def do_cca(X, y, X_orig, n_components=10, permutations=10):
'''
Performs a CCA using components
Projects scores back to edge space
'''
cca = CCA(n_components=n_components)
cca.fit(X, y)
# save the latent component correlation
cca.mode_r = []
for component in range(n_components):
cca.mode_r.append(
np.corrcoef(cca.x_scores_[:, component],
cca.y_scores_[:, component])[0, 1])
# correlate behaviour with LC score
cca.y_score_correlation = np.zeros((np.shape(y)[1], n_components))
for component in range(n_components):
for beh in range(np.shape(y)[1]):
cca.y_score_correlation[beh, component] = np.corrcoef(
y[:, beh].T, cca.y_scores_[:, component])[0, 1]
# correlate edges with LC score
cca.x_score_correlation = np.zeros((np.shape(X_orig)[1], n_components))
for component in range(n_components):
cca.x_score_correlation[:, component] = np.corrcoef(
cca.x_scores_[:, component], X_orig.T)[1::, 0]
# non parametric max T tests for component significance
max_r = []
for perm in tqdm(range(permutations)):
#shuffle the behaviour for each permutation
y_shuffle = shuffle(y)
#perform a new CCA with shuffled data
cca_perm = []
cca_perm = CCA(n_components=n_components)
cca_perm.fit(X, y_shuffle)
# save the latent component correlation
mode_r_perm = []
for component in range(n_components):
mode_r_perm.append(
np.corrcoef(cca_perm.x_scores_[:, component],
cca_perm.y_scores_[:, component])[0, 1])
# take the max r value
max_r.append(np.max(mode_r_perm))
# Compute adjusted p-values via percentile
p_adj = []
for component in range(n_components):
p_adj.append(np.mean(max_r >= cca.mode_r[component]))
return cca, p_adj
def run(data, template, reference):
n_components = 1
corrs = []
cca = CCA(n_components)
for target in range(0, len(st.frequencies)):
results = []
X = data
cor, _, xweights, yweights = su.find_correlation_for_one_pair(
cca, 1, data, template[target, :, :])
cor_ref, _, xweights_ref, yweights_ref = su.find_correlation_for_one_pair(
cca, 1, data, reference[target, :, :])
cor_ref_tem, _, xweights_ref_tem, yweights_ref_tem = su.find_correlation_for_one_pair(
cca, 1, np.squeeze(template[target, :, :]),
reference[target, :, :])
corr_t = get_cor_template(X, template[target, :, :], xweights_ref,
xweights_ref)
corr_ref_temp = get_cor_template(X, template[target, :, :],
xweights_ref_tem, xweights_ref_tem)
corr_temp = get_cor_template(np.squeeze(template[target, :, :]),
template[target, :, :], xweights,
yweights)
results = [cor, cor_ref, corr_t, corr_ref_temp, corr_temp]
corsum = map(lambda x: np.sign(x) * x**2, results)
corsum = list(corsum)
corsum = np.sum(corsum)
corrs.append(np.sum(corsum))
pre = np.argmax(corrs)
return pre
def fit(self, model):
"""Fits the model and creates a (random) orthogonal transformation.
Args:
model: string; a value in ["logistic", "svm", "linear", "svr", "cca"]
"""
if model == 'linear':
self.mod = LinearRegression()
elif model == 'logistic':
self.mod = LogisticRegression(penalty='none',
class_weight='balanced',
solver='saga')
#self.mod = LogisticRegression()
elif model == 'svr':
self.mod = SVR(kernel='linear')
elif model == 'svm':
self.mod = SVC(C=1.0, kernel='linear')
elif model == 'cca':
self.mod = CCA(n_components=1,
scale=True,
max_iter=500,
tol=1e-06,
copy=True)
self.mod.intercept_ = 0.0
self.mod.fit(self.Xrel, self.Yrel)
# now compute T with a random orthogonal basis
# todo potential bug: what to do with the intercept_?
w0 = self.mod.coef_ # + self.mod.intercept_
if len(w0.shape) < 2:
w0 = w0.reshape(1, -1)
w0 = w0 / np.linalg.norm(w0)
Wcompl = null_space(w0)
self.T = np.hstack((w0.transpose(), Wcompl))
assert np.allclose(self.T.transpose().dot(self.T),
np.eye(self.T.shape[0])), "self.T not orthonormal."
def compCorrCoefs(self, learningSet, EEGSignals):
n_components = 1
cca = CCA(n_components)
#print(EEGSignals.shape)
'''
correlation14 = abs(np.corrcoef(np.mean(learningSet[0:3].T, axis=1),np.mean(EEGSignals.T, axis=1))[0, 1])
correlation28 = abs(np.corrcoef(np.mean(learningSet[3:6].T, axis=1),np.mean(EEGSignals.T, axis=1))[0, 1])
correlation8 = abs(np.corrcoef(np.mean(learningSet[6:9].T, axis=1),np.mean(EEGSignals.T, axis=1))[0, 1])
print(learningSet[0][0],learningSet[1][0],learningSet[2][0])
for i in range(0,9,3):
print(abs(np.corrcoef(learningSet[i].T,EEGSignals[int(i/3)].T)[0, 1]),
abs(np.corrcoef(learningSet[i+1].T,EEGSignals[int(i/3)].T)[0, 1]),
abs(np.corrcoef(learningSet[i+2].T,EEGSignals[int(i/3)].T)[0, 1]))
print("---")
'''
cca.fit(learningSet[0:3].T, EEGSignals.T)
U, V = cca.transform(learningSet[0:3].T, EEGSignals.T)
correlation14 = abs(np.corrcoef(U.T, V.T)[0, 1])
cca.fit(learningSet[3:6].T, EEGSignals.T)
U, V = cca.transform(learningSet[3:6].T, EEGSignals.T)
correlation28 = abs(np.corrcoef(U.T, V.T)[0, 1])
cca.fit(learningSet[6:9].T, EEGSignals.T)
U, V = cca.transform(learningSet[6:9].T, EEGSignals.T)
correlation8 = abs(np.corrcoef(U.T, V.T)[0, 1])
return correlation14, correlation28, correlation8
def train_eval(self, train_index, test_index, ignore_eval=False):
normalized_train, normalized_test = normalize_by_train(self.source[train_index], self.source[test_index])
if self.comp is not None:
if self.use_scikit is not None:
if self.use_scikit == 'cca':
dim_reduction = CCA(n_components=self.comp)
else:
dim_reduction = PCA(n_components=self.comp)
# fit cca according to train data only
dim_reduction.fit(normalized_train, self.target[train_index])
# convert source into lower dimensional representation
normalized_train = dim_reduction.transform(normalized_train)
normalized_test = dim_reduction.transform(normalized_test)
else:
_, wa, _ = tutorial_on_cca(normalized_train, self.target[train_index])
normalized_train = normalized_train @ wa[:, :self.comp]
normalized_test = normalized_test @ wa[:, :self.comp]
model = self.build_model()
model.fit(normalized_train, self.target[train_index])
prediction = model.predict(normalized_test)
# res_df.to_csv(f"{self.out_name}/res1.csv")
if not ignore_eval:
return self.evaluate_regression(prediction, test_index)
else:
return prediction
def __init__(self,
model_name,
model_type,
n_clusters=None,
n_components=None,
n_lag=None,
regularisation=None):
self.n_lag = n_lag
self.model_name = model_name
self.model_type = model_type
# self.n_clusters = n_clusters
# self.clustering = NeuronClustering(self.n_clusters, signal_correlation)
if model_name == 'cca':
self.n_components = n_components
self.model = CCA(n_components=self.n_components)
elif model_name == 'linear-regression':
if regularisation is None:
self.model = LinearRegression()
elif regularisation == 'l1':
self.model = Lasso()
elif regularisation == 'l2':
self.model = Ridge()
elif regularisation == 'l1l2':
self.model = ElasticNet()
else:
raise NotImplementedError
def test_cca():
"""Test CCA."""
# Compare results with Matlab
# x = np.random.randn(1000, 11)
# y = np.random.randn(1000, 9)
# x = demean(x).squeeze()
# y = demean(y).squeeze()
mat = loadmat('./tests/data/ccadata.mat')
x = mat['x']
y = mat['y']
A2 = mat['A2']
B2 = mat['B2']
A1, B1, R = nt_cca(x, y) # if mean(A1(:).*A2(:))<0; A2=-A2; end
X1 = np.dot(x, A1)
Y1 = np.dot(y, B1)
C1 = tscov(np.hstack((X1, Y1)))[0]
# Sklearn CCA
cca = CCA(n_components=9, scale=False, max_iter=1e6)
X2, Y2 = cca.fit_transform(x, y)
# C2 = tscov(np.hstack((X2, Y2)).T)[0]
# import matplotlib.pyplot as plt
# f, (ax1, ax2) = plt.subplots(2, 1)
# ax1.imshow(C1)
# ax2.imshow(C2)
# plt.show()
# assert_almost_equal(C1, C2, decimal=4)
# Compare with matlab
X2 = np.dot(x, A2)
Y2 = np.dot(y, B2)
C2 = tscov(np.hstack((X2, Y2)))[0]
assert_almost_equal(C1, C2)
def compute_SVCCA(activation1, activation2):
'''
activation1 - Activation array 1 as a numpy array of size n X m1
activation2 - Activation array 2 as a numpy array of size n X m2
'''
pca_r = 40 # value from Shi et al NeurIPS 2019
n = activation1.shape[0]
assert n == activation2.shape[
0], "Size of activation arrays are different!!"
if pca_r > activation1.shape[1]:
print(
"Activation 1 array has less neurons.. changing number of PCs to ",
activation1.shape[1])
pca_r = activation1.shape[1]
if pca_r > activation2.shape[1]:
print(
"Activation 2 array has less neurons.. changing number of PCs to ",
activation2.shape[1])
pca_r = activation2.shape[1]
pca1 = PCA(n_components=pca_r)
red_activation1 = pca1.fit_transform(activation1)
pca2 = PCA(n_components=pca_r)
red_activation2 = pca2.fit_transform(activation2)
cca = CCA(n_components=pca_r)
red_activation1_c, red_activation2_c = cca.fit_transform(
red_activation1, red_activation2)
corr_values = np.zeros(pca_r)
for idx in range(pca_r):
corr_values[idx] = np.corrcoef(
red_activation1_c[:, idx],
red_activation2_c[:, idx])[0, 1] # get the off-diagonal element
return np.mean(corr_values)
def perform(arrs):
blocks_cnt = sum([arr.shape[1] * arr.shape[2] for arr in arrs])
X = np.zeros((blocks_cnt, 16))
Y = np.zeros((blocks_cnt, 64))
for c in range(3):
for i, arr in enumerate(arrs):
height = arr.shape[1]
width = arr.shape[2]
for y in range(height):
for x in range(width):
X[y * width + x] = np.hstack(
[arr[c][y][x - 1][:][-1], arr[c][y - 1][x][-1][:]])
Y[y * width + x] = arr[c][y][x].ravel()
X_mc = (X - X.mean()) / (X.std())
Y_mc = (Y - Y.mean()) / (Y.std())
ca = CCA(n_components=1)
ca.fit(X_mc, Y_mc)
print(f'\nColor {c}:')
weights = ca.x_weights_.ravel()
print(weights.shape)
print(', '.join(map(lambda a: str(a), weights)))
print(ca.n_iter_)
def cca_classify(X_eeg_signals, Yi_frequency_signals):
cca = CCA(1)
corr_results = []
for fr in range(0, Yi_frequency_signals.shape[0]):
X = X_eeg_signals
Yi = Yi_frequency_signals[fr, :, :]
#计算X与Yi之间的相关性
cca.fit(X.T, np.squeeze(Yi).T)
X_train_r, Yi_train_r = cca.transform(X.T, np.squeeze(Yi).T)
corr = np.corrcoef(X_train_r[:, 0], Yi_train_r[:, 0])[0, 1]
#得出X与每个Yi的相关性
corr_results.append(corr)
if corr_results[np.argmax(corr_results)] > 0.50:
#设置阈值
global index
global all_data
classify_result = np.argmax(corr_results) + 1
print(corr_results)
index += 1
#保存数据
TT = pd.DataFrame(X_eeg_signals)
all_data = all_data.append(np.transpose(TT[1:9]))
if index == 50:
#保存数据
all_data = pd.DataFrame(all_data)
all_data.to_csv('./j_8_all_data.csv', index=False)
return classify_result
else:
return -1
def Initialize(self):
self.configs = self.Config[0]
self.Channellist = self.configs['Channellist']
self.Trigger = {'state': 0, 'code': 0} #注册trigger
self.eeg = np.empty(0)
self.trigger = np.empty(0)
# self.mudname = rootdir+'/src/mud/MUD.mat'
# self.sigparm = sio.loadmat(self.mudname)
# self.FilterLR = self.sigparm['FilterLR']
n = 3
MdB = 20
bprange = np.array([6.0, 35.0])
Ws = bprange / (self.configs['SamplingRate'] / 2)
self.b, self.a = scipy.signal.iirfilter(n, Ws, rs=MdB,
ftype='cheby2') # Hd_Bandpass
self.prodata = np.empty(0)
self.frequency = [9, 11.7, 14.5]
t = np.arange(0.005, 5.5, 0.005)
self.Y = {}
for i in range(len(self.frequency)):
y = np.array([
np.sin(2 * np.pi * self.frequency[i] * t),
np.cos(2 * np.pi * self.frequency[i] * t),
np.sin(4 * np.pi * self.frequency[i] * t),
np.cos(4 * np.pi * self.frequency[i] * t),
np.sin(6 * np.pi * self.frequency[i] * t),
np.cos(6 * np.pi * self.frequency[i] * t)
])
self.Y[str(i)] = y
self.rank_min = min(len(self.Channellist), self.Y[str(0)].shape[0])
self.cca = CCA(n_components=self.rank_min)
self.score_threshold = 0.25 #####load .mat
def doCCA(metrics, color):
inp = np.array([metrics[m] for m in metricsInput2]).T.astype(float)
out = np.array([metrics[m] for m in metricsOutput2]).T.astype(float)
inp0 = np.zeros(len(metricsInput2))
out0 = np.zeros(len(metricsOutput2))
inp = np.vstack((inp, inp0))
out = np.vstack((out, out0))
cca = CCA(n_components=1, scale=False)
cca.fit(inp, out)
inp_cca = inp.dot(cca.x_weights_)
out_cca = out.dot(cca.y_weights_)
# Create linear regression object
regr = linear_model.LinearRegression()
# Train the model using the training sets
regr.fit(inp_cca, out_cca)
cca_regr = regr.predict(inp_cca)
# The coefficients
print('Coefficients: \n', regr.coef_)
plt.scatter(inp_cca, out_cca, c=color)
plt.plot(inp_cca, cca_regr, color=color, linewidth=0.5)
logging.info('cca')
logging.info(cca.x_rotations_)
logging.info(cca.y_rotations_)
def distance(s1, s2, type="dist", x=None):
# TODO: type check
if (type == "dist"):
mat_1 = np.matmul(
np.matmul(s1, np.linalg.solve(np.matmul(np.transpose(s1), s1))),
np.transpose(s1))
mat_2 = np.matmul(
np.matmul(s2, np.linalg.solve(np.matmul(np.transpose(s2), s2))),
np.transpose(s2))
return math.sqrt(np.sum(np.subtract(mat_1, mat_2))**2)
if (type == "trace"):
mat_1 = np.matmul(
np.matmul(s1, np.linalg.solve(np.matmul(np.transpose(s1), s1))),
np.transpose(s1))
mat_2 = np.matmul(
np.matmul(s2, np.linalg.solve(np.matmul(np.transpose(s2), s2))),
np.transpose(s2))
return np.sum(np.diag(np.matmul(mat_1, mat_2))) / helper.n_col(s1)
if (type == "canonical"):
if (x == None):
raise Exception("x must be specified if use type = 'canonical'")
if (helper.n_col(x) != helper.n_row(s1)):
raise Exception("Dimension of x is not correct.")
cca = CCA(n_components=1)
return np.mean(cca.fit(np.matmul(x, s1), np.matmul(x, s2)))
def __init__(self):
# 알고리즘 이름
self._name = 'cca'
# 기본 경로
self._f_path = os.path.abspath(
os.path.join(os.path.dirname(os.path.abspath(__file__)),
os.pardir))
# 경고 메시지 삭제
warnings.filterwarnings('ignore')
# 원본 데이터 로드
data = pd.read_csv(self._f_path +
"/regression/resource/regression_sample.csv",
sep=",",
encoding="utf-8")
# 학습 및 테스트 데이터 분리
self._x = (data["year"] <= 2017)
self._y = (data["year"] >= 2018)
# 학습 데이터 분리
self._x_train, self._y_train = self.preprocessing(data[self._x])
# 테스트 데이터 분리
self._x_test, self._y_test = self.preprocessing(data[self._y])
# 모델 선언
self._model = CCA()
# 모델 학습
self._model.fit(self._x_train, self._y_train)
def train_eval(self, train_index, test_index):
train_source, test_source = self.source[train_index], self.source[
test_index]
train_target, test_target = self.target[train_index], self.target[
test_index]
train_source, test_source = scale_train_test(train_source, test_source)
train_target, _ = scale_train_test(train_target, test_target)
# rho, w_t, w_s, _ = evaluate_cca_wa_wb(train_target, train_source)
cca = CCA(n_components=min(train_source.shape[1],
train_target.shape[1]),
max_iter=1000)
cca.fit(train_source, train_target)
w_s = cca.x_rotations_
w_t = cca.y_rotations_
predicted_target = test_source @ w_s @ np.linalg.pinv(w_t)
predicted_target = unscale_prediction(train_target, predicted_target)
if self.target_encoder is not None:
test_target = self.original_target[test_index]
predicted_target = self.target_encoder.decode(
torch.as_tensor(predicted_target)).detach().numpy()
scores = np.zeros(self.original_target.shape[1])
for i in range(self.original_target.shape[1]):
predicted = predicted_target[:, i]
actual = test_target[:, i]
r, pval = pearsonr(predicted, actual)
scores[i] = r
return scores
def rdc1(x, y, k=10, s=0.2):
if len(x.shape) == 1: x = x.reshape((-1, 1))
if len(y.shape) == 1: y = y.reshape((-1, 1))
cx = np.column_stack([rankdata(xc, method='ordinal')
for xc in x.T]) / float(x.size)
cy = np.column_stack([rankdata(yc, method='ordinal')
for yc in y.T]) / float(y.size)
# Add a vector of ones so that w.x + b is just a dot product
O = np.ones(cx.shape[0])
X = np.column_stack([cx, O])
Y = np.column_stack([cy, O])
Rx = (s / X.shape[1]) * np.random.randn(X.shape[1], k)
Ry = (s / Y.shape[1]) * np.random.randn(Y.shape[1], k)
X = np.dot(X, Rx)
Y = np.dot(Y, Ry)
X = np.sin()
"""print rcancor(np.sin(X),np.sin(Y))
return 0"
"""
cca = CCA(n_components=1)
xc, yc = cca.fit_transform(X, Y)
result = np.corrcoef(xc.T, yc.T)[0, 1]
print(result)
def visualize_with_cca(X, y, title):
cca = CCA(n_components=2)
cca.fit(X, y)
X_cca = cca.transform(X)
Xax = X_cca[:, 0]
Yax = X_cca[:, 1]
labels = (y > 0).astype(int)
cdict = {0: 'red', 1: 'green'}
labl = {0: 'home_loss', 1: 'home_win'}
marker = {0: '*', 1: 'o'}
alpha = {0: .3, 1: .5}
fig, ax = plt.subplots(figsize=(7, 5))
fig.patch.set_facecolor('white')
for l in np.unique(labels):
ix = np.where(labels == l)
ax.scatter(Xax[ix],
Yax[ix],
c=cdict[l],
s=40,
label=labl[l],
marker=marker[l],
alpha=alpha[l])
plt.xlabel("First Principal Component", fontsize=14)
plt.ylabel("Second Principal Component", fontsize=14)
plt.legend()
plt.title(title)
plt.show()
def find_correlation_cca_method1(signal, reference_signals, n_components=2):
r"""
Perform canonical correlation analysis (CCA)
Reference: https://github.com/aaravindravi/Brain-computer-interfaces/blob/master/notebook_12_class_cca.ipynb
Args:
signal : ndarray, shape (channel,time)
Input signal in time domain
reference_signals : ndarray, shape (len(flick_freq),2*num_harmonics,time)
Required sinusoidal reference templates corresponding to the flicker frequency for SSVEP classification
n_components : int, default: 2
number of components to keep (for sklearn.cross_decomposition.CCA)
Returns:
result : array, size: len(flick_freq)
Probability for each reference signals
Dependencies:
CCA : sklearn.cross_decomposition.CCA
np : numpy package
"""
cca = CCA(n_components)
corr = np.zeros(n_components)
result = np.zeros(reference_signals.shape[0])
for freq_idx in range(0, reference_signals.shape[0]):
cca_x = signal.T
cca_y = np.squeeze(reference_signals[freq_idx, :, :]).T
cca.fit(cca_x, cca_y)
a, b = cca.transform(cca_x, cca_y)
for ind_val in range(0, n_components):
corr[ind_val] = np.corrcoef(a[:, ind_val], b[:, ind_val])[0, 1]
result[freq_idx] = np.max(corr)
return result
def CCA_project_vectors(args,
src_dico,
tgt_dico,
src_full,
tgt_full,
src_train,
tgt_train,
NUM_dim=100):
print('Exporting embeddings...')
OutputDir = "output/{}-{}/".format(args.src_lang, args.tgt_lang)
if not os.path.exists(OutputDir):
os.makedirs(OutputDir)
cca = CCA(n_components=NUM_dim)
print("Fitting...")
cca.fit(src_train, tgt_train)
print(cca.get_params())
X_c, Y_c = cca.transform(src_full, tgt_full)
src_out, tgt_out = utils.norm_embeddings(X_c), utils.norm_embeddings(Y_c)
print("Exporting embeddings...")
utils.export_embeddings(src_dico[0], src_out,
OutputDir + 'projected.{}'.format(args.src_lang))
utils.export_embeddings(tgt_dico[0], tgt_out,
OutputDir + 'projected.{}'.format(args.tgt_lang))
print("work over!")
def __init__(self, controller=None, num_targets=8, num_seconds=3):
self.controller = controller
self.num_targets = num_targets
self.num_seconds = num_seconds # number of seconds for a stimulus cycle
self.sampling_rate = 128.0
# frequencies calculated by frames/len(array) as seen in flicker_patterns.txt
# or frequencies are set from paper we are basing our experiment from
# check for either 4 or 8 targets
if self.num_targets == 8:
self.frequencies = np.asarray(
[43.0, 37.0, 29.0, 21.0, 17.0, 11.0, 8.0, 5.0])
elif self.num_targets == 4:
# up, down, right, left from paper
# NOTE: using old csv files with these new frequencies is not valid
self.frequencies = [15.0, 12.0, 8.57, 5.45]
else:
print("cca did not get a good target number.")
# prediction should be targets 1 to num_targets, not 0 to num_targets - 1
# based on command_to_keyboard action
self.prediction = None
self.ref_signals = []
self.getAllReferenceSignals()
self.cca = CCA(n_components=1)
self.fig = None
self.ax = None
self.plotter = None
self.filter_obj = Filter()
def plot_subfigure(X, Y, subplot, title, transform):
if transform == "pca":
X = PCA(n_components=2).fit_transform(X)
elif transform == "cca":
X = CCA(n_components=2).fit(X[0:len(Y),:], Y).transform(X)
else:
raise ValueError
min_x = np.min(X[:, 0])
max_x = np.max(X[:, 0])
min_y = np.min(X[:, 1])
max_y = np.max(X[:, 1])
plt.subplot(1, 2, subplot)
plt.title(title)
zero_class = np.where(Y[:, 0])
one_class = np.where(Y[:, 1])
two_class = np.where(Y[:, 2])
three_class = np.where(Y[:, 3])
plt.scatter(X[:, 0], X[:, 1], s=40, c='gray')
plt.scatter(X[zero_class, 0], X[zero_class, 1], s=160, edgecolors='b',
facecolors='none', linewidths=2, label='Class 1')
def map_spaces(self, algo, src_mapped_embed=None, trg_mapped_embed=None):
# (There may be duplicates in self.shared_vocab_src and/or self.shared_vocab_trg,
# swap_vocab can be used to only inspect one-to-one translations)
src_embed = self.model_src[self.shared_vocab_src]
trg_embed = self.model_trg[self.shared_vocab_trg]
os.makedirs(algo, exist_ok=True)
if algo == "procrustes":
logging.info(
"Calculating Rotation Matrix (Procrustes Problem) and applying it to first embedding"
)
#ortho, _ = orthogonal_procrustes(src_embed, trg_embed)
# does the same as
u, _, vt = np.linalg.svd(trg_embed.T.dot(src_embed))
w = vt.T.dot(u.T)
self.model_src.vectors.dot(w, out=self.model_src.vectors)
elif algo == "noise":
logging.info(
"Calculating Rotation Matrix with noise aware algorithm and applying it to first embedding"
)
transform_matrix, alpha, clean_indices, noisy_indices = noise_aware(
src_embed, trg_embed)
#write cleaned vocab to file
with open("vocab.clean.txt", 'w') as v:
for src, trg in np.asarray(self.shared_vocab)[clean_indices]:
v.write("{}\t{}\n".format(src, trg))
self.model_src.vectors.dot(transform_matrix,
out=self.model_src.vectors)
logging.info("Percentage of clean indices: {}".format(alpha))
elif algo == "cca":
logging.info(
"Calculating Mapping based on CCA and applying it to both embeddings"
)
cca = CCA(n_components=100, max_iter=5000)
cca.fit(src_embed, trg_embed)
self.model_src.vectors, self.model_trg.vectors = cca.transform(
self.model_src.vectors, self.model_trg.vectors)
elif algo == "gcca":
logging.info(
"Calculating Mapping based on GCCA and applying it to both embeddings"
)
gcca = GCCA()
gcca.fit([src_embed, trg_embed])
transform_l = gcca.transform_as_list(
(self.model_src.vectors, self.model_trg.vectors))
# gcca computes positive and negative correlations (eigenvalues), sorted in ascending order.
# We are only interested in the positive portion
self.model_src.vectors = transform_l[0][:, 100:]
self.model_trg.vectors = transform_l[1][:, 100:]
# save transformed model(s)
if src_mapped_embed:
self.model_src.save(os.path.join(algo, src_mapped_embed))
if trg_mapped_embed:
self.model_trg.save(os.path.join(algo, trg_mapped_embed))
def rdc_cca(indexes):
i, j, rdc_features = indexes
cca = CCA(n_components=1, max_iter=CCA_MAX_ITER)
X_cca, Y_cca = cca.fit_transform(rdc_features[i], rdc_features[j])
rdc = np.corrcoef(X_cca.T, Y_cca.T)[0, 1]
# logger.info(i, j, rdc)
return rdc
def new_cca_ds(A, B, n_components=1):
# http://onlinelibrary.wiley.com/doi/10.1002/cem.2637/abstract
model = CCA(n_components=n_components, scale=False).fit(B, A)
F1 = np.linalg.pinv(model.x_scores_).dot(model.y_scores_)
F2 = np.linalg.pinv(model.y_scores_).dot(A)
P = ct.multi_dot((model.x_weights_, F1, F2))
return P, B.dot(P)
def fit_cca(self, outfile=''):
# fits linear CCA constraint and replaces pretrained name embeddings with CCA transformed embeddings
self.load_embeddings()
self.extract_pretrained_prototype_embeddings()
items, vectors = zip(
*[(k, v) for k, v in self.pretrained_prototype_embeddings.items()
if k in self.exemplar_to_concept])
concept_embs = Reach(vectors, items)
train_vectors = []
for x in items:
train_vectors.append(self.train_embeddings[x])
train_vectors = Reach.normalize(train_vectors)
cca = CCA(n_components=self.train_embeddings.size, max_iter=10000)
cca.fit(train_vectors, concept_embs.norm_vectors)
# transform all name embeddings using the CCA mapping
all_name_embeddings = deepcopy(self.pretrained_name_embeddings)
items = [x for _, x in sorted(all_name_embeddings.indices.items())]
projected_name_embeddings = cca.transform(
all_name_embeddings.norm_vectors)
new_name_embeddings = Reach(projected_name_embeddings, items)
self.pretrained_name_embeddings = new_name_embeddings
self.load_embeddings()
if outfile:
with open('{}_cca.p', 'wb') as f:
pickle.dump(cca, f)
def get_cca(X, Y, n_comp=10):
cca = CCA(n_components=n_comp)
print("X.shape", X.shape)
print("Y.shape", Y.shape)
x_scores, y_scores = cca.fit_transform(X, Y)
# Manual Transform
X -= cca.x_mean_
X /= cca.x_std_
Y -= cca.y_mean_
Y /= cca.y_std_
calc_scores_x = np.dot(X, cca.x_rotations_)
calc_scores_y = np.dot(Y, cca.y_rotations_)
# id_x = cca.x_rotations_ @ linalg.pinv2(cca.x_rotations_)
# id_y = cca.y_rotations_ @ linalg.pinv2(cca.y_rotations_)
print("x_scores.shape", x_scores.shape)
print("y_scores.shape", y_scores.shape)
correlations = np.diag(
np.corrcoef(x_scores, y_scores, rowvar=False)[:n_comp, n_comp:])
calc_correlations = np.diag(
np.corrcoef(calc_scores_x, calc_scores_y, rowvar=False)[:n_comp,
n_comp:])
print(correlations)
print(calc_correlations)
return x_scores, y_scores
def main(args):
(training_file, label_file, test_file, u_file, e, c, output_file,
components) = args
X_training = load_feat(training_file)
n = len(X_training)
U = load_feat(u_file)
y_training = [float(line.strip()) for line in open(label_file)]
U = np.asarray(U)
X_training = np.asarray(X_training)
#X = preprocessing.normalize(X, norm='l2')
y_training = np.asarray(y_training)
X_test = load_feat(test_file)
y_test = [float(line.strip()) for line in open(test_label)]
X_test = np.asarray(X_test)
X_test[np.isnan(X_test)] = 0.0
#test_X = preprocessing.normalize(test_X, norm='l2')
y_test = np.asarray(y_test)
s = min(len(X_training), len(U))
cca = CCA(n_components=components, max_iter=50)
(X_cca, U_cca) = cca.fit_transform(X_training[:s], U[:s])
X_test_cca = cca.transform(X_test)
svr = SVR(C=c, epsilon=e, kernel='rbf')
svr.fit(X_cca, y_training[:s])
pred = svr.predict(X_test_cca)
with open(output_file, 'w') as output:
for p in pred:
print >> output, p
return
def cca(m1, m2, preprocessing=None):
"""
Use CCA to decompose two views and plot result.
Params:
m1, m2: Every column is a example with every row as a feature.
preprocessing: If None, we don't do pre-processing; if 'orth', we adjust center to 0 and perform PCA.
"""
# Adjust means to be 0 and perform PCA.
if preprocessing == "orth":
# Zero means.
m1 -= np.mean(m1, axis=1, keepdims=True)
# print("m1=", np.sum(m1, axis=1))
m2 -= np.mean(m2, axis=1, keepdims=True)
# PCA.
cca = CCA(n_components=3, max_iter=100)
cca.fit(m1.T, m2.T)
X_c = cca.transform(m1.T)
fig, ax = plt.subplots()
ax.set_title('Fig.2.(c)')
# ax.set_color_cycle(['blue', 'green', 'red'])
ax.set_prop_cycle('color', ['blue', 'red', 'green'])
ax.plot(X_c)
# ax.plot(Y_c)
plt.show()
def cca_coef(X, Y):
'''
Apply CCA method to compute inter_channel correlation
:param X: data 1 (n_events, n_epochs, n_chans, n_times)
:param Y: data 2 (actually equal to data 1)
'''
cca = CCA(n_components=1)
def fbcca(eeg, list_freqs, fs, num_harms=3, num_fbs=5):
fb_coefs = np.power(np.arange(1,num_fbs+1),(-1.25)) + 0.25
num_targs, _, num_smpls = eeg.shape #40 taget (means 40 fre-phase combination that we want to predict)
y_ref = cca_reference(list_freqs, fs, num_smpls, num_harms)
cca = CCA(n_components=1) #initilize CCA
# result matrix
r = np.zeros((num_fbs,num_targs))
results = np.zeros(num_targs)
for targ_i in range(num_targs):
test_tmp = np.squeeze(eeg[targ_i, :, :]) #deal with one target a time
for fb_i in range(num_fbs): #filter bank number, deal with different filter bank
testdata = filterbank(test_tmp, fs, fb_i) #data after filtering
for class_i in range(num_targs):
refdata = np.squeeze(y_ref[class_i, :, :]) #pick corresponding freq target reference signal
test_C, ref_C = cca.fit_transform(testdata.T, refdata.T)
# len(row) = len(observation), len(column) = variables of each observation
# number of rows should be the same, so need transpose here
# output is the highest correlation linear combination of two sets
r_tmp, _ = pearsonr(np.squeeze(test_C), np.squeeze(ref_C)) #return r and p_value, use np.squeeze to adapt the API
r[fb_i, class_i] = r_tmp
rho = np.dot(fb_coefs, r) #weighted sum of r from all different filter banks' result
tau = np.argmax(rho) #get maximum from the target as the final predict (get the index)
results[targ_i] = tau #index indicate the maximum(most possible) target
return results
