def RunSparseCodingScikit(q):
totalTimer = Timer()
# Load input dataset.
inputData = np.genfromtxt(self.dataset[0], delimiter=',')
dictionary = np.genfromtxt(self.dataset[1], delimiter=',')
# Get all the parameters.
l = re.search("-l (\d+)", options)
l = 0 if not l else int(l.group(1))
try:
with totalTimer:
# Perform Sparse Coding.
model = SparseCoder(dictionary=dictionary,
transform_algorithm='lars',
transform_alpha=l)
code = model.transform(inputData)
except Exception as e:
q.put(-1)
return -1
time = totalTimer.ElapsedTime()
q.put(time)
return time
def sparse_code(self, X, W):
'''
Given data matrix X and dictionary matrix W, find
code matrix H such that W*H approximates X
args:
X (numpy array): data matrix with dimensions: data_dim (d) x samples (n)
W (numpy array): dictionary matrix with dimensions: data_dim (d) x topics (r)
returns:
H (numpy array): code matrix with dimensions: topics (r) x samples(n)
'''
if DEBUG:
print('sparse_code')
print('X.shape:', X.shape)
print('W.shape:', W.shape, '\n')
# extract matrix dimensions from X, W
# and initialize H with appropriate dimensions
d, n = np.shape(X)
d, r = np.shape(W)
H = np.zeros([n, r])
# initialize the SparseCoder with W as its dictionary
# then find H such that X \approx W*H
coder = SparseCoder(dictionary=W.T,
transform_n_nonzero_coefs=None,
transform_alpha=2,
transform_algorithm='lasso_lars',
positive_code=False)
H = coder.transform(X.T)
# transpose H before returning to undo the preceding transpose on X
return H.T
def main():
images, labels = load_labeled_training(flatten=True)
images = standardize(images)
unl = load_unlabeled_training(flatten=True)
unl = standardize(unl)
test = load_public_test(flatten=True)
test = standardize(test)
shuffle_in_unison(images, labels)
#d = DictionaryLearning().fit(images)
d = MiniBatchDictionaryLearning(n_components=500, n_iter=500, verbose=True).fit(images)
s = SparseCoder(d.components_)
proj_test = s.transform(images)
pt = s.transform(test)
#kpca = KernelPCA(kernel="rbf")
#kpca.fit(unl)
#test_proj = kpca.transform(images)
#pt = kpca.transform(test)
#spca = SparsePCA().fit(unl)
#test_proj = spca.transform(images)
#pt = spca.transform(test)
svc = SVC()
scores = cross_validation.cross_val_score(svc, proj_test, labels, cv=10)
print scores
print np.mean(scores)
print np.var(scores)
svc.fit(proj_test, labels)
pred = svc.predict(pt)
write_results(pred, '../svm_res.csv')
def RunSparseCodingScikit():
totalTimer = Timer()
# Load input dataset.
inputData = np.genfromtxt(self.dataset[0], delimiter=',')
dictionary = np.genfromtxt(self.dataset[1], delimiter=',')
# Get all the parameters.
opts = {}
if "lambda" in options:
opts["transform_alpha"] = options.pop("lambda")
if "max_iterations" in options:
opts["max_iter"] = options.pop("max_iterations")
opts["transform_algorithm"] = "lars"
opts["dictionary"] = dictionary
if len(options) > 0:
Log.Fatal("Unknown parameters: " + str(options))
raise Exception("unknown parameters")
try:
with totalTimer:
# Perform Sparse Coding.
model = SparseCoder(**opts)
code = model.transform(inputData)
except Exception as e:
return -1
return totalTimer.ElapsedTime()
def get_activations(stft, dico, n_nonzero_coefs=None):
coder = SparseCoder(
dictionary=dico.T,
transform_n_nonzero_coefs=n_nonzero_coefs,
transform_algorithm="lasso_cd",
positive_code=True)
return coder.transform(stft.T).T
def omp_sparse(dictionary, train_data):
dictionary = dictionary.transpose()
w = SparseCoder(dictionary, transform_algorithm='omp')
t = w.transform(train_data)
return t
def RunSparseCodingScikit(q):
totalTimer = Timer()
# Load input dataset.
inputData = np.genfromtxt(self.dataset[0], delimiter=',')
dictionary = np.genfromtxt(self.dataset[1], delimiter=',')
# Get all the parameters.
opts = {}
if "lambda" in options:
opts["transform_alpha"] = options.pop("lambda")
if "max_iterations" in options:
opts["max_iter"] = options.pop("max_iterations")
opts["transform_algorithm"] = "lars"
opts["dictionary"] = dictionary
if len(options) > 0:
Log.Fatal("Unknown parameters: " + str(options))
raise Exception("unknown parameters")
try:
with totalTimer:
# Perform Sparse Coding.
model = SparseCoder(**opts)
code = model.transform(inputData)
except Exception as e:
q.put(-1)
return -1
time = totalTimer.ElapsedTime()
q.put(time)
return time
def predict(self, X_test):
n_samples = X_test.shape[1]
X = X_test
D = self.D
W = self.W
D = np.nan_to_num(D)
D1 = np.sum(np.abs(D)**2, axis=0)**(1. / 2)
for i in range(D.shape[1]):
D[:, i] = D[:, i] / D1[i]
#print(D,X)
print("predicting")
coder = SparseCoder(dictionary=np.transpose(D),
transform_n_nonzero_coefs=self.n_nonzero,
transform_algorithm='omp')
Z = coder.transform(np.transpose(X))
#print(Z.shape, Z)
#print(np.count_nonzero(Z))
#print(np.count_nonzero(Z, axis=0))
pred = np.zeros((n_samples, ))
for i in range(n_samples):
pred[i] = np.argmax(np.dot(W, Z[i, :]))
return pred
def joint_sparse_code_tensor(self, X, W):
'''
Given data matrix X and dictionary matrix W, find
code matrix H such that W*H approximates X
args:
X (numpy array): data matrix with dimensions: features (d) x samples (n)
W (numpy array): dictionary matrix with dimensions: features (d) x topics (r)
returns:
H (numpy array): code matrix with dimensions: topics (r) x samples(n)
'''
if DEBUG:
print('sparse_code')
print('X.shape:', X.shape)
print('W.shape:', W.shape, '\n')
# initialize the SparseCoder with W as its dictionary
# then find H such that X \approx W*H
if self.alpha == None:
coder = SparseCoder(dictionary=W.T, transform_n_nonzero_coefs=None,
transform_alpha=2, transform_algorithm='lasso_lars', positive_code=True)
else:
coder = SparseCoder(dictionary=W.T, transform_n_nonzero_coefs=None,
transform_alpha=self.alpha, transform_algorithm='lasso_lars', positive_code=True)
# alpha = L1 regularization parameter.
H = coder.transform(X.T)
# transpose H before returning to undo the preceding transpose on X
return H
def sparse_coding_with_LSC(Y_labelled,D,H,W,gamma,lamda):
_Y = np.vstack((Y_labelled, np.sqrt(gamma) * H))
_D = np.vstack((D, np.sqrt(gamma) * W))
coder = SparseCoder(dictionary=_D.T,transform_alpha=lamda/2., transform_algorithm='lasso_cd')
X =(coder.transform(_Y.T)).T
return X
def test_sparse_coder():
db = coco.COCO('/media/zawlin/ssd/coco/annotations/captions_train2014.json')
ilsvrc_word2vec = zl.load('ilsvrc_word2vec')
test_word2vec = zl.load('test_word2vec')
D = []
idx2word = []
for i in xrange(len(ilsvrc_word2vec.keys())):
idx2word.append(ilsvrc_word2vec.keys()[i])
D.append(ilsvrc_word2vec[ilsvrc_word2vec.keys()[i]])
idx2word = np.array(idx2word)
D = np.array(D)
print 'loading word2vec'
model = gensim.models.Word2Vec.load_word2vec_format('./GoogleNews-vectors-negative300.bin', binary=True)
coder = SparseCoder(dictionary=D, transform_n_nonzero_coefs=5,
transform_alpha=None, transform_algorithm='lasso_cd')
#random.shuffle(db.anns)
for k in db.anns:
cap = db.anns[k]['caption']
splited = cap.split(' ')
for s in splited:
if s.lower() in model:
y = model[s.lower()]
x = coder.transform(y.reshape(1,-1))[0]
print '%s = %s'%(s.lower(),idx2word[np.argsort(x)[::-1]])
print x[np.argsort(x)[::-1]]
c = raw_input('press q to quit')
if c== 'q':
exit(0)
def compute_residual(query, test):
"""Compute the residual of the sparse coding representation (RSCR)
Args:
query (torch.Tensor): Query LR
test (torch.Tensor): Test LR
Returns:
float: RSCR
"""
D = test.squeeze().detach().numpy()
Y = query.detach().numpy()
X = query
if Y.ndim < 2:
Y = Y.reshape(1, -1)
if D.ndim < 2:
D = D.reshape(1, -1)
coder = SparseCoder(dictionary=D, transform_algorithm='lasso_lars')
A = coder.transform(Y)
A = torch.FloatTensor(A)
D = torch.FloatTensor(D)
RSCR = torch.norm(X - torch.mm(A, D)).item()
return RSCR
def SparseRepresentation(X,D,A_mean,param,l1,transform_n_nonzero_coefs):
if len(D.shape)>2:
featureDim=D.shape[1]
D=D.transpose(1,0,2).reshape((featureDim,-1))
# par = Params()
# par.lambda1 = param.lambda1;
# par.lambda2 = param.lambda2;
# D=np.hstack((D[0],D[1],D[2],D[3],D[4]))
# A_ini = np.ones((transform_n_nonzero_coefs, X.shape[1]))
# par.A_mean = A_mean[0]
# if D.shape[0] >= D.shape[1]:
# w, v = LA.eig(D.T @ D)
# par.c = 1.05 * w.max()
# else:
# w, v = LA.eig(D @ D.T)
# par.c = 1.05 * w.max()
# opts = Coef_Update_Test(X, D, A_ini, par)
# A_test = opts.A
# P = LA.inv(D.T @ D + l1 * np.eye(D.shape[1])) @ D.T
# A_test = P @ X
coder = SparseCoder(dictionary=D.T, transform_n_nonzero_coefs=transform_n_nonzero_coefs,
transform_algorithm="omp")
A_test = (coder.transform(X.T)).T
A_test_nonzero=A_test
# A_test_nonzero=np.empty((transform_n_nonzero_coefs, X.shape[1]))
# for i in range(A_test.shape[1]):
# A_test_nonzero[:,i] = A_test[:,i][A_test[:,i] != 0]
return A_test_nonzero
def reconstruct_image_color(self,
loading,
recons_resolution=1,
if_save=True):
print('reconstructing given network...')
'''
Reconstruct original color image using lerned CP dictionary atoms
'''
A = self.data # A.shape = (row, col, 3)
CPdict = self.out(loading)
k = self.patch_size
W = np.zeros(shape=(3 * k**2, self.n_components))
for j in np.arange(self.n_components):
W[:, j] = CPdict.get('A' + str(j)).reshape(-1, 1)[:, 0]
A_matrix = A.reshape(-1, A.shape[1]) # (row, col, 3) --> (3row, col)
[m, n] = A_matrix.shape
A_recons = np.zeros(shape=A.shape)
A_overlap_count = np.zeros(shape=(A.shape[0], A.shape[1]))
k = self.patch_size
t0 = time()
c = 0
num_rows = np.floor(
(A_recons.shape[0] - k) / recons_resolution).astype(int)
num_cols = np.floor(
(A_recons.shape[1] - k) / recons_resolution).astype(int)
for i in np.arange(0, A_recons.shape[0] - k, recons_resolution):
for j in np.arange(0, A_recons.shape[1] - k, recons_resolution):
patch = A[i:i + k, j:j + k, :]
patch = patch.reshape((-1, 1))
coder = SparseCoder(dictionary=W.T,
transform_n_nonzero_coefs=None,
transform_alpha=1,
transform_algorithm='lasso_lars',
positive_code=True)
# alpha = L1 regularization parameter. alpha=2 makes all codes zero (why?)
code = coder.transform(patch.T)
patch_recons = np.dot(W, code.T).T
patch_recons = patch_recons.reshape(k, k, 3)
# now paint the reconstruction canvas
for x in itertools.product(np.arange(k), repeat=2):
c = A_overlap_count[i + x[0], j + x[1]]
A_recons[i + x[0], j +
x[1], :] = (c * A_recons[i + x[0], j + x[1], :] +
patch_recons[x[0], x[1], :]) / (c + 1)
A_overlap_count[i + x[0], j + x[1]] += 1
# progress status
print('reconstructing (%i, %i)th patch out of (%i, %i)' %
(i / recons_resolution, j / recons_resolution, num_rows,
num_cols))
print('Reconstructed in %.2f seconds' % (time() - t0))
print('A_recons.shape', A_recons.shape)
if if_save:
np.save('Video_dictionary/video_recons_color', A_recons)
plt.imshow(A_recons)
return A_recons
def parallel_sc(args):
(dico, p) = args
coder = SparseCoder(dictionary=dico, transform_algorithm='omp')
#by default, number of non zero coefficients is 0.1 * n_features
#Zeyde et al.: 3 non zero coefficients !
code = coder.transform(p).astype(np.float32)
return code
def code1(self, data, max_iter=None, errors=False):
'''
Sparse codes a single feature
Requires that the dictionary is already trained
'''
if self.codebook is None:
self.codebook = SparseCoder(self.codebook_comps.T,n_jobs=4)
return self.codebook.transform(data.reshape(1,-1)).ravel()
def test_sparse_coder_dtype_match(data_type, transform_algorithm):
# Verify preserving dtype for transform in sparse coder
n_components = 6
rng = np.random.RandomState(0)
dictionary = rng.randn(n_components, n_features)
coder = SparseCoder(dictionary.astype(data_type),
transform_algorithm=transform_algorithm)
code = coder.transform(X.astype(data_type))
assert code.dtype == data_type
def sparse_code_proximal(self, X, W, a1, a2):
'''
Given data matrix X and dictionary matrix W, find
code matrix H and noise matrix S such that
H, S = argmin ||X - WH - S||_{F}^2 + \alpha ||H||_{1} + \beta ||S||_{1}
Uses proximal gradient
G = [H \\ S']
V = [W, b I] (so that VG = WH + bS')
Then solve
min_{G,V} |X-VG|_{F} + \alpha |G|_{1} =
= min_{H,S'} |X - HW - bS'|_{F}^2 + \alpha |H|_{1} + \alpha |S'|_{1}
= min_{H,S} |X - HW - S|_{F}^2 + \alpha |H|_{1} + (\alpha/b)|S|_{1}
using constrained LASSO
args:
X (numpy array): data matrix with dimensions: features (d) x samples (n)
W (numpy array): dictionary matrix with dimensions: features (d) x topics (r)
returns:
H (numpy array): code matrix with dimensions: topics (r) x samples(n)
S (numpy array): noise matrix with dimensions: features (d) x samples (n)
'''
if DEBUG:
print('sparse_code')
print('X.shape:', X.shape)
print('W.shape:', W.shape, '\n')
# initialize the SparseCoder with W as its dictionary
# H_new = LASSO with W as dictionary
# S_new = LASSO with id (d x d) as dictionary
# Y_new = Y + (W H_new + S_new - S) : Dual variable
### Initialization
d, n = X.shape
r = self.n_components
### Augmented dictionary matrix for proximal gradient
V = np.hstack((W, a2*np.identity(d)))
### Proximal sparse coding by constrained LASSO
coder = SparseCoder(dictionary=V.T, transform_n_nonzero_coefs=None,
transform_alpha=a1, transform_algorithm='lasso_lars', positive_code=True)
G = coder.transform(X.T)
G = G.T
# transpose G before returning to undo the preceding transpose on X
### Read off H and S from V
H = G[0:r, :]
S = a2*G[r:, :]
return H, S
def sc_result_analysis():
"""
对稀疏编码的结果进行分析i
:return:
"""
sc_file = open('./tmp_file/30_dictionary.pickle', 'rb')
sc_list = cPickle.load(sc_file)
classified_file = open('./tmp_file/30_class_result.pickle', 'rb')
(classified_feature, classified_patch) = cPickle.load(classified_file)
model_file = open('./tmp_file/30_kmeans_pca_model.pickle', 'rb')
(k_means, pca) = cPickle.load(model_file)
sc_file.close()
classified_file.close()
model_file.close()
# ========================================================================
for i in range(5):
k = i
#v_feature = pca.transform(classified_feature[1][k]).reshape((-1,))
v_feature = classified_feature[3][k]
v_patch = classified_patch[3][k]
feature_dict = sc_list[0][:, :144]
patch_dict = sc_list[0][:, 144:]
#v_feature = feature_dict[0]
#v_patch = patch_dict[0]
coder = SparseCoder(dictionary=feature_dict, transform_algorithm='omp',
transform_alpha=0.01, n_jobs=2, transform_n_nonzero_coefs=1)
weight = coder.transform(v_feature)
v_patch = v_patch.reshape((9, 9))
result = np.dot(weight, patch_dict).reshape((9, 9))
mask = weight != 0
print weight[mask]
mask = mask[0]
print len(patch_dict[mask])
print len(patch_dict[mask])
patch_show(patch_dict[mask],[0,0,0.45,0.45],1)
ax2 = plt.axes([0, 0.5, 0.45, 0.45])
ax2.imshow(result, interpolation="none", cmap=cm.gray)
ax2 = plt.axes([0.5, 0.5, 0.45, 0.45])
ax2.imshow(v_patch, interpolation="none", cmap=cm.gray)
plt.show()
def train(DWA_all, D_all, W_all, A_all, Cs, labels, file_paths,
inds_of_file_path, train_number, start_init_number, update_times,
update_index, n_classes, n_atoms, n_features, lambda_init,
the_lambda, transform_n_nonzero_coefs, omp_tag):
for j in range(n_classes):
if j == 0:
print(update_index)
sys.stdout.flush()
coder = SparseCoder(
dictionary=D_all.T,
transform_n_nonzero_coefs=transform_n_nonzero_coefs,
transform_algorithm='omp')
label_indexs_for_update = inds_of_file_path[j][:train_number]
new_index = [
label_indexs_for_update[(update_index + start_init_number) %
train_number]
]
new_label = labels[new_index][0]
lab_index = j
im_vec = load_img(file_paths[new_index][0])
im_vec = im_vec / 255.
new_y = np.array(im_vec, dtype=float)
new_y = preprocessing.normalize(new_y.T, norm='l2').T
new_y = norm_Ys(new_y)
new_y = new_y.reshape(n_features, 1)
new_h = np.zeros((n_classes, 1))
new_h[lab_index, 0] = 1
new_q = np.zeros((n_atoms * n_classes, 1))
new_q[n_atoms * lab_index:n_atoms * (lab_index + 1), 0] = 1
new_yhq = np.vstack((new_y, new_h, new_q))
new_x = None
if omp_tag == "true":
new_x = (coder.transform(new_y.T)).T
if omp_tag == "wzz":
new_x = transform(D_all, new_y, transform_n_nonzero_coefs)
the_C = Cs
the_u = (1 / the_lambda) * np.dot(the_C, new_x)
gamma = 1 / (1 + np.dot(new_x.T, the_u))
the_r = new_yhq - np.dot(DWA_all, new_x)
new_C = (1 / the_lambda) * the_C - gamma * np.dot(the_u, the_u.T)
new_DWA = DWA_all + gamma * np.dot(the_r, the_u.T)
DWA_all = new_DWA
part_lambda = (1 - update_index / update_times)
the_lambda = 1 - (1 -
lambda_init) * part_lambda * part_lambda * part_lambda
D_all = DWA_all[0:D_all.shape[0], :]
W_all = DWA_all[D_all.shape[0]:D_all.shape[0] + W_all.shape[0], :]
A_all = DWA_all[D_all.shape[0] + W_all.shape[0]:, :]
D_all = preprocessing.normalize(D_all.T, norm='l2').T
W_all = preprocessing.normalize(W_all.T, norm='l2').T
A_all = preprocessing.normalize(A_all.T, norm='l2').T
DWA_all = np.vstack((D_all, W_all, A_all))
return DWA_all, D_all, W_all, A_all, the_lambda
def predict_svm(self, img): gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) gray = cv2.resize(gray, (100,100)) gray = np.array(gray).reshape(1,100*100) #X = self.pca.transform(gray) D = np.array(self.pca.components_) dictionary = SparseCoder(D, transform_algorithm='omp', transform_n_nonzero_coefs=10, transform_alpha=None) features = dictionary.transform(gray) label = self.svm.predict(features) score = None return label, score
def sparse_codifier(y,D, transform_algo = 'omp', transform_n_nonzero_coefs = 2):
"""
Encodes an input vector y into an output sparse vector x,
based on a given dictionary D
Retired, to delete.
"""
print('y.shape = %s ' % (y.shape,))
print('Dictionary shape = %s' % (D.shape,))
coder = SparseCoder(dictionary = D, transform_algorithm = transform_algo)
x = coder.transform(y)
print('x.shape = %s' % (x.shape,))
return x
def sparse_codifier(y, D, transform_algo='omp', transform_n_nonzero_coefs=2):
"""
Encodes an input vector y into an output sparse vector x,
based on a given dictionary D
Retired, to delete.
"""
print('y.shape = %s ' % (y.shape, ))
print('Dictionary shape = %s' % (D.shape, ))
coder = SparseCoder(dictionary=D, transform_algorithm=transform_algo)
x = coder.transform(y)
print('x.shape = %s' % (x.shape, ))
return x
def reconstruct_image_color(self, path, recons_resolution=1):
print('reconstructing given network...')
'''
Note: For WAN data, the algorithm reconstructs the normalized WAN matrix A/np.max(A).
Scale the reconstructed matrix B by np.max(A) and compare with the original network.
'''
A = self.read_img_as_array(path) # A.shape = (row, col, 3)
A_matrix = A.reshape(-1, A.shape[1]) # (row, col, 3) --> (3row, col)
[m, n] = A_matrix.shape
A_recons = np.zeros(shape=A.shape)
A_overlap_count = np.zeros(shape=(A.shape[0], A.shape[1]))
k = self.patch_size
t0 = time()
c = 0
num_rows = np.floor(
(A_recons.shape[0] - k) / recons_resolution).astype(int)
num_cols = np.floor(
(A_recons.shape[1] - k) / recons_resolution).astype(int)
for i in np.arange(0, A_recons.shape[0] - k, recons_resolution):
for j in np.arange(0, A_recons.shape[1] - k, recons_resolution):
patch = A[i:i + k, j:j + k, :]
patch = patch.reshape((-1, 1))
# print('patch.shape', patch.shape)
coder = SparseCoder(dictionary=self.W.T,
transform_n_nonzero_coefs=None,
transform_alpha=1,
transform_algorithm='lasso_lars',
positive_code=True)
# alpha = L1 regularization parameter. alpha=2 makes all codes zero (why?)
code = coder.transform(patch.T)
patch_recons = np.dot(self.W, code.T).T
patch_recons = patch_recons.reshape(k, k, 3)
# now paint the reconstruction canvas
for x in itertools.product(np.arange(k), repeat=2):
c = A_overlap_count[i + x[0], j + x[1]]
A_recons[i + x[0], j +
x[1], :] = (c * A_recons[i + x[0], j + x[1], :] +
patch_recons[x[0], x[1], :]) / (c + 1)
A_overlap_count[i + x[0], j + x[1]] += 1
# progress status
print('reconstructing (%i, %i)th patch out of (%i, %i)' %
(i / recons_resolution, j / recons_resolution, num_rows,
num_cols))
print('Reconstructed in %.2f seconds' % (time() - t0))
print('A_recons.shape', A_recons.shape)
np.save('Image_dictionary/img_recons_color', A_recons)
plt.imshow(A_recons)
return A_recons
def __sparse_encode(self, D, test_X):
"""
Z.shape = (test_size, atoms_size=dict_size)
"""
coder = SparseCoder(dictionary=D,
transform_algorithm='lasso_cd',
transform_alpha=self.coder_alpha)
coder.fit(test_X)
Z = coder.transform(test_X)
return Z
def reconstruct_network(self, path, recons_iter=100):
print('reconstructing given network...')
'''
Note: For WAN data, the algorithm reconstructs the normalized WAN matrix A/np.max(A).
Scale the reconstructed matrix B by np.max(A) and compare with the original network.
'''
A = self.A
[N, N] = A.shape
A_recons = np.zeros(shape=(N, N))
A_overlap_count = np.zeros(shape=(N, N))
B = self.path_adj(self.k1, self.k2)
k = self.k1 + self.k2 + 1 # size of the network patch
x0 = np.random.choice(np.arange(0, N))
emb = self.tree_sample(B, x0)
t0 = time()
c = 0
for t in np.arange(recons_iter):
patch, emb = self.get_single_patch_glauber(B, emb)
coder = SparseCoder(dictionary=self.W.T,
transform_n_nonzero_coefs=None,
transform_alpha=0,
transform_algorithm='lasso_lars',
positive_code=True)
# alpha = L1 regularization parameter. alpha=2 makes all codes zero (why?)
# This only occurs when sparse coding a single array
code = coder.transform(patch.T)
patch_recons = np.dot(self.W, code.T).T
patch_recons = patch_recons.reshape(k, k)
for x in itertools.product(np.arange(k), repeat=2):
a = emb[x[0]]
b = emb[x[1]]
j = A_overlap_count[a, b]
A_recons[a, b] = (j * A_recons[a, b] +
patch_recons[x[0], x[1]]) / (j + 1)
# A_recons[a,b] = A_recons[a,b] + patch_recons[x[0], x[1]]
A_overlap_count[a, b] += 1
# progress status
if 100 * t / recons_iter % 1 == 0:
print(t / recons_iter * 100)
print('Reconstructed in %.2f seconds' % (time() - t0))
np.save(
'Network_dictionary/WAN/twain_recons' + "_" + str(self.k1) +
str(self.k2) + "_" + str(self.n_components), A_recons)
return A_recons
def F(self, x, B, A=None):
'''
Calculates a gradient-derived matrix (A-prime) based on a known dictionary and a known
signal matrix
Parameters
-----------
x : numpy array: base data used for reconstruction
B : numpy array: Dictionary
A : numpy array: Singal matrix
Return
------
acts: numpy array: derived matrix (A-prime)
'''
# 4b
B = np.asarray(B)
A = np.asarray(A)
coder = SparseCoder(dictionary=B.T,
transform_alpha=self.rp,
transform_algorithm='lasso_cd')
comps, acts = librosa.decompose.decompose(x, transformer=coder)
acts = self._pos_constraint(acts)
return acts
def generateWalk2vecSC(D, times, n, vs, l):
coder = SparseCoder(dictionary=D.components_, transform_alpha=l)
path = input("Enter filename to save datasets...\n")
with open(path, 'w') as f:
try:
for i in range(times):
v = coder.transform(vs[i * n:(i + 1) * n])
v = np.average(v, axis=0)
f.write(str(v.tolist()) + '\n')
sys.stdout.write("\r{}/{} finished".format(i + 1, times))
except:
os.remove(path)
traceback.print_exc()
exit(1)
print("\rData successfully generated")
def test_sparse_coder_estimator_clone():
n_components = 12
rng = np.random.RandomState(0)
V = rng.randn(n_components, n_features) # random init
V /= np.sum(V**2, axis=1)[:, np.newaxis]
coder = SparseCoder(dictionary=V,
transform_algorithm='lasso_lars',
transform_alpha=0.001)
cloned = clone(coder)
assert id(cloned) != id(coder)
np.testing.assert_allclose(cloned.dictionary, coder.dictionary)
assert id(cloned.dictionary) != id(coder.dictionary)
assert cloned.n_components_ == coder.n_components_
assert cloned.n_features_in_ == coder.n_features_in_
data = np.random.rand(n_samples, n_features).astype(np.float32)
np.testing.assert_allclose(cloned.transform(data), coder.transform(data))
def predict_joint_single(self, data, a1):
k = self.patch_size
L = self.prediction_length
A = data # A.shape = (self.data.shape[0], k-L, self.data.shape[2])
# A_recons = np.zeros(shape=(A.shape[0], k, A.shape[2]))
# W_tensor = self.W.reshape((k, A.shape[0], -1))
# print('A.shape', A.shape)
W_tensor = self.W.reshape(
(self.data.shape[0], k, self.data.shape[2], -1))
# print('W.shape', W_tensor.shape)
# for missing data, not needed for the COVID-19 data set
# extract only rows of nonnegative values (disregarding missing entries) (negative = N/A)
J = np.where(np.min(A, axis=(0, 1)) >= -1)
A_pos = A[:, :, J]
# print('A_pos', A_pos)
# print('np.min(A)', np.min(A))
W_tensor = W_tensor[:, :, J, :]
W_trimmed = W_tensor[:, 0:k - L, :, :]
W_trimmed = W_trimmed.reshape((-1, self.n_components))
patch = A_pos
# print('patch', patch)
patch = patch.reshape((-1, 1))
# print('patch.shape', patch.shape)
# print('patch', patch)
coder = SparseCoder(dictionary=W_trimmed.T,
transform_n_nonzero_coefs=None,
transform_alpha=a1,
transform_algorithm='lasso_lars',
positive_code=True)
# alpha = L1 regularization parameter
code = coder.transform(patch.T)
patch_recons = np.dot(
self.W,
code.T).T # This gives prediction on the last L missing entries
patch_recons = patch_recons.reshape(-1, k, A.shape[2])
# now paint the reconstruction canvas
# only add the last predicted value
A_recons = patch_recons[:, k - 1, :]
return A_recons[:, np.newaxis, :]
def sparse_code_affine(self, X, W, a1, num_blocks):
'''
Given data matrix X and dictionary matrix W, find
code matrix H and affine translations for each blocks in X so that X \approx WH + block-translation.
For the case when X has a single block, this is X \approx WH + bI.
Use alternating optimization -- fix b, find H by LASSO; fix H, find b by MSE, which gives b = mean(X-WH).
args:
X (numpy array): data matrix with dimensions: features (d) x samples (n)
W (numpy array): dictionary matrix with dimensions: features (d) x topics (r)
returns:
H (numpy array): code matrix with dimensions: topics (r) x samples(n)
S (numpy array): noise matrix with dimensions: features (d) x samples (n). (d) rows are partitioned into
"num_blocks" blocks, in which the entries of S are constant.
'''
if DEBUG:
print('sparse_code')
print('X.shape:', X.shape)
print('W.shape:', W.shape, '\n')
# initialize the SparseCoder with W as its dictionary
# H_new = LASSO with W as dictionary
# S_new = LASSO with id (d x d) as dictionary
# Y_new = Y + (W H_new + S_new - S) : Dual variable
block_iter = 1
nb = num_blocks
H = []
S = np.zeros(shape=X.shape)
for step in np.arange(block_iter):
### Optimize H by constrained LASSO
coder = SparseCoder(dictionary=W.T, transform_n_nonzero_coefs=None,
transform_alpha=a1, transform_algorithm='lasso_lars', positive_code=True)
H = coder.transform((X - S).T)
H = H.T
### Optimiez S by solving MSE
l = np.floor(X.shape[0]/nb).astype(int)
for i in np.arange(l):
Y = X - W @ H
print('Y', np.sum(Y))
print('S', np.sum(S))
S[nb*i : nb*(i+1), 0] = np.mean(Y[nb*i : nb*(i+1), 0]) # solution to block-MSE
S[nb * l:, 0] = np.mean((X - W @ H)[nb * l:, 0]) # solution to block-MSE
return H, S
def test_sparse_coder_parallel_mmap():
# Non-regression test for:
# https://github.com/scikit-learn/scikit-learn/issues/5956
# Test that SparseCoder does not error by passing reading only
# arrays to child processes
rng = np.random.RandomState(777)
n_components, n_features = 40, 64
init_dict = rng.rand(n_components, n_features)
# Ensure that `data` is >2M. Joblib memory maps arrays
# if they are larger than 1MB. The 4 accounts for float32
# data type
n_samples = int(2e6) // (4 * n_features)
data = np.random.rand(n_samples, n_features).astype(np.float32)
sc = SparseCoder(init_dict, transform_algorithm='omp', n_jobs=2)
sc.fit_transform(data)
def cluster_sk_sparse_coder(content):
""" x """
_config = SparseCoder(n_components=content['n_components'],
alpha=content['alpha'],
ridge_alpha=content['ridge_alpha'],
max_iter=content['max_iter'],
tol=content['tol'],
method=content['sk_method'],
n_jobs=-1)
_result = _config.fit_transform(content['data'])
return httpWrapper(
json.dumps({
'result': _result.tolist(),
'components': _config.components_.tolist(),
'error': _config.error_,
'iter': _config.n_iter_
}))
def test_sparse_coder_parallel_mmap():
# Non-regression test for:
# https://github.com/scikit-learn/scikit-learn/issues/5956
# Test that SparseCoder does not error by passing reading only
# arrays to child processes
rng = np.random.RandomState(777)
n_components, n_features = 40, 64
init_dict = rng.rand(n_components, n_features)
# Ensure that `data` is >2M. Joblib memory maps arrays
# if they are larger than 1MB. The 4 accounts for float32
# data type
n_samples = int(2e6) // (4 * n_features)
data = np.random.rand(n_samples, n_features).astype(np.float32)
sc = SparseCoder(init_dict, transform_algorithm='omp', n_jobs=2)
sc.fit_transform(data)
def test_max_iter():
def ricker_function(resolution, center, width):
"""Discrete sub-sampled Ricker (Mexican hat) wavelet"""
x = np.linspace(0, resolution - 1, resolution)
x = (
(2 / (np.sqrt(3 * width) * np.pi**0.25))
* (1 - (x - center) ** 2 / width**2)
* np.exp(-((x - center) ** 2) / (2 * width**2))
)
return x
def ricker_matrix(width, resolution, n_components):
"""Dictionary of Ricker (Mexican hat) wavelets"""
centers = np.linspace(0, resolution - 1, n_components)
D = np.empty((n_components, resolution))
for i, center in enumerate(centers):
D[i] = ricker_function(resolution, center, width)
D /= np.sqrt(np.sum(D**2, axis=1))[:, np.newaxis]
return D
transform_algorithm = "lasso_cd"
resolution = 1024
subsampling = 3 # subsampling factor
n_components = resolution // subsampling
# Compute a wavelet dictionary
D_multi = np.r_[
tuple(
ricker_matrix(
width=w, resolution=resolution, n_components=n_components // 5
)
for w in (10, 50, 100, 500, 1000)
)
]
X = np.linspace(0, resolution - 1, resolution)
first_quarter = X < resolution / 4
X[first_quarter] = 3.0
X[np.logical_not(first_quarter)] = -1.0
X = X.reshape(1, -1)
# check that the underlying model fails to converge
with pytest.warns(ConvergenceWarning):
model = SparseCoder(
D_multi, transform_algorithm=transform_algorithm, transform_max_iter=1
)
model.fit_transform(X)
# check that the underlying model converges w/o warnings
with pytest.warns(None) as record:
model = SparseCoder(
D_multi, transform_algorithm=transform_algorithm, transform_max_iter=2000
)
model.fit_transform(X)
assert not [w.message for w in record]
def genD_X(p):
Site = p[0]
Y = p[1][0]
D = p[1][1]
n_nonzero = 1
algo = 'omp'
#Y_mod=np.reshape(Y, (1,N))
Y_mod = Y.T
D_mod = D.T
#D = np.random.randn(K,N)
coder = SparseCoder(dictionary=D_mod, transform_n_nonzero_coefs=n_nonzero, transform_alpha=None, transform_algorithm=algo)
X = coder.transform(Y_mod)
X_mod = X.T
#X-returned has shape (k,s)
#shape of Y is (n,s) while D is (n,k)
return (Site, (Y,D,X_mod))
def slp_train_svm(folder, output):
path = str.format('{0}/*', folder)
data = []
cls = []
c = 0
for folder in glob.glob(path):
for img_path in glob.glob(folder + '/*.jpg'):
print 'Loading: ', img_path
img = cv2.imread(img_path, 0)
img = np.array(img).reshape(img.shape[0]*img.shape[1])
data.append(img)
cls.append(c)
c = c + 1
data = np.array(data)
pca = PCA(n_components=10)
pca.fit(data)
D = np.array(pca.components_)
dictionary = SparseCoder(D, transform_algorithm='omp', transform_n_nonzero_coefs=10, transform_alpha=None)
features = dictionary.transform(data)
svm = LinearSVC()
svm.fit(features, cls)
pickle.dump(svm, open(output, 'wb'))
def RunSparseCodingScikit(q):
totalTimer = Timer()
# Load input dataset.
inputData = np.genfromtxt(self.dataset[0], delimiter=',')
dictionary = np.genfromtxt(self.dataset[1], delimiter=',')
# Get all the parameters.
l = re.search("-l (\d+)", options)
l = 0 if not l else int(l.group(1))
try:
with totalTimer:
# Perform Sparse Coding.
model = SparseCoder(dictionary=dictionary, transform_algorithm='lars',
transform_alpha=l)
code = model.transform(inputData)
except Exception as e:
q.put(-1)
return -1
time = totalTimer.ElapsedTime()
q.put(time)
return time
def fit(self, X, tuple_size=3, n_tuples=692,
weights=True, expand=True,
thres=0.7, cutoff=3):
"""
Discovers topics and used them as a dictionary for sparse-coding.
"""
models = self.discover_topics(X,
tuple_size=tuple_size,
n_tuples=n_tuples,
weights=weights,
expand=expand,
thres=thres,
cutoff=cutoff)
self.coder = SparseCoder(dictionary=normalize(models.toarray()),
transform_algorithm='lasso_lars',
split_sign=True,
n_jobs=4)
class SMHClassifier(BaseEstimator):
"""
SMH-based classifier.
"""
def __init__(self, tuple_size=3, n_tuples=692,
wcc=None, ovr_thres=0.7):
self.tuple_size = tuple_size
if wcc:
self.wcc = wcc
self.n_tuples = log(0.5) / log(1.0 - pow(wcc, tuple_size))
else:
self.n_tuples = n_tuples
def discover_topics(self, X, tuple_size=3, n_tuples=692,
weights=True, expand=True,
thres=0.7, cutoff=3):
"""
Discovers topics from a text corpus.
"""
ifs = array_to_listdb(X)
mined = ifs.mine(tuple_size=tuple_size,
num_tuples=n_tuples,
weights=weights,
expand=expand)
mined.cutoff(min=cutoff)
models = mined.cluster_mhlink(thres=thres)
return models
def fit(self, X, tuple_size=3, n_tuples=692,
weights=True, expand=True,
thres=0.7, cutoff=3):
"""
Discovers topics and used them as a dictionary for sparse-coding.
"""
models = self.discover_topics(X,
tuple_size=tuple_size,
n_tuples=n_tuples,
weights=weights,
expand=expand,
thres=thres,
cutoff=cutoff)
self.coder = SparseCoder(dictionary=normalize(models.toarray()),
transform_algorithm='lasso_lars',
split_sign=True,
n_jobs=4)
def fit_transform(self, X, tuple_size=3, n_tuples=692,
weights=None, expand=None,
thres=0.7, cutoff=3):
"""
Discovers topics and used them as a dictionary to sparse-code
the documents.
"""
self.fit(X, tuple_size=tuple_size, n_tuples=n_tuples, weights=weights, expand=expand,
thres=thres, cutoff=cutoff)
return self.coder.fit_transform(X.todense())
def transform(self, X):
"""
Sparse-code a given set of documents from the
discovered topics.
"""
return self.coder.transform(X.todense())
def scskl_reconstruction(data,mask,D):
output = np.zeros(data.shape)
fmap = np.zeros((D.shape[0]))
#fdata = np.zeros((data.shape[0],data.shape[1],data.shape[2],D.shape[0]))
px = np.int(np.around(np.power(D.shape[1],1/3))) #patch size (is assumed to be isotropic)
hpx = np.floor(px/2).astype(int)
nblock = 2 # number of block per dimension
subsize = np.ceil(np.array(data.shape) / nblock).astype(int)
med = np.median(data)
currentblock = 1
for x in range(np.ceil(data.shape[0]/subsize[0]).astype(int)):
xmin = x*subsize[0]
xmax = np.min((data.shape[0],(x+1)*subsize[0]))
for y in range(np.ceil(data.shape[1]/subsize[1]).astype(int)):
ymin = y*subsize[1]
ymax = np.min((data.shape[1],(y+1)*subsize[1]))
for z in range(np.ceil(data.shape[2]/subsize[2]).astype(int)):
zmin = z*subsize[2]
zmax = np.min((data.shape[2],(z+1)*subsize[2]))
print('Processing block : ',currentblock)
currentblock+=1
#Enlarge subimage to take into account block effect due to non-overlapping patches
xmin2 = np.max((0,xmin-hpx))
xmax2 = np.min((data.shape[0],xmax+hpx))
ymin2 = np.max((0,ymin-hpx))
ymax2 = np.min((data.shape[1],ymax+hpx))
zmin2 = np.max((0,zmin-hpx))
zmax2 = np.min((data.shape[2],zmax+hpx))
subdata = data[xmin2:xmax2,ymin2:ymax2,zmin2:zmax2]
submask = mask[xmin2:xmax2,ymin2:ymax2,zmin2:zmax2]
p = mp.array_to_patches(subdata,patch_shape=(px,px,px),normalization=False)
pm = mp.array_to_patches(submask,patch_shape=(px,px,px),normalization=False)
#remove patch we dont want to process
index = ~np.all(pm==0,axis=1)
subp = p[index]
subp -= med
if subp.shape[0] > 0:
print('Number of patches to process: ',subp.shape[0])
#Currently, there is a bug when using n_jobs>1 (https://github.com/scikit-learn/scikit-learn/issues/5956)
coder = SparseCoder(dictionary=D, transform_algorithm='omp')
code = coder.transform(subp).astype(np.float32)
fmap += np.sum((np.fabs(code)>0),axis=0)
subp = np.dot(code, D)
subp += med
p[index] = subp
suboutput = mp.patches_to_array(patches=p, patch_shape=(px,px,px), array_shape=subdata.shape)
tmpoutput = np.empty(data.shape)
tmpoutput[xmin2:xmax2,ymin2:ymax2,zmin2:zmax2]= suboutput
output[xmin:xmax,ymin:ymax,zmin:zmax] = tmpoutput[xmin:xmax,ymin:ymax,zmin:zmax]
# for a in range(D.shape[0]):
# for s in range(subp.shape[0]):
# subp[s,:] = code[s,a]
# p.fill(0)
# p[index] = subp
# fa = mp.patches_to_array(patches=p, patch_shape=(px,px,px), array_shape=subdata.shape)
# to = np.empty(data.shape)
# to[xmin2:xmax2,ymin2:ymax2,zmin2:zmax2]= fa
# fdata[xmin:xmax,ymin:ymax,zmin:zmax,a] = to[xmin:xmax,ymin:ymax,zmin:zmax]
#plt.bar(range(0,D.shape[0]), fmap)
#print('points in mask: ',np.sum(mask!=0))
#print('Number of non zero elements: ',np.sum(fmap)/np.sum(mask!=0))
#plt.show()
# return (output,fdata)
return output
def __init__(
self,
which_set,
frame_length,
overlap=0.5,
frames_per_example=1,
start=0,
stop=None,
audio_only=True,
n_prev_phones=0,
n_next_phones=0,
samples_to_predict=1,
filter_fn=None,
rng=_default_seed,
b=1.019,
step=64,
n_channels=64,
):
"""
Parameters
----------
which_set : str
Either "train", "valid" or "test"
frame_length : int
Number of acoustic samples contained in a frame
overlap : int, optional
Number of overlapping acoustic samples for two consecutive frames.
Defaults to 0, meaning frames don't overlap.
frames_per_example : int, optional
Number of frames in a training example. Defaults to 1.
start : int, optional
Starting index of the sequences to use. Defaults to 0.
stop : int, optional
Ending index of the sequences to use. Defaults to `None`, meaning
sequences are selected all the way to the end of the array.
audio_only : bool, optional
Whether to load only the raw audio and no auxiliary information.
Defaults to `False`.
rng : object, optional
A random number generator used for picking random indices into the
design matrix when choosing minibatches.
"""
self.frame_length = frame_length
if overlap < 1.0:
self.overlap = overlap * frame_length
else:
self.overlap = overlap
self.frames_per_example = frames_per_example
self.offset = self.frame_length - self.overlap
self.audio_only = audio_only
self.n_prev_phones = n_prev_phones
self.n_next_phones = n_next_phones
self.samples_to_predict = samples_to_predict
self.b = b
self.step = step
self.n_channels = n_channels
print "Frame length %d, overlap %d" % (self.frame_length, self.overlap)
# Initializing the dictionary
self.D = numpy.r_[
tuple(
gammatone_matrix(self.b, fc, self.frame_length, self.step)
for fc in erb_space(150, 8000, self.n_channels)
)
]
print "Using dictionary with shape", self.D.shape
self.coder = SparseCoder(
dictionary=self.D, transform_n_nonzero_coefs=None, transform_alpha=None, transform_algorithm="omp"
)
# RNG initialization
if hasattr(rng, "random_integers"):
self.rng = rng
else:
self.rng = numpy.random.RandomState(rng)
# Load data from disk
self._load_data(which_set)
examples_per_sequence = [0]
for sequence_id, samples_sequence in enumerate(self.raw_wav):
print "Sentence %d/%d" % (sequence_id, len(self.raw_wav))
X = segment_axis(samples_sequence, frame_length, overlap, end="pad")
X = numpy.hanning(self.frame_length) * X
self.raw_wav[sequence_id] = scipy.sparse.csr_matrix(self.coder.transform(X))
# TODO: change me
# Generate features/targets/phones/phonemes/words map
num_frames = self.raw_wav[sequence_id].shape[0]
num_examples = num_frames - self.frames_per_example
examples_per_sequence.append(num_examples)
self.cumulative_example_indexes = numpy.cumsum(examples_per_sequence)
self.samples_sequences = self.raw_wav
numpy.save("%s_sparse_frames.npy" % which_set, self.samples_sequences)
self.num_examples = self.cumulative_example_indexes[-1]
# DataSpecs
features_space = VectorSpace(dim=self.D.shape[0] * self.frames_per_example)
features_source = "features"
def features_map_fn(indexes):
rval = []
for sequence_index, example_index in self._fetch_index(indexes):
rval.append(
self.samples_sequences[sequence_index][
example_index : example_index + self.frames_per_example
].todense()
)
return rval
targets_space = VectorSpace(dim=self.D.shape[0])
targets_source = "targets"
def targets_map_fn(indexes):
rval = []
for sequence_index, example_index in self._fetch_index(indexes):
rval.append(self.samples_sequences[sequence_index][example_index + self.frames_per_example].todense())
return rval
space_components = [features_space, targets_space]
source_components = [features_source, targets_source]
map_fn_components = [features_map_fn, targets_map_fn]
batch_components = [None, None]
space = CompositeSpace(space_components)
source = tuple(source_components)
self.data_specs = (space, source)
self.map_functions = tuple(map_fn_components)
self.batch_buffers = batch_components
# Defaults for iterators
self._iter_mode = resolve_iterator_class("shuffled_sequential")
self._iter_data_specs = (CompositeSpace((features_space, targets_space)), (features_source, targets_source))
from scipy.io import wavfile
from scikits.talkbox import segment_axis
resolution = 160
step = 8
b = 1.019
n_channels = 64
overlap = 80
# Compute a multiscale dictionary
D_multi = np.r_[tuple(gammatone_matrix(b, fc, resolution, step) for
fc in erb_space(150, 8000, n_channels))]
# Load test signal
fs, y = wavfile.read('/home/jfsantos/data/TIMIT/TRAIN/DR1/FCJF0/SA1.WAV')
y = y / 2.0**15
Y = segment_axis(y, resolution, overlap=overlap, end='pad')
Y = np.hanning(resolution) * Y
# segments should be windowed and overlap
coder = SparseCoder(dictionary=D_multi, transform_n_nonzero_coefs=None, transform_alpha=1., transform_algorithm='omp')
X = coder.transform(Y)
density = len(np.flatnonzero(X))
out= np.zeros((np.ceil(len(y)/resolution)+1)*resolution)
for k in range(0, len(X)):
idx = range(k*(resolution-overlap),k*(resolution-overlap) + resolution)
out[idx] += np.dot(X[k], D_multi)
squared_error = np.sum((y - out[0:len(y)]) ** 2)
wavfile.write('reconst_%d_%d.wav'%(resolution,overlap), fs, np.asarray(out, dtype=np.float32))
class TIMITSparseGenerator(Dataset):
"""
Frame-based TIMIT dataset
"""
_default_seed = (17, 2, 946)
# Mean and standard deviation of the acoustic samples from the whole
# dataset (train, valid, test).
_mean = 0.0035805809921434142
_std = 542.48824133746177
def __init__(
self,
which_set,
frame_length,
overlap=0.5,
frames_per_example=1,
start=0,
stop=None,
audio_only=True,
n_prev_phones=0,
n_next_phones=0,
samples_to_predict=1,
filter_fn=None,
rng=_default_seed,
b=1.019,
step=64,
n_channels=64,
):
"""
Parameters
----------
which_set : str
Either "train", "valid" or "test"
frame_length : int
Number of acoustic samples contained in a frame
overlap : int, optional
Number of overlapping acoustic samples for two consecutive frames.
Defaults to 0, meaning frames don't overlap.
frames_per_example : int, optional
Number of frames in a training example. Defaults to 1.
start : int, optional
Starting index of the sequences to use. Defaults to 0.
stop : int, optional
Ending index of the sequences to use. Defaults to `None`, meaning
sequences are selected all the way to the end of the array.
audio_only : bool, optional
Whether to load only the raw audio and no auxiliary information.
Defaults to `False`.
rng : object, optional
A random number generator used for picking random indices into the
design matrix when choosing minibatches.
"""
self.frame_length = frame_length
if overlap < 1.0:
self.overlap = overlap * frame_length
else:
self.overlap = overlap
self.frames_per_example = frames_per_example
self.offset = self.frame_length - self.overlap
self.audio_only = audio_only
self.n_prev_phones = n_prev_phones
self.n_next_phones = n_next_phones
self.samples_to_predict = samples_to_predict
self.b = b
self.step = step
self.n_channels = n_channels
print "Frame length %d, overlap %d" % (self.frame_length, self.overlap)
# Initializing the dictionary
self.D = numpy.r_[
tuple(
gammatone_matrix(self.b, fc, self.frame_length, self.step)
for fc in erb_space(150, 8000, self.n_channels)
)
]
print "Using dictionary with shape", self.D.shape
self.coder = SparseCoder(
dictionary=self.D, transform_n_nonzero_coefs=None, transform_alpha=None, transform_algorithm="omp"
)
# RNG initialization
if hasattr(rng, "random_integers"):
self.rng = rng
else:
self.rng = numpy.random.RandomState(rng)
# Load data from disk
self._load_data(which_set)
examples_per_sequence = [0]
for sequence_id, samples_sequence in enumerate(self.raw_wav):
print "Sentence %d/%d" % (sequence_id, len(self.raw_wav))
X = segment_axis(samples_sequence, frame_length, overlap, end="pad")
X = numpy.hanning(self.frame_length) * X
self.raw_wav[sequence_id] = scipy.sparse.csr_matrix(self.coder.transform(X))
# TODO: change me
# Generate features/targets/phones/phonemes/words map
num_frames = self.raw_wav[sequence_id].shape[0]
num_examples = num_frames - self.frames_per_example
examples_per_sequence.append(num_examples)
self.cumulative_example_indexes = numpy.cumsum(examples_per_sequence)
self.samples_sequences = self.raw_wav
numpy.save("%s_sparse_frames.npy" % which_set, self.samples_sequences)
self.num_examples = self.cumulative_example_indexes[-1]
# DataSpecs
features_space = VectorSpace(dim=self.D.shape[0] * self.frames_per_example)
features_source = "features"
def features_map_fn(indexes):
rval = []
for sequence_index, example_index in self._fetch_index(indexes):
rval.append(
self.samples_sequences[sequence_index][
example_index : example_index + self.frames_per_example
].todense()
)
return rval
targets_space = VectorSpace(dim=self.D.shape[0])
targets_source = "targets"
def targets_map_fn(indexes):
rval = []
for sequence_index, example_index in self._fetch_index(indexes):
rval.append(self.samples_sequences[sequence_index][example_index + self.frames_per_example].todense())
return rval
space_components = [features_space, targets_space]
source_components = [features_source, targets_source]
map_fn_components = [features_map_fn, targets_map_fn]
batch_components = [None, None]
space = CompositeSpace(space_components)
source = tuple(source_components)
self.data_specs = (space, source)
self.map_functions = tuple(map_fn_components)
self.batch_buffers = batch_components
# Defaults for iterators
self._iter_mode = resolve_iterator_class("shuffled_sequential")
self._iter_data_specs = (CompositeSpace((features_space, targets_space)), (features_source, targets_source))
def _fetch_index(self, indexes):
digit = numpy.digitize(indexes, self.cumulative_example_indexes) - 1
return zip(digit, numpy.array(indexes) - self.cumulative_example_indexes[digit])
def _load_data(self, which_set):
"""
Load the TIMIT data from disk.
Parameters
----------
which_set : str
Subset of the dataset to use (either "train", "valid" or "test")
"""
# Check which_set
if which_set not in ["train", "valid", "test"]:
raise ValueError(
which_set + " is not a recognized value. " + "Valid values are ['train', 'valid', 'test']."
)
# Create file paths
timit_base_path = os.path.join(os.environ["PYLEARN2_DATA_PATH"], "timit/readable")
raw_wav_path = os.path.join(timit_base_path, which_set + "_x_raw.npy")
# Load data. For now most of it is not used, as only the acoustic
# samples are provided, but this is bound to change eventually.
# Set-related data
self.raw_wav = serial.load(raw_wav_path)
# self.scaler = serial.load(scaler_path)
def _validate_source(self, source):
"""
Verify that all sources in the source tuple are provided by the
dataset. Raise an error if some requested source is not available.
Parameters
----------
source : `tuple` of `str`
Requested sources
"""
for s in source:
try:
self.data_specs[1].index(s)
except ValueError:
raise ValueError("the requested source named '" + s + "' " + "is not provided by the dataset")
def get_data_specs(self):
"""
Returns the data_specs specifying how the data is internally stored.
This is the format the data returned by `self.get_data()` will be.
.. note::
Once again, this is very hacky, as the data is not stored that way
internally. However, the data that's returned by `TIMIT.get()`
_does_ respect those data specs.
"""
return self.data_specs
def get(self, source, indexes):
"""
.. todo::
WRITEME
"""
if type(indexes) is slice:
indexes = numpy.arange(indexes.start, indexes.stop)
self._validate_source(source)
rval = []
for so in source:
batch = self.map_functions[self.data_specs[1].index(so)](indexes)
batch_buffer = self.batch_buffers[self.data_specs[1].index(so)]
dim = self.data_specs[0].components[self.data_specs[1].index(so)].dim
if batch_buffer is None or batch_buffer.shape != (len(batch), dim):
batch_buffer = numpy.zeros((len(batch), dim), dtype=batch[0].dtype)
for i, example in enumerate(batch):
batch_buffer[i] = example
rval.append(batch_buffer)
return tuple(rval)
@functools.wraps(Dataset.iterator)
def iterator(self, mode=None, batch_size=None, num_batches=None, rng=None, data_specs=None, return_tuple=False):
"""
.. todo::
WRITEME
"""
if data_specs is None:
data_specs = self._iter_data_specs
# If there is a view_converter, we have to use it to convert
# the stored data for "features" into one that the iterator
# can return.
space, source = data_specs
if isinstance(space, CompositeSpace):
sub_spaces = space.components
sub_sources = source
else:
sub_spaces = (space,)
sub_sources = (source,)
convert = []
for sp, src in safe_zip(sub_spaces, sub_sources):
convert.append(None)
# TODO: Refactor
if mode is None:
if hasattr(self, "_iter_subset_class"):
mode = self._iter_subset_class
else:
raise ValueError("iteration mode not provided and no default " "mode set for %s" % str(self))
else:
mode = resolve_iterator_class(mode)
if batch_size is None:
batch_size = getattr(self, "_iter_batch_size", None)
if num_batches is None:
num_batches = getattr(self, "_iter_num_batches", None)
if rng is None and mode.stochastic:
rng = self.rng
return FiniteDatasetIterator(
self,
mode(self.num_examples, batch_size, num_batches, rng),
data_specs=data_specs,
return_tuple=return_tuple,
convert=convert,
)
# List the different sparse coding methods in the following format:
# (title, transform_algorithm, transform_alpha, transform_n_nozero_coefs)
estimators = [('OMP', 'omp', None, 15, 'navy'),
('Lasso', 'lasso_cd', 2, None, 'turquoise'), ]
lw = 2
plt.figure(figsize=(13, 6))
for subplot, (D, title) in enumerate(zip((D_fixed, D_multi),
('fixed width', 'multiple widths'))):
plt.subplot(1, 2, subplot + 1)
plt.title('Sparse coding against %s dictionary' % title)
plt.plot(y, lw=lw, linestyle='--', label='Original signal')
# Do a wavelet approximation
for title, algo, alpha, n_nonzero, color in estimators:
coder = SparseCoder(dictionary=D, transform_n_nonzero_coefs=n_nonzero,
transform_alpha=alpha, transform_algorithm=algo)
x = coder.transform(y.reshape(1, -1))
density = len(np.flatnonzero(x))
x = np.ravel(np.dot(x, D))
squared_error = np.sum((y - x) ** 2)
plt.plot(x, color=color, lw=lw,
label='%s: %s nonzero coefs,\n%.2f error'
% (title, density, squared_error))
# Soft thresholding debiasing
coder = SparseCoder(dictionary=D, transform_algorithm='threshold',
transform_alpha=20)
x = coder.transform(y.reshape(1, -1))
_, idx = np.where(x != 0)
x[0, idx], _, _, _ = np.linalg.lstsq(D[idx, :].T, y)
x = np.ravel(np.dot(x, D))
def encode_kmeans_sparsecode(df, km, algo='lasso_cd', alpha=1, split=False):
centroids = km.cluster_centers_
D = [centroids[i]/np.linalg.norm(centroids[i]) for i in range(len(centroids))]
D = np.array(D)
sc = SparseCoder(D, transform_algorithm=algo, transform_alpha=alpha, split_sign=split)
return pd.DataFrame(sc.transform(df))
# List the different sparse coding methods in the following format:
# (title, transform_algorithm, transform_alpha, transform_n_nozero_coefs)
estimators = [('OMP', 'omp', None, 15, 'navy'),
('Lasso', 'lasso_cd', 2, None, 'turquoise'), ]
lw = 2
plt.figure(figsize=(13, 6))
for subplot, (D, title) in enumerate(zip((D_fixed, D_multi),
('fixed width', 'multiple widths'))):
plt.subplot(1, 2, subplot + 1)
plt.title('Sparse coding against %s dictionary' % title)
plt.plot(y, lw=lw, linestyle='--', label='Original signal')
# Do a wavelet approximation
for title, algo, alpha, n_nonzero, color in estimators:
coder = SparseCoder(dictionary=D, transform_n_nonzero_coefs=n_nonzero,
transform_alpha=alpha, transform_algorithm=algo)
x = coder.transform(y)
density = len(np.flatnonzero(x))
x = np.ravel(np.dot(x, D))
squared_error = np.sum((y - x) ** 2)
plt.plot(x, color=color, lw=lw,
label='%s: %s nonzero coefs,\n%.2f error'
% (title, density, squared_error))
# Soft thresholding debiasing
coder = SparseCoder(dictionary=D, transform_algorithm='threshold',
transform_alpha=20)
x = coder.transform(y)
_, idx = np.where(x != 0)
x[0, idx], _, _, _ = np.linalg.lstsq(D[idx, :].T, y)
x = np.ravel(np.dot(x, D))
def getSparseCodes(dataset,Dict):
print Dict.shape
print dataset.shape
coder=SparseCoder(Dict,transform_algorithm='lasso_lars')
return coder.transform(dataset)
class SparseCoding(object):
def __init__(self, log_lev='INFO', sparse_dim_rat=None, name='',
dist_beta=0.1, dist_sigma=0.005, display=0):
LOG.setLevel(log_lev)
self.name = name
self.codebook_comps = None
self.active_set = None
self.min_coeff = max([1,
co.CONST['sparse_fss_min_coeff']])
self.min_coeff_rat = co.CONST['sparse_fss_min_coeff_rat']
self.gamma = co.CONST['sparse_fss_gamma']
self.rat = None
if isinstance(self.gamma, str):
if self.gamma.starts_with('var'):
try:
self.rat = [float(s) for s in str.split() if
co.type_conv.isfloat(s)][0]
except IndexError:
self.rat = None
self.inp_features = None
self.sparse_features = None
self.basis_constraint = 1
self.inv_codebook_comps = None
self.res_codebook_comps = None
self.max_iter = 500
self.dict_max_iter = 300
self.display = display
self.prev_err = 0
self.curr_error = 0
self.allow_big_vals = False
self.sparse_dim_rat = sparse_dim_rat
if sparse_dim_rat is None:
self.sparse_dim_rat = co.CONST['sparse_dim_rat']
self.theta = None
self.prev_sparse_feats = None
self.flush_flag = False
self.sparse_feat_list = None
self.inp_feat_list = None
self.codebook = None
self.time = []
def flush_variables(self):
'''
Empty variables
'''
self.active_set = None
self.theta = None
self.codebook_comps = None
self.inp_features = None
self.inp_feat_list = None
self.sparse_features = None
self.flush_flag = True
self.res_codebook_comps = None
self.prev_err = 0
self.curr_error = 0
self.lbds = 0.5*np.ones(self.sparse_dim)
def initialize(self, feat_dim,
init_codebook_comps=None):
'''
Initialises B dictionary and s
'''
self.sparse_dim = self.sparse_dim_rat * feat_dim
if init_codebook_comps is not None:
if (init_codebook_comps.shape[0] == feat_dim and
init_codebook_comps.shape[1] == self.sparse_dim_rat *
feat_dim):
self.codebook_comps = init_codebook_comps.copy()
else:
raise Exception('Wrong input of initial B matrix, the dimensions' +
' should be ' + str(feat_dim) + 'x' +
str(self.sparse_dim) + ', not ' +
str(init_codebook_comps.shape[0]) + 'x' +
str(init_codebook_comps.shape[1]))
if (self.codebook_comps is None) or self.flush_flag:
LOG.warning('Non existent codebook, manufactuning a random one')
self.codebook_comps = random.random((feat_dim, self.sparse_dim))
if (self.sparse_features is None) or self.flush_flag:
self.sparse_features = zeros((self.sparse_dim, 1))
self.theta = zeros(self.sparse_dim)
self.active_set = zeros((self.sparse_dim), bool)
self.sparse_features = zeros((self.sparse_dim, 1))
self.flush_flag = False
self.is_trained = False
def object_val_calc(self, codebook_comps, ksi, gamma, theta, vecs):
'''
Calculate objective function value
'''
_bs_ = np.dot(codebook_comps, vecs)
square_term = 0.5 * npsum((ksi - _bs_)**2, axis=0)
res = (square_term + gamma * dot(theta.T, vecs)).ravel()
return res
def feature_sign_search_algorithm(self,
inp_features,
acondtol=1e-3,
ret_error=False,
display_error=False,
max_iter=0,
single=False, timed=True,
starting_points=None,
training=False):
'''
Returns sparse features representation
'''
self.min_coeff_rat = co.CONST['sparse_fss_min_coeff_rat']
self.min_coeff = max([self.min_coeff,
self.min_coeff_rat *
np.size(inp_features)])
if self.inp_feat_list is not None:
self.inp_feat_list.append(inp_features.ravel())
else:
self.inp_feat_list = [inp_features.ravel()]
self.inp_features = inp_features.copy().reshape((-1,1))
# Step 1
btb = dot(self.codebook_comps.T, self.codebook_comps)
btf = dot(self.codebook_comps.T, self.inp_features)
if self.rat is not None:
self.gamma = np.max(np.abs(-2 * btf)) * self.rat
gamma = self.gamma
if starting_points is not None:
self.sparse_features = starting_points.reshape((self.sparse_dim,
1))
self.theta = np.sign(self.sparse_features)
self.active_set[:] = False
self.active_set[self.sparse_features.ravel()!=0] = True
step2 = 0
else:
step2 = 1
count = 0
prev_objval = 0
if max_iter == 0:
max_iter = self.max_iter
else:
self.max_iter = max_iter
self.prev_sparse_feats = None
prev_error = 0
initial_energy = compute_lineq_error(inp_features, 0,
0)
interm_error = initial_energy
SPLOG.info('Initial Signal Energy: ' + str(initial_energy))
SPLOG.info('Initial nonzero elements number: ' +
str(np.sum(inp_features!=0)))
converged = False
for count in range(self.max_iter):
# Step 2
if step2:
zero_coeffs = (self.sparse_features == 0)
qp_der_outfeati = 2 * \
(dot(btb, self.sparse_features)
- btf) * zero_coeffs.reshape((-1,1))
i = argmax(npabs(qp_der_outfeati))
if (npabs(qp_der_outfeati[i]) > gamma
or npsum(self.active_set) < self.min_coeff):
self.theta[i] = -sign(qp_der_outfeati[i])
self.active_set[i] = True
# Step 3
codebook_comps_h = self.codebook_comps[:, self.active_set]
sparse_feat_h = self.sparse_features[self.active_set].reshape(
(-1,1))
theta_h = self.theta[self.active_set].reshape((-1,1))
_q_ = dot(codebook_comps_h.T, self.inp_features) - gamma * theta_h / 2.0
codebook_comps_h2 = dot(codebook_comps_h.T, codebook_comps_h)
rank = matrix_rank(codebook_comps_h2)
zc_search = True
if rank == codebook_comps_h2.shape[0]:
new_sparse_f_h = np.linalg.solve(codebook_comps_h2, _q_)
else:
u,s,v = np.linalg.svd(codebook_comps_h2)
col_space = u[:, :rank]
null_space = u[:, rank:]
#Check if q belongs in column space, ie the projection of
#q in the column space is q itself
q_proj = np.zeros_like(_q_).reshape(-1, 1)
for i in range(col_space.shape[1]):
col = col_space[:,i].reshape(-1, 1)
q_proj+=((dot(_q_.reshape(1,-1),col) /
np.dot(col.T, col).astype(float))*col)
'''
LOG.info('q|Projection: ' +
str(np.concatenate((_q_.reshape(-1,1),q_proj),axis=1)))
LOG.info('Projection Energy: '+ str(np.sum(q_proj**2)))
LOG.info('Distance between q and projection: '+str(np.linalg.norm(q_proj.ravel()-_q_.ravel())))
'''
if np.allclose(q_proj.ravel()-_q_.ravel(), 0, atol=1.e-6):
new_sparse_f_h = dot(pinv(codebook_comps_h2),_q_)
else:
#direction z in nullspace of codebook_comps_h2 can not be
#perpendicular to _q_, because then _q_ = C(codebook_comps_h2),
#which was proven not to hold.
#I take the principal vector that belongs in null_space of
#codebook_comps_h2 and add it to the current sparse_feat_h
#so that to search for zerocrossings
#inside the line constructed
# by this vector and sparse_feat_h, which has direction,
# belonging to null_space of codebook_comps_h2
tmp_sparse_f_h = sparse_feat_h + dot(null_space,
np.ones((null_space.shape[1],1)))
zero_points_lin_par = sparse_feat_h / (sparse_feat_h
-
tmp_sparse_f_h).astype(float)
# find _t_ that corresponds to the closest zero crossing to
# sparse_feat_h
_t_ind = np.argmin(np.abs(zero_points_lin_par[
np.isfinite(zero_points_lin_par)]))
_t_ = zero_points_lin_par[
np.isfinite(zero_points_lin_par)][_t_ind]
null_vec = _t_ * tmp_sparse_f_h + (1 - _t_) * sparse_feat_h
new_sparse_f_h = null_vec
zc_search = False
if (np.prod(sign(sparse_feat_h) != sign(new_sparse_f_h))
and zc_search):
zero_points_lin_par = sparse_feat_h / (sparse_feat_h -
new_sparse_f_h).astype(float)
zero_points_lin_par = concatenate((zero_points_lin_par[
((zero_points_lin_par > 0) *
(zero_points_lin_par < 1)).astype(bool)][:], array([1])), axis=0)
_t_ = zero_points_lin_par
null_vecs = _t_ * new_sparse_f_h + (1 - _t_) * sparse_feat_h
objvals = self.object_val_calc(codebook_comps_h, self.inp_features, gamma,
theta_h,
null_vecs).flatten()
objval_argmin = argmin(objvals)
objval = np.min(objvals)
new_sparse_f_h = null_vecs[:, objval_argmin][:, None].copy()
else:
objval = self.object_val_calc(codebook_comps_h, self.inp_features, gamma, theta_h,
new_sparse_f_h)
self.sparse_features[self.active_set] = new_sparse_f_h.copy()
self.active_set[self.active_set] = np.logical_not(
isclose(new_sparse_f_h, 0))
if npsum(self.active_set) < self.min_coeff:
step2 = 1
continue
self.theta = sign(self.sparse_features)
# Step 4
nnz_coeff = self.sparse_features != 0
# a
new_qp_der_outfeati = 2 * (dot(btb, self.sparse_features) - btf)
cond_a = (new_qp_der_outfeati +
gamma * sign(self.sparse_features)) * nnz_coeff
'''
if np.abs(objval) - np.abs(prev_objval) > 100 and not\
self.allow_big_vals and not count == 0:
if self.prev_sparse_feats is not None:
SPLOG.info('Current Objective Function value: ' +
str(np.abs(objval)))
SPLOG.info('Previous Objective Function value: ' +
str(np.abs(prev_objval)))
SPLOG.info('Problem with big values of inv(B^T*B)' +
',you might want to increase atol' +
' or set flag allow_big_vals to true' +
' (this might cause' +
' problems)')
SPLOG.info('Reverting to previous iteration result ' +
'and exiting loop..')
self.sparse_features = self.prev_sparse_feats.ravel()
break
else:
LOG.error('Current Objective Function value: ' +
str(np.abs(objval)))
LOG.error('Previous Objective Function value: ' +
str(np.abs(prev_objval)))
LOG.error('Problem with big values of inv(B^T*B),increase atol' +
' or set flag allow_big_vals to true (this might cause' +
' serious convergence problems)')
LOG.error('Exiting as algorithm has not produced any'
+ ' output results.')
exit()
'''
prev_objval = objval
self.prev_sparse_feats = self.sparse_features
if allclose(cond_a, 0, atol=acondtol):
# go to cond b:
z_coeff = self.sparse_features == 0
cond_b = npabs(new_qp_der_outfeati * z_coeff) <= gamma
if npsum(cond_b) == new_qp_der_outfeati.shape[0]:
self.sparse_features = self.sparse_features.reshape((-1,1))
converged = True
break
else:
# go to step 2
step2 = 1
else:
# go to step 3
step2 = 0
if count % 10 == 0:
interm_error = compute_lineq_error(
self.inp_features, self.codebook_comps,
self.sparse_features)
if interm_error == prev_error or interm_error > initial_energy:
converged=True
break
else:
prev_error = interm_error
SPLOG.info('\t Epoch:' + str(count))
SPLOG.info('\t\t Intermediate Error=' +
str(interm_error))
if interm_error < 0.001:
converged=True
SPLOG.info('Small error, asssuming convergence')
break
'''
if initial_energy < interm_error:
if not training:
LOG.warning('FSS Algorithm did not converge, using pseudoinverse' +
' of provided codebook instead')
if self.inv_codebook_comps is None:
self.inv_codebook_comps = pinv(self.codebook_comps)
self.sparse_features=dot(self.inv_codebook_comps,self.inp_features).ravel()
else:
SPLOG.info('FSS Algorithm did not converge,' +
' removing sample from training dataset...')
self.sparse_features = None
return (interm_error), False, initial_energy
else:
'''
if not converged:
SPLOG.info('FSS Algorithm did not converge' +
' in the given iterations')
else:
SPLOG.info('Successful Convergence')
SPLOG.info('\tFinal error: ' + str(interm_error))
SPLOG.info('\tNumber of nonzero elements: ' +
str(np.sum(self.sparse_features!=0)))
if not single:
if self.sparse_feat_list is None:
self.sparse_feat_list = [self.sparse_features.ravel()]
else:
self.sparse_feat_list.append(self.sparse_features.ravel())
if ret_error:
return (compute_lineq_error(self.inp_features, self.codebook_comps,
self.sparse_features),
True, initial_energy)
self.sparse_features = self.sparse_features.ravel()
return None, True, None
def lagrange_dual(self, lbds, ksi, _s_, basis_constraint):
'''
Lagrange dual function for the minimization problem
<ksi> is input, <_s_> is sparse,
'''
lbds[lbds==0] = 10**(-18) #the drawback of this method
self.ksist = dot(ksi, _s_.T)
interm_result = inv(
dot(_s_, _s_.T) + diag(lbds.ravel()))
LOG.debug('Computed Lagrange Coefficients:\n'+str(np.unique(lbds)))
res = ((dot(ksi.T,ksi)).trace() -
(dot(dot(self.ksist, interm_result), self.ksist.T)).trace() -
(basis_constraint * diag(lbds.ravel())).trace())
return -res # minimizing negative = maximizing positive
def lagrange_dual_grad(self, lbds, ksi, _s_, basis_constraint):
'''
Gradient of lagrange dual function, w.r.t. elf.codebook_comps,
self.sparse_feat_list,
self.are_sparsecoded_inp) = self.pickle.load(inp)
s_
'''
# lbds=lbds.flatten()
interm_result = inv(
dot(_s_, _s_.T) + diag(lbds.ravel()))
interm_result = dot(self.ksist, interm_result)
interm_result = dot(interm_result.T,interm_result)
res = diag(interm_result) - basis_constraint
return -res # minimizing negative = maximizing positive
def lagrange_dual_hess(self, lbds, ksi, _s_, basis_constraint):
'''
It is not used, but it is here in case numpy solver gets also
the hessian as input
'''
interm_result = inv(
dot(_s_, _s_.T) + diag(lbds.ravel()))
interm_result1 = dot(self.ksist, interm_result)
res = -2 * dot(interm_result1.T, interm_result1) * interm_result
return -res #minimizing negative = maximizing positive
# pylint: disable=no-member
def conj_grad_dict_compute(self):
'''
Function to train nxm matrix using truncated newton method
'''
options = {'disp':True}
'''
if self.res_codebook_comps is None:
self.res_codebook_comps = self.codebook_comps
LOG.info(self.res_codebook_comps.shape)
'''
res = minimize(self.lagrange_dual,
self.lbds.copy(),
method='Newton-CG',
jac=self.lagrange_dual_grad,
#hess=self.lagrange_dual_hess,
#bounds=np.array(([(10**(-18), 10**10)] *
# self.sparse_feat_list.shape[0])),
#stepmx=50.0,
#maxCGit=20,
#maxfun=100,
options=options,
#fmin=0.1,
#ftol=0.1,
#xtol=0.001,
#rescale=1.5,
args=(self.are_sparsecoded_inp.copy(),
self.sparse_feat_list.copy(),
self.basis_constraint)
)
LOG.info(res)
self.lbds = res.x
LOG.info(np.unique(self.lbds))
interm_result = (self.lbds+
dot(self.sparse_feat_list,
self.sparse_feat_list.T))
LOG.info(np.linalg.rank(interm_result))
codebook_comps = dot(inv(interm_result),
self.ksist.T).T
return codebook_comps
# pylint: enable=no-member
def train_sparse_dictionary(self, data, sp_opt_max_iter=200,
init_traindata_num=200, incr_rate=2,
min_iterations=3, init_codebook_comps=None,
log_lev=None, n_jobs=4):
if log_lev is not None:
LOG.setLevel(log_lev)
self.codebook_comps = DictionaryLearning(
n_components=self.sparse_dim_rat * data.shape[1],
alpha=co.CONST['sparse_alpha'],
verbose=1, n_jobs=n_jobs).fit(data).components_.T
@timeit
def code1(self, data, max_iter=None, errors=False):
'''
Sparse codes a single feature
Requires that the dictionary is already trained
'''
if self.codebook is None:
self.codebook = SparseCoder(self.codebook_comps.T,n_jobs=4)
return self.codebook.transform(data.reshape(1,-1)).ravel()
def train_sparse_dictionary1(self, data, sp_opt_max_iter=200,
init_traindata_num=200, incr_rate=2,
min_iterations=3, init_codebook_comps=None,
debug=False):
'''
<data> is a numpy array, holding all the features(of single kind) that
are required to train the sparse dictionary, with dimensions
[n_features, n_samples]. The sparse dictionary is trained with a random
subset of <data>, which is increasing in each iteration with rate
<incr_rate> , along with the max iterations <sp_opt_max_iter> of feature
sign search algorithm. <min_iterations> is the least number of
iterations of the dictionary training, after total data is processed.
'''
self.sparse_dim = min(data.shape) * self.sparse_dim_rat
self.flush_variables()
try:
import progressbar
except:
LOG.warning('Install module progressbar2 to get informed about the'
+' feature sign search algorithm progress')
pass
self.initialize(data.shape[0], init_codebook_comps=init_codebook_comps)
iter_count = 0
retry_count = 0
LOG.info('Training dictionary: ' + self.name)
LOG.info('Minimum Epochs number after total data is processed:' + str(min_iterations))
reached_traindata_num = False
reached_traindata_count = 0
computed = data.shape[1] * [None]
retry = False
lar_approx = False
while True:
LOG.info('Epoch: ' + str(iter_count))
loaded = False
self.sparse_feat_list = None
self.inp_feat_list = None
if debug and iter_count == 0:
LOG.warning('Debug is on, loading data from first FSS execution')
try:
with open(self.name+' debug_sparse.pkl','r') as inp:
(self.codebook_comps,
self.sparse_feat_list,
self.are_sparsecoded_inp) = pickle.load(inp)
loaded=True
except (IOError, EOFError):
LOG.warning('Not existent '+self.name
+' debug_sparse.pkl')
if not loaded:
train_num = min(int(init_traindata_num *
(incr_rate) ** iter_count),
data.shape[1])
if train_num == data.shape[1] and not reached_traindata_num:
reached_traindata_num = True
LOG.info('Total data is processed')
if reached_traindata_num:
reached_traindata_count += 1
LOG.info('Number of samples used: ' + str(train_num))
ran = rand.sample(xrange(data.shape[1]), train_num)
feat_sign_max_iter = min(1000,
sp_opt_max_iter * incr_rate ** iter_count)
LOG.info('Feature Sign Search maximum iterations allowed:'
+ str(feat_sign_max_iter))
try:
format_custom_text = progressbar.FormatCustomText(
'Mean Initial Error: %(mean_init_energy).4f,'+
' Mean Final Error: %(mean).4f ,Valid Samples Ratio: %(valid).2f',
dict(
mean_init_energy=0,
mean=0,
valid=0
),
)
pbar = progressbar.ProgressBar(max_value=train_num - 1,
redirect_stdout=True,
widgets=[progressbar.widgets.Percentage(),
progressbar.widgets.Bar(),
format_custom_text])
errors=True
sum_error = 0
sum_energy = 0
except UnboundLocalError:
pbar = None
errors = False
pass
are_sparsecoded = []
if pbar is not None:
iterat = pbar(enumerate(ran))
else:
iterat = enumerate(ran)
for count, sample_count in iterat:
fin_error, valid, init_energy = self.feature_sign_search_algorithm(
data[:, sample_count],
max_iter=feat_sign_max_iter,
ret_error=errors,training=True,
starting_points=computed[sample_count])
are_sparsecoded.append(True)
try:
if iter_count > 0 and valid:
#do not trust first iteration sparse features, before
#having trained the codebooks at least once
computed[sample_count] = self.sparse_feat_list[-1]
except (TypeError,AttributeError):
pass
if valid and pbar and errors:
sum_error += fin_error
mean_error = sum_error/float(sum(are_sparsecoded))
sum_energy += init_energy
mean_init_energy = sum_energy/float(sum(are_sparsecoded))
if pbar is not None:
format_custom_text.update_mapping(mean_init_energy=
mean_init_energy,
mean=mean_error,
valid=sum(are_sparsecoded)
/float(len(are_sparsecoded)))
self.initialize(data.shape[0])
self.inp_feat_list = np.transpose(np.array(self.inp_feat_list))
self.sparse_feat_list = np.array(self.sparse_feat_list).T
are_sparsecoded = np.array(
are_sparsecoded).astype(bool)
retry = np.sum(are_sparsecoded) < 1 / 3.0 * (are_sparsecoded).size
self.are_sparsecoded_inp = self.inp_feat_list[:, are_sparsecoded]
if debug and iter_count==0:
LOG.warning('Debug is on, saving debug_sparse.pkl')
with open(self.name + ' debug_sparse.pkl','w') as out:
pickle.dump((self.codebook_comps,
self.sparse_feat_list,
self.are_sparsecoded_inp), out)
prev_error = compute_lineq_error(self.are_sparsecoded_inp, self.codebook_comps,
self.sparse_feat_list)
if not lar_approx:
dictionary = self.conj_grad_dict_compute()
curr_error = compute_lineq_error(
self.are_sparsecoded_inp,
dictionary,
self.sparse_feat_list)
LOG.info('Reconstruction Error: ' + str(curr_error))
if loaded:
mean_init_energy=0
mean_error = 0
if curr_error > prev_error or mean_error>1000 or retry or lar_approx:
if (prev_error > 100 or mean_error>1000
or retry or lar_approx):
if retry_count == 2 or lar_approx:
if iter_count != 0:
iter_count = 0
lar_approx = True
init_traindata_num = data.shape[1]
continue
LOG.warning('Training has high final error but' +
' reached maximum retries. No codebook can'
+ ' be produced with the fast method,'+
' using Lagrange Dual, as input'+
' sparsecoded data S is'
+' ill-conditioned (too low' +
' rank of the STS).'+
' Least Angle Regression Method '+
' will be used')
self.codebook_comps = DictionaryLearning(
self.sparse_dim,
fit_algorithm='lars',
code_init=self.inp_feat_list.T).fit(
self.are_sparsecoded_inp.T).components_.T
curr_error = compute_lineq_error(
self.are_sparsecoded_inp,
self.codebook_comps,
self.sparse_feat_list)
LOG.info('Reconstruction Error using LARS: '
+ str(curr_error))
if curr_error > 1000:
LOG.info('LARS method did not converge,' +
' no codebook is produced.')
self.is_trained = False
self.codebook_comps = None
else:
break
LOG.warning('Training of codebook ' + self.name + ' completed with no success,'+
' reinitializing (Retry:' + str(retry_count + 1) + ')')
self.flush_variables()
self.initialize(data.shape[0])
computed = data.shape[1] * [None]
retry_count += 1
iter_count = -1
reached_traindata_count = 0
reached_traindata_num = False
elif (np.isclose(prev_error,curr_error,atol=0.1)
and reached_traindata_num and
reached_traindata_count > min_iterations):
break
if curr_error < 0.5 and reached_traindata_num:
break
if (reached_traindata_num and
reached_traindata_count > min_iterations and
iter_count >= 0):
break
iter_count += 1
self.codebook_comps = dictionary
self.inp_feat_list = None
self.sparse_feat_list = None
self.is_trained = True
@timeit
def code(self, data, max_iter=None, errors=False):
'''
Sparse codes a single feature
Requires that the dictionary is already trained
'''
if max_iter is None:
max_iter = co.CONST['sparse_fss_max_iter']
self.initialize(data.size)
self.feature_sign_search_algorithm(data.ravel(), max_iter=max_iter,
single=True, display_error=errors,
ret_error=errors)
return self.sparse_features
def multicode(self, data, max_iter=None, errors=False):
'''
Convenience method for sparsecoding multiple features.
<data> is assumed to have dimensions [n_features, n_samples]
output has dimensions [n_sparse, n_samples]
'''
feat_dim = 0
for datum in data:
if datum is not None:
feat_dim = len(datum)
if feat_dim == 0 :
raise Exception('Bad Input, full of nans')
sparse_features = np.zeros((len(data),
self.sparse_dim_rat* feat_dim))
for count in range(len(data)):
if data[count] is not None and np.prod(np.isfinite(data[count][:])):
sparse_features[count, :] = self.code(data[count][:],
max_iter, errors).ravel()
else:
sparse_features[count, :] = np.nan
return sparse_features
