def test_kmm(source, target):
kmm_kernel = 'rbf'
B = 1000
print('Kernel mean matching')
# from kernel_mean_matching import eprimical_kmm as ekmm
# coef_s,_ = ekmm(target, source, kern = kmm_kernel, B = B)
from kernel_mean_matching import kernel_mean_matching as kmm
coef_s = kmm(target, source, kern=kmm_kernel, B=B)
tr = coef_s > 0.1
tr = tr.reshape(tr.shape[0], )
from embedders import embedding, variable_embedder
n_components = 2
embedding_type = "no_embedding"
embedder = embedding(embedding_type, n_cmp=n_components, n_ngb=10)
if embedding_type == "autoencoder":
split = 0.3
cut = np.floor(target.shape[0] * (1 - split)).astype(int)
test_X = target[:cut, :]
val_X = target[cut:, :]
cut = np.floor(x_s.shape[0] * (1 - split)).astype(int)
test_C = source[:cut, :]
val_C = source[cut:, :]
emb_c, emb_val_c, emb_x, emb_val_x = variable_embedder(embedder,\
[test_C, test_X],[val_C, val_X])
source = np.concatenate((emb_c, emb_val_c), axis=0)
target = np.concatenate((emb_x, emb_val_x), axis=0)
else:
source, target, = variable_embedder(embedder, [source, target])
marker_size = 5
l1, = plt.plot(source[:, 0],
source[:, 1],
'o',
color='red',
label='source')
plt.setp(l1, markersize=marker_size)
l2, = plt.plot(target[:, 0],
target[:, 1],
'o',
color='blue',
label='target')
plt.setp(l2, markersize=marker_size)
marker_size = 2
l1, = plt.plot(source[:, 0],
source[:, 1],
'o',
color='red',
label='source')
plt.setp(l1, markersize=marker_size)
marker_size = 1
l3, = plt.plot(source[tr, 0], source[tr, 1], 'o', color='k', label='ssbc')
plt.setp(l3, markersize=marker_size)
plt.show()
def eprimical_kmm_emb(target_samples,
source_samples,
kern='rbf',
B=1,
embedder_type='autoencoder',
n_components=30):
from embedders import embedding, variable_embedder
embedder = embedding(embedder_type, n_cmp=n_components, n_ngb=10)
if embedder_type != "autoencoder":
X, C = variable_embedder(embedder, [target_samples, source_samples])
else:
from autoencoder import autoencoder
split = 0.3
cut = np.floor(target_samples.shape[0] * (1 - split)).astype(int)
test_X = target_samples[:cut, :]
val_X = target_samples[cut:, :]
cut = np.floor(source_samples.shape[0] * (1 - split)).astype(int)
test_C = source_samples[:cut, :]
val_C = source_samples[cut:, :]
test_X, val_X, test_C, val_C\
= variable_embedder(embedder,\
[test_X, test_C],[val_X, val_C])
X = np.concatenate((test_X, val_X), axis=0)
C = np.concatenate((test_C, val_C), axis=0)
# =============================================================================
# from embedders import variable_embedder
# embeded_output = variable_embedder(embedder, (source_samples, target_samples))
# C = embeded_output[0]
# X = embeded_output[1]
# =============================================================================
if embedder_type == 'tsne' or embedder_type == "autoencoder":
X = X.astype(np.double)
C = C.astype(np.double)
coef_s = kernel_mean_matching(X, C, kern=kern, B=B)
coef_t = []
# coef_s, coef_t = eprimical_kmm(X, C , kern=kern, B=B)
return coef_s, coef_t
def embedd():
##############################################################################
from embedders import embedding, variable_embedder
n_components = 5
embedding_type = "spectral"
embedder = embedding(embedding_type, n_cmp=n_components, n_ngb=30)
##############################################################################
if embedding_type == "autoencoder":
emb_train_x_s, emb_train_val_x_s, emb_train_x_t, emb_train_val_x_t\
= variable_embedder(embedder,\
[train_x_s, train_x_t],[val_x_s, val_x_t])
else:
emb_train_x_s, emb_train_val_x_s, emb_train_x_t, emb_train_val_x_t\
= variable_embedder(embedder,\
[train_x_s, val_x_s, train_x_t, val_x_t])
# if embedding_type == "autoencoder":
# emb_train_x_s, emb_train_val_x_s\
# = embedder.fit_transform(train_x_s, val_x_s)
# else:
# emb_train_x_s, emb_train_val_x_s\
# = variable_embedder(embedder, [train_x_s, val_x_s])
plt.scatter(emb_train_x_s[:, 0],
emb_train_x_s[:, 1],
c=cmap_s,
label='source',
marker="o",
s=5)
plt.show()
plt.scatter(emb_train_x_t[:, 0],
emb_train_x_t[:, 1],
c=cmap_t,
label='source',
marker="o",
s=5)
plt.show()
import matplotlib.pyplot as plt
import matplotlib.colors as colors
import matplotlib.cm as cmx
from load_data import load
from kernel_mean_matching import eprimical_kmm_emb as ekmm_emb
from kernel_mean_matching import kernel_mean_matching as kmm
train_x_s, train_y_s, val_x_s, val_y_s, train_x_t, train_y_t, \
val_x_t, val_y_t, test_x_t, test_y_t = load(unlabele_target_percentage = 1)
##############################################################################
from embedders import embedding, variable_embedder
n_components = 2
embedding_type = "no_embedding"
embedder = embedding(embedding_type, n_cmp=n_components, n_ngb=10)
##############################################################################
if embedding_type == "autoencoder":
train_x_s, val_x_s, train_x_t, val_x_t\
= variable_embedder(embedder,\
[train_x_s, train_x_t],[val_x_s, val_x_t])
test_x_t = embedder.predict(test_x_t)
else:
train_x_s, val_x_s, train_x_t, val_x_t, test_x_t \
= variable_embedder(embedder,[train_x_s, val_x_s,
train_x_t, val_x_t, test_x_t])
##############################################################################
treshhold = 0.5
##############################################################################
###### ssbc
print('Kernel mean matching')
# =============================================================================
print("Increazing the power level of the source floor with 50dB")
train_x_s[train_x_s != 110] = train_x_s[train_x_s != 110] + 80
val_x_s[val_x_s != 110] = val_x_s[val_x_s != 110] + 80
#normalize the vectors of Xs and Ys
def normalizer(X):
mean_X = np.mean(X)
std_X = np.sqrt(np.var(X))
X = (X-mean_X) / std_X
return X
from embedders import embedding
embedding_type = "no_embedding"
embedder = embedding(embedding_type)
# =============================================================================
emb_x_s = embedder.fit_transform(train_x_s)
emb_x_t = embedder.fit_transform(train_x_t)
emb_val_x_s = embedder.fit_transform(val_x_s)
emb_val_x_t = embedder.fit_transform(val_x_t)
emb_test_x = embedder.fit_transform(test_x_t)
# =============================================================================
num_inputs = emb_x_s.shape[1]# input layer size
# =============================================================================
model_obj = my_models(num_inputs, dropout = dropout_pr)
model = model_obj.build_model()
model = model_obj.fit(emb_x_s, train_y_s, emb_val_x_s, val_y_s,
scale = NN_scaling)
from plotters import plot_scatter_colored from plotters import plot_cmp from sklearn import preprocessing from keras.utils import plot_model from load_data import load train_x_s, train_y_s, val_x_s, val_y_s, train_x_t, train_y_t, \ val_x_t, val_y_t, test_x_t, test_y_t = load() load_data_flag = False from embedders import variable_embedder, embedding embedding_type = "pca" embedder = embedding(embedding_type, n_cmp = 519) train_x_s, val_x_s, train_x_t, val_x_t, test_x_t = variable_embedder(embedder,\ [train_x_s, val_x_s, train_x_t, val_x_t, test_x_t]) fine_tuning = True %run -i naive_learning.py fine_tuning = False %run -i naive_learning.py %run -i metric_learning_training.py %run -i sample_selection_bias_by_unlabeled_tranining.py %run -i transform_features.py %run -i transformed_sample_selection.py
def plot_embeding(x_s,
x_t,
coef_s,
train_y_s=None,
train_y_t=None,
fig_name=None):
import numpy as np
treshhold = 0.9
# treshhold = np.mean(coef_s)
coef_ind = coef_s > treshhold
coef_ind = coef_ind.reshape(coef_ind.shape[0], )
from embedders import embedding, variable_embedder
n_components = 2
embedding_type = "mds"
embedder = embedding(embedding_type, n_cmp=n_components, n_ngb=10)
if embedding_type == "autoencoder":
split = 0.3
cut = np.floor(x_t.shape[0] * (1 - split)).astype(int)
test_X = x_t[:cut, :]
val_X = x_t[cut:, :]
cut = np.floor(x_s.shape[0] * (1 - split)).astype(int)
test_C = x_s[:cut, :]
val_C = x_s[cut:, :]
emb_c, emb_val_c, emb_x, emb_val_x = variable_embedder(embedder,\
[test_C, test_X],[val_C, val_X])
emb_train_x_s = np.concatenate((emb_c, emb_val_c), axis=0)
emb_train_x_t = np.concatenate((emb_x, emb_val_x), axis=0)
else:
emb_train_x_s, emb_train_x_t, = variable_embedder(embedder, [x_s, x_t])
##############################################################################
import matplotlib.pyplot as plt
import seaborn as sns
sns.set_style("whitegrid")
bg_color = 'white'
fg_color = 'black'
f = plt.figure(figsize=(10, 10), facecolor=bg_color, edgecolor=fg_color)
if train_y_s is not None:
cmap_s = cmap2d(train_y_s)
cmap_t = cmap2d(train_y_t)
else:
cmap_s = 'b'
cmap_t = 'k'
l00 = plt.scatter(emb_train_x_s[:, 0],
emb_train_x_s[:, 1],
c=cmap_s,
label='source',
marker="o",
s=3)
l01 = plt.scatter(emb_train_x_t[:, 0],
emb_train_x_t[:, 1],
c=cmap_t,
label='target',
marker="^",
s=3)
l22 = plt.scatter(emb_train_x_s[coef_ind, 0],
emb_train_x_s[coef_ind, 1],
c='r',
label='ssbc',
s=0.5)
if fig_name:
plt.savefig(fig_name + ".svg", dpi=1200)
else:
plt.show()