protocol=pkl.HIGHEST_PROTOCOL)
def load(self, repo, filename):
with closing(open(os.path.join(repo, filename), 'rb')) as f:
self.dico_fisher = pkl.load(f)
# temporary
#self.dico_fisher= dict([('conv_1_bias', self.dico_fisher['conv_1_bias']), ('conv_2_bias', self.dico_fisher['conv_2_bias'])])
def build_mean(self):
if self.dico_fisher is None:
print('you need to compute the fisher information first')
return {}
layers = self.network.layers
dico_mean = {}
for layer in layers:
if re.match('convolution(.*)', layer.name):
nb_layer = layer.name.split('_')[1]
# attention the bias is separated from the weigths
W, b = layer.get_weights()
dico_mean['conv_'+nb_layer] = W.flatten()
dico_mean['conv_'+nb_layer+'_bias'] = b.flatten()
elif re.match('dense_(.*)', layer.name):
W, b = layer.get_weights()
dico_mean[layer.name]=np.concatenate([W.flatten(), b.flatten()], axis=0)
elif re.match('batchnormalization_(.*)', layer.name):
gamma = layer.gamma.get_value()
beta = layer.beta.get_value()
dico_mean[layer.name]=np.concatenate([gamma.flatten(), beta.flatten()], axis=0)
return dico_mean
def copy_weights(self, model, dico_weights):
layers = model.layers
for layer in layers:
if re.match('convolution(.*)input(.*)', layer.name):
continue
if re.match('convolution(.*)', layer.name):
nb_layer = layer.name.split('_')[1]
name_W = 'conv_'+nb_layer
if name_W in dico_weights.keys():
W = layer.W;
W_value = dico_weights['conv_'+nb_layer].astype('float32')
W_shape = W.shape.eval()
W.set_value(W_value.reshape(W_shape))
name_b = 'conv_'+nb_layer+'_bias'
if name_b in dico_weights.keys():
b = layer.b
b_value = dico_weights['conv_'+nb_layer+'_bias'].astype('float32')
b_shape = b.shape.eval()
b.set_value(b_value.reshape(b_shape))
elif re.match('dense_(.*)', layer.name):
W = layer.W; b = layer.b
W_shape = layer.W.shape.eval()
b_shape = layer.b.shape.eval()
if layer.name in dico_weights.keys():
params_W_b = dico_weights[layer.name].astype('float32')
split = len(params_W_b) - np.prod(b_shape)
W_value = params_W_b[:split]
b_value = params_W_b[split:]
W.set_value(W_value.reshape(W_shape))
b.set_value(b_value.reshape(b_shape))
elif re.match('batchnormalization_(.*)', layer.name):
gamma = layer.gamma
beta = layer.beta
gamma_shape = gamma.shape.eval()
beta_shape = beta.shape.eval()
gamma_split = np.prod(gamma_shape)
if layer.name in dico_weights.keys():
params_gamma_beta = dico_weights[layer.name].astype('float32')
gamma_value = params_gamma_beta[:gamma_split]
beta_value = params_gamma_beta[gamma_split:]
gamma.set_value(gamma_value.reshape(gamma_shape))
beta.set_value(beta_value.reshape(beta_shape))
def fisher_sample(self):
if self.sampler is None:
print("preprocessing")
if self.dico_fisher is None:
print('you need to compute the fisher information first')
return
self.sampler = Sampling(self.build_mean(), self.dico_fisher)
print('sampling ok')
config = self.network.get_config()
if self.network.__class__.__name__=='Sequential':
new_model = Sequential.from_config(config)
else:
new_model = Model.from_config(config)
new_params = self.sampler.sample()
"""
means = self.sampler.mean
for key in means:
if np.max(np.abs(means[key] - new_params[key]))==0:
print key
print('kikou')
import pdb; pdb.set_trace()
"""
#tmp_prob = self.sampler.prob(new_params)
new_model.compile(loss=self.network.loss,
optimizer=str.lower(self.network.optimizer.__class__.__name__),
metrics = self.network.metrics)
new_model.set_weights(self.network.get_weights())
self.copy_weights(new_model, new_params)
return new_model
"""
def sample_ensemble(self, data, N=2, nb_classe=1):
(X_train, Y_train), (X_test, Y_test) = data
def evaluate(model, data_train):
(X_train,Y_train) = data_train
yPreds = model.predict(X_train)
yPred = np.argmax(yPreds, axis=1)
yTrue = Y_train
return metrics.accuracy_score(yTrue, yPred)
models = [self.network] + [self.fisher_sample() for i in range(N)]
probabilities = []
data_train = (X_train, Y_train)
import pdb
for model in models:
var = evaluate(model, data_train)
probabilities.append(var)
#probabilities = [evaluate(model, (X_train, Y_train)) for model in models]
probabilities /= sum(probabilities)
yPreds_ensemble = np.mean([alpha*model.predict(X_test) for alpha, model in zip(probabilities,models)], axis=0)
yPred = np.argmax(yPreds_ensemble, axis=1)
yTrue = Y_test
accuracy = metrics.accuracy_score(yTrue, yPred) * 100
print("Accuracy : ", accuracy)
"""
def sample_ensemble(self, data, N=2, nb_classe=1):
(X_train, Y_train), (X_test, Y_test) = data
models = [self.fisher_sample() for i in range(N)]
N-=1
yPreds_ensemble = np.array([model.predict(X_test) for model in models]) #(N+1, 10000, 10)
committee = np.mean(yPreds_ensemble, axis=0) # shape(10000, 10)
def kl_divergence(committee_proba, member_proba):
# for a fixed i
# for a fixed j
yPred=[]
for n in range(len(Y_test)):
predict = []
for i in range(10):
proba_i_C = committee_proba[n,i]
predict_i=0
for j in range(N+1):
proba_i_j = member_proba[j,n, i]
predict_i += np.log(proba_i_j)*np.log(proba_i_j/proba_i_C)
predict.append(predict_i)
yPred.append(np.argmin(predict))
return np.array(yPred).astype('uint8')
yPred = kl_divergence(committee, yPreds_ensemble)
yTrue = Y_test
accuracy = metrics.accuracy_score(yTrue, yPred) * 100
print("Accuracy : ", accuracy)
def correlation_sampling(self, data, N=3, n=5):
(X_train, Y_train), (X_test, Y_test) = data
predict_ensemble = [self.network.predict(X_train)]
ensemble_model = [self.network]
for l in range(N):
p_ensemble = np.mean(predict_ensemble, axis=0)
models = [self.fisher_sample() for p in range(n)]
predict = [model.predict(X_train) for model in models]
def evaluate(model, data_train):
(X_train,Y_train) = data_train
yPreds = model.predict(X_train)
yPred = np.argmax(yPreds, axis=1)
yTrue = Y_train
return metrics.accuracy_score(yTrue, yPred)
def kl(p_network, p_model):
kl=[]
n = len(Y_train)
for i in range(n):
kl_n = 0
for j in range(10):
kl_n += p_network[i,j]*np.log(p_network[i,j]/p_model[i,j])
kl.append(kl_n)
return np.mean(kl)
index = np.argmax([kl(p_ensemble, predict[j]) for j in range(n)])
model = models[index]
ensemble_model.append(model)
predict_ensemble.append(predict[index])
# Accuracy
#Ypred = np.argmax(np.mean([network.predict(X_test) for network in [self.network, model]], axis=0), axis=1)
Ypred = np.argmax(np.mean([model.predict(X_test) for model in ensemble_model]))
Ytrue = Y_test
print(metrics.accuracy_score(Ytrue, Ypred))
def sample_ensemble(self, data, N=2, nb_classe=1):
(X_train, Y_train), (X_test, Y_test) = data
models = [self.network]; scores=[1.]
for i in range(N):
model, score = self.fisher_sample()
models.append(model)
scores.append(score)
#models = [self.network] + [self.fisher_sample() for i in range(N)]
min_prob = np.min(scores)
scores -= min_prob
scores = np.exp(scores)
print(scores)
"""
def evaluate(model, data_train):
(X_train,Y_train) = data_train
yPreds = model.predict(X_train)
yPred = np.argmax(yPreds, axis=1)
yTrue = Y_train
return metrics.accuracy_score(yTrue, yPred)
"""
#probabilities = [evaluate(model, (X_train, Y_train)) for model in models]
#probabilities /= sum(probabilities)
#yPreds_ensemble = np.array([alpha*model.predict(X_test) for alpha, model in zip(scores,models)]) #(N+1, 10000, 10)
#yPreds_network = np.mean(self.network.predict(X_test), axis=0)
yPred = np.argmax(np.mean(np.array([alpha*model.predict(X_test) for alpha, model in zip(scores,models)]), axis=0), axis=1) #(N+1, 10000, 10)
"""
yPred = []
for n in range(len(Y_test)):
label = np.argmax(yPreds_ensemble[:,n,:]) % (N+1)
yPred.append(label)
if label !=Y_test[n]:
print((np.argmax(yPreds_ensemble[0,n]), label, Y_test[n]))
"""
yTrue = Y_test
accuracy = metrics.accuracy_score(yTrue, yPred) * 100
print("Accuracy : ", accuracy)
