def test_cp_decomposition(self):
test_data1 = T.array([[[0, 1, 3, 4], [4, 0, 2, 1], [4, 2, 3, 4]],
[[2, 4, 2, 3], [3, 3, 2, 4], [2, 3, 0, 2]]])
test_data2 = T.array([[[3, 1, 1, 2], [1, 0, 3, 2], [3, 4, 0, 2]],
[[1, 2, 3, 3], [2, 3, 1, 0], [1, 2, 0, 2]]])
factors, lamda = cp(test_data1,
r=3,
stop_iter=500,
tol=1e-6,
normalize_factor=True,
random_seed=44)
P, fit, itr = cp_als(dtensor(test_data1), 3, init='random')
T.testing.assert_array_almost_equal(reconstruct_tensor(
factors, lamda, (2, 3, 4)),
P.toarray(),
decimal=0)
factors, lamda = cp(test_data2,
r=3,
stop_iter=500,
tol=1e-6,
normalize_factor=True,
random_seed=44)
P, fit, itr = cp_als(dtensor(test_data2), 3, init='random')
T.testing.assert_array_almost_equal(reconstruct_tensor(
factors, lamda, (2, 3, 4)),
P.toarray(),
decimal=0)
def get_cp_factors(layer, rank, cin, cout, kernel_size, **kwargs):
weights, bias = get_weights_and_bias(layer)
w_h = None
w_w = None
w_cin = None
w_cout = None
if isinstance(layer, keras.Sequential):
w_cout, w_cin, w_h, w_w = recompress_ncpd_tensor(weights, new_rank=rank, max_cycle=500, return_fit=False,
tensor_format="cpd")
elif isinstance(layer, keras.layers.Conv2D):
P, _, _ = cp_als(dtensor(weights), rank, init="random")
w_cin, w_cout, w_h, w_w = extract_weights_tensors(P)
# Reshape to proper kernel sizes.
w_h = w_h.T.reshape((rank, 1, kernel_size[0], 1))
w_w = w_w.T.reshape((rank, 1, 1, kernel_size[1]))
w_cin = w_cin.T.reshape((rank, cin, 1, 1))
w_cout = w_cout.reshape([cout, rank, 1, 1])
# Reorder to TensorFlow order.
w_cin, w_cout = [to_tf_kernel_order(w) for w in [w_cin, w_cout]]
# The middle layers are depthwise it should have order
# [rank, 1, kernel_size, kernel_size].
# This reorders it correctly from TensorFlow order to PyTorch order
# w_h, w_w = [depthwise_to_pytorch_kernel_order(w) for w in [w_h, w_w]].
w_h, w_w = [to_tf_kernel_order(w) for w in [w_h, w_w]]
return [w_cin, w_h, w_w, w_cout], [None, None, None, bias]
def get_cp_factors(layer, rank, cin, cout, kernel_size, **kwargs):
weights, bias = get_weights_and_bias(layer)
w_cin = None
w_z = None
w_cout = None
if isinstance(layer, keras.Sequential):
w_cout, w_cin, w_z = recompress_ncpd_tensor(weights,
new_rank=rank,
max_cycle=500,
return_fit=False,
tensor_format="cpd")
elif isinstance(layer, keras.layers.Conv2D):
P, _, _ = cp_als(dtensor(weights), rank, init="random")
w_cin, w_cout, w_z = extract_weights_tensors(P)
if w_cin is None or w_z is None or w_cout is None:
raise CompressionError()
# Reshape to the proper PyTorch shape order.
w_cin = w_cin.T.reshape((rank, cin, 1, 1))
w_z = w_z.T.reshape((rank, 1, *kernel_size))
w_cout = w_cout.reshape((cout, rank, 1, 1))
# Reorder to TensorFlow order.
w_cin, w_z, w_cout = [to_tf_kernel_order(w) for w in [w_cin, w_z, w_cout]]
return [w_cin, w_z, w_cout], [None, None, bias]
def test_cp_reconstruction(self):
data = T.array([[[3, 1, 1, 2], [1, 0, 3, 2], [3, 4, 0, 2]],
[[1, 2, 3, 3], [2, 3, 1, 0], [1, 2, 0, 2]]])
tensor = dtensor(data)
P, fit, itr = cp_als(tensor, 3, init='random')
T.testing.assert_array_almost_equal(
P.toarray(), reconstruct_tensor(P.U, P.lmbda, (2, 3, 4)))
def tensor_decomp(X):
print("CP-ALS Decomposition.")
T = dtensor(XX)
P, fit, itr, exectimes = cp_als(T, 2, init='nvecs')
proj = P.U[2]
fproj = np.abs(np.fft.fft(proj, axis=0))[:XX.shape[-1] // 2, :]
return fproj, proj
def get_cp_factors(self):
if self.pretrained is not None:
mat_dict = scipy.io.loadmat(self.pretrained)
if mat_dict['R'][0][0] != self.rank:
print('WRONG FACTORS, do not correspond to desired rank')
PU_z, PU_cout, PU_cin = [Ui[0] for Ui in mat_dict['P_bals_epc_U']]
Plmbda = mat_dict['P_bals_epc_lambda'].ravel()
f_cin = np.array(PU_cin)
f_cout = np.array(PU_cout)
f_z = (np.array(PU_z) * (Plmbda))
else:
if '__getitem__' in dir(self.layer):
f_cout_old, f_cin_old, f_z_old = self.weight
f_cout_old = np.array(f_cout_old)
f_cin_old = np.array(f_cin_old)
f_z_old = np.array(f_z_old)
bias = self.bias
f_cout, f_cin, f_z = recompress_ncpd_tensor(
[f_cout_old, f_cin_old, f_z_old],
new_rank=self.rank,
max_cycle=500,
return_fit=False,
tensor_format='cpd')
else:
if self.weight.is_cuda:
self.weight = self.weight.cpu()
if self.bias is not None:
self.bias = self.bias.cpu()
weights = dtensor(self.weight)
bias = self.bias
T = dtensor(self.weight)
P, fit, itr, exectimes = cp_als(T, self.rank, init='random')
f_cout = np.array(P.U[0])
f_cin = np.array(P.U[1])
f_z = (np.array(P.U[2]) * (P.lmbda))
f_cin = torch.FloatTensor(
np.reshape(f_cin.T, [self.rank, self.cin, 1, 1])).contiguous()
f_z = torch.FloatTensor(
np.reshape(f_z.T, [self.rank, 1, *self.kernel_size])).contiguous()
f_cout = torch.FloatTensor(
np.reshape(f_cout, [self.cout, self.rank, 1, 1])).contiguous()
return [f_cin, f_z, f_cout], [None, None, bias]
def __init__(self, data, rank=8):
"""
Args:
data: np.ndarray : the underlying multi-dimensional array
rank:
Returns:
BaseTensor object
"""
self.rank = rank
self.shape = data.shape
self.ktensor = cp_als(dtensor(data), (self.rank, ) * 3)[0]
def _splitrank3(h, verbose=False):
# this fixes a bug in scikit-tensor
if np.sum(np.abs(h)) < 1.e-30:
return tuple(np.zeros(s) for s in h.shape) + (np.zeros_like(h), )
P, fit, itr, exectimes = cp_als(dtensor(h.copy()), 1)
hx, hy, hz = [(P.lmbda[0])**(1. / 3) * np.array(P.U[i])[:, 0]
for i in range(3)]
if verbose:
print("lambdas= %s \nfit = %s \niterations= %s " % (P.lmbda, fit, itr))
return hx, hy, hz, P.toarray()
def load_data():
all_subjects = range(1, 24)
X = []
y = []
ids_test = []
label_count = []
n_basis = 10
all_U0 = []
all_U2 = []
for n, subject in enumerate(all_subjects):
if subject < 17:
filename = 'data/train_subject%02d.mat' % subject
else:
filename = 'data/test_subject%02d.mat' % subject
print("Loading", filename)
data = loadmat(filename, squeeze_me=True)
XX = window_filter_baseline(data['X'])
mask = get_outlier_mask(XX)
T = dtensor(XX)
r = cp_als(T, rank=n_basis)
U0 = r[0].U[0]
U1 = r[0].U[1]
U2 = r[0].U[2]
X.append(XX)
all_U0.append(U0)
all_U2.append(U2)
if subject < 17:
yy = data['y'].ravel()
y.append(yy)
else:
ids = data['Id']
ids_test.append(ids)
label_count += [subject] * len(XX)
all_U0 = np.vstack(all_U0)
all_U2 = np.vstack(all_U2)
X = np.vstack(X)
y = np.concatenate(y)
cv = StratifiedShuffleSplit(yy, n_iter=50, test_size=.1)
selection_pipe = Pipeline([('scaler', StandardScaler()),
('estimator', LassoLarsCV(cv=cv))])
selection_pipe.fit(all_U0[:y.shape[0]], y * 2 - 1)
weights = selection_pipe.steps[1][1].coef_
selected = np.where(weights != 0)[0]
proj = all_U2[:, selected].T
ids_test = np.concatenate(ids_test)
from IPython import embed; embed()
return np.dot(X, proj), y, ids_test, label_count
def decompose_tensor(filters):
""" filters is of type input feature maps, output feature maps, wxh of filter
Output is a structure P which contains lambda, U{1}, U{2}, U{3}
"""
# Set logging to DEBUG to see CP-ALS information
logging.basicConfig(level=logging.DEBUG)
print filters.shape
filters = np.array(filters)
print filters.shape
print filters.dtype
nbr_filters = filters.shape[0]
fwidth = filters.shape[2]
fheight = filters.shape[3]
Pstruct = []
for chanel in range(filters.shape[1]):
filter_for_channel = filters[:,chanel,:,:]
filter_for_channel.reshape(nbr_filters, fwidth, fheight)
filter_for_channel = np.swapaxes(filter_for_channel, 0,2);
print 'Number of filters ', nbr_filters
print 'filter_for channel shape ', filter_for_channel.shape
fig, axes = plt.subplots(nrows=5, ncols=4)
fig.tight_layout()
for f in xrange(nbr_filters):
img = filter_for_channel[:,:,f]
plt.subplot(5,4,f)
plt.imshow(img)
plt.show(block=False)
T = dtensor(filter_for_channel);
rank = np.floor(nbr_filters*0.6);
print 'rank is ', rank
session = pymatlab.session_factory()
session.putvalue('A',rank)
del session
## P.U, P.lmbda
print 'P U0,U1,U2, lambda sizes: ', P.U[0].size, P.U[1].size, P.U[2].size, P.lmbda
print 'fit was ', fit
Pstruct.append(P)
#dtensor(ktensor(U).toarray())
print np.allclose(T, P.totensor())
U = [np.random.rand(i,3) for i in (20, 10, 14)]
Tn = dtensor(ktensor(U).toarray())
P, fit, itr, _ = cp_als(Tn, 10)
print 'P U0,U1,U2, lambda sizes: ', P.U[0].size, P.U[1].size, P.U[2].size, P.lmbda
print 'fit was ', fit
print np.allclose(Tn, P.totensor())
return Pstruct
def approx_CP_R(value, R):
if value.ndim < 2:
return value
T = dtensor(value)
P, fit, itr, exetimes = cp_als(T, R, init='random')
Y = None
for i in range(R):
y = P.lmbda[i]
o = None
for l in range(T.ndim):
o = P.U[l][:,i] if o is None else np.outer(o, P.U[l][:,i])
y = y * o
Y = y if Y is None else Y+y
return Y
def corcondia(tensor, components=1):
k = components
X = tensor
X_approx_ks, fit, itr, exectimes = cp_als(X, k, init='random')
A = X_approx_ks.U[0]
B = X_approx_ks.U[1]
C = X_approx_ks.U[2]
x = X_approx_ks.totensor()
Ua, Sa, Va = np.linalg.svd(A)
Ub, Sb, Vb = np.linalg.svd(B)
Uc, Sc, Vc = np.linalg.svd(C)
SaI = np.zeros((Ua.shape[0], Va.shape[0]), float)
np.fill_diagonal(SaI, Sa)
SbI = np.zeros((Ub.shape[0], Vb.shape[0]), float)
np.fill_diagonal(SbI, Sb)
ScI = np.zeros((Uc.shape[0], Vc.shape[0]), float)
np.fill_diagonal(ScI, Sc)
SaI = np.linalg.pinv(SaI)
SbI = np.linalg.pinv(SbI)
ScI = np.linalg.pinv(ScI)
y = kronecker([Ua.transpose(), Ub.transpose(), Uc.transpose()], x)
z = kronecker([SaI, SbI, ScI], y)
G = kronecker([Va.transpose(), Vb.transpose(), Vc.transpose()], z)
# print G
C = np.full((k, k, k), 0)
for i in range(k):
for j in range(k):
for l in range(k):
if i == j == l:
C[i][j][l] = 1
c = 0
for i in range(k):
for j in range(k):
for l in range(k):
c += float(G[i][j][l] - C[i][j][l])**2.0
cc = 100 * (1 - (c / float(k)))
return round(cc)
def get_cp_factors(self):
if '__getitem__' in dir(self.layer):
f_cout_old, f_cin_old, f_h_old, f_w_old = self.weight
f_cout_old = np.array(f_cout_old)
f_cin_old = np.array(f_cin_old)
f_h_old = np.array(f_h_old)
f_w_old = np.array(f_w_old)
bias = self.bias
f_cout, f_cin, f_h, f_w = recompress_ncpd_tensor(
[f_cout_old, f_cin_old, f_h_old, f_w_old],
new_rank=self.rank,
max_cycle=500,
return_fit=False,
tensor_format='cpd')
else:
if self.weight.is_cuda:
self.weight = self.weight.cpu()
if self.bias is not None:
self.bias = self.bias.cpu()
weights = dtensor(self.weight)
bias = self.bias
T = dtensor(weights)
P, fit, itr, exectimes = cp_als(T, self.rank, init='random')
f_w = (np.array(P.U[3]) * (P.lmbda))
f_h = np.array(P.U[2])
f_cin = np.array(P.U[1])
f_cout = np.array(P.U[0])
f_h = torch.FloatTensor(
np.reshape(f_h.T,
(self.rank, 1, self.kernel_size[0], 1))).contiguous()
f_w = torch.FloatTensor(
np.reshape(f_w.T,
[self.rank, 1, 1, self.kernel_size[1]])).contiguous()
f_cin = torch.FloatTensor(
np.reshape(f_cin.T, [self.rank, self.cin, 1, 1])).contiguous()
f_cout = torch.FloatTensor(
np.reshape(f_cout, [self.cout, self.rank, 1, 1])).contiguous()
return [f_cin, f_h, f_w, f_cout], [None, None, None, bias]
def prepare_models(LAYER, R, NET_PATH, NET_NAME, INPUT_DIM):
PATH = NET_PATH
NET_PREFIX = PATH + NET_NAME
input_dim = INPUT_DIM
model = load_model(NET_PREFIX + '.prototxt')
ind = find_layer_by_name(model, LAYER)
new_model = accelerate_model(model, ind, R)
save_model(new_model, NET_PREFIX + '_accelerated.prototxt')
new_deploy = create_deploy_model(new_model, input_dim)
save_model(new_deploy, NET_PREFIX + '_accelerated_deploy.prototxt')
deploy = create_deploy_model(model, input_dim)
save_model(deploy, NET_PREFIX + '_deploy.prototxt')
net = caffe.Classifier(NET_PREFIX + '_deploy.prototxt',
NET_PREFIX + '.caffemodel')
fast_net = caffe.Classifier(NET_PREFIX + '_accelerated_deploy.prototxt',
NET_PREFIX + '.caffemodel')
l = ind - 1 #layer index in deploy version
weights = net.layers[l].blobs[0].data
bias = net.layers[l].blobs[1]
T = dtensor(weights)
P, fit, itr, exectimes = cp_als(T, R, init='random')
f_x = (np.array(P.U[3]) * (P.lmbda)).T
f_y = np.array(P.U[2]).T
f_c = np.array(P.U[1]).T
f_n = np.array(P.U[0])
n = model.layer[ind].convolution_param.num_output
d = model.layer[ind].convolution_param.kernel_size[0]
c = weights.shape[1]
f_y = np.reshape(f_y, [R, 1, d, 1])
f_x = np.reshape(f_x, [R, 1, 1, d])
f_c = np.reshape(f_c, [R, c, 1, 1])
f_n = np.reshape(f_n, [n, R, 1, 1])
np.copyto(fast_net.layers[l].blobs[0].data, f_c)
np.copyto(fast_net.layers[l + 1].blobs[0].data, f_y)
np.copyto(fast_net.layers[l + 2].blobs[0].data, f_x)
np.copyto(fast_net.layers[l + 3].blobs[0].data, f_n)
np.copyto(fast_net.layers[l + 3].blobs[1].data, bias.data)
fast_net.save(NET_PREFIX + '_accelerated.caffemodel')
def claculateCP(self,T,rank):
logging.warning("claculateCP")
P, fit, itr, exectimes = cp_als (T, rank, init='random', fit_method='full')
logging.warning("*****************Tensor*****************************")
logging.warning(T.shape)
logging.warning("******************** U1 represent the subjects *************************")
logging.warning( P.U[0].shape)
#print(P.U[0])
logging.warning("********************U2 represent the objects **************************")
logging.warning(P.U[1].shape)
#print(P.U[1])
logging.warning("********************U3 represent the predicats **************************")
logging.warning( P.U[2].shape)
#print(P.U[2])
logging.warning("********************L represent the subjects **************************")
#print(P.lmbda)
logging.warning("Done with "+itr.__str__()+" iterations")
return P
def cp_conv(weights, img, iter, w_shp, size, N_i, N_j, rank):
# Define parameters
(F, C, X, Y) = w_shp
(N, C, H, W) = size
# Instantiate 4D tensor for input
input = T.tensor4(name='input')
# Initialize shared variable for weights.
weights = weights.eval()
W_tensor = dtensor(weights)
# Apply CP-Decomposition on the clustered weight tensor
P, fit, itr, exectimes = cp_als(W_tensor, rank, init='random')
output = None
for k in range(rank):
T_F = theano.shared(np.reshape(P.U[0][:,k], (F, 1, 1, 1)), name='F_{0}'.format(k))
T_C = theano.shared(np.reshape(P.U[1][:,k], (1, C, 1, 1)), name='C_{0}'.format(k))
T_X = theano.shared(np.reshape(P.U[2][:,k], (1, 1, X, 1)), name='X_{0}'.format(k))
T_Y = theano.shared(np.reshape(P.U[3][:,k], (1, 1, 1, Y)), name='Y_{0}'.format(k))
# Apply convolution on each dimension individually
conv_C = conv.conv2d(input, T_C)
conv_X = conv.conv2d(conv_C, T_X)
conv_Y = conv.conv2d(conv_X, T_Y)
conv_F = conv.conv2d(conv_Y, T_F)
output = output + conv_F if output else conv_F
# Map Theano function
f = theano.function([input], output, profile=False)
# Execute Theano function
times = []
for i in range(iter):
start = time.time()
filtered_img = f(img)
done = time.time()
times.append(done-start)
avg1 = np.mean(times)
return filtered_img, avg1
def prepare_models(LAYER, R, NET_PATH, NET_NAME, INPUT_DIM):
PATH = NET_PATH
NET_PREFIX = PATH + NET_NAME
input_dim = INPUT_DIM
model = load_model(NET_PREFIX + '.prototxt')
ind = find_layer_by_name(model, LAYER)
new_model = accelerate_model(model, ind, R)
save_model(new_model, NET_PREFIX + '_accelerated.prototxt')
new_deploy = create_deploy_model(new_model, input_dim)
save_model(new_deploy, NET_PREFIX + '_accelerated_deploy.prototxt')
deploy = create_deploy_model(model, input_dim)
save_model(deploy, NET_PREFIX + '_deploy.prototxt')
net = caffe.Classifier(NET_PREFIX + '_deploy.prototxt', NET_PREFIX + '.caffemodel')
fast_net = caffe.Classifier(NET_PREFIX + '_accelerated_deploy.prototxt', NET_PREFIX + '.caffemodel')
l = ind - 1#layer index in deploy version
weights = net.layers[l].blobs[0].data
bias = net.layers[l].blobs[1]
T = dtensor(weights)
P, fit, itr, exectimes = cp_als(T, R, init='random')
f_x = (np.array(P.U[3])*(P.lmbda)).T
f_y = np.array(P.U[2]).T
f_c = np.array(P.U[1]).T
f_n = np.array(P.U[0])
n = model.layer[ind].convolution_param.num_output # OUTPUT
d = model.layer[ind].convolution_param.kernel_size[0] # KERNEL SIZE
c = weights.shape[1] # INPUT
f_c = np.reshape(f_c, [R, c, 1, 1]) # 1
f_y = np.reshape(f_y, [R, 1, d, 1]) # 2
f_x = np.reshape(f_x, [R, 1, 1, d]) # 3
f_n = np.reshape(f_n, [n, R, 1, 1]) # 4
np.copyto(fast_net.layers[l].blobs[0].data, f_c)
np.copyto(fast_net.layers[l+1].blobs[0].data, f_y)
np.copyto(fast_net.layers[l+2].blobs[0].data, f_x)
np.copyto(fast_net.layers[l+3].blobs[0].data, f_n)
np.copyto(fast_net.layers[l+3].blobs[1].data, bias.data)
fast_net.save(NET_PREFIX + '_accelerated.caffemodel')
def tcca(self, Views, var_matrix, cov_ten, reduce_to_dim):
var_matrix_inverse = list()
for v in range(self.number_of_views):
var_matrix_inverse.append(
self.root_inverse(var_matrix[v]) +
np.eye(var_matrix[v].shape[0]))
M_ten = self.ttm(cov_ten, var_matrix_inverse)
M_ten = dtensor(M_ten)
P, fit, itr = cp_als(M_ten, reduce_to_dim, max_iter=self.max_iter)
H = list()
for v in range(self.number_of_views):
H.append(np.dot(var_matrix_inverse[v], P.U[v]))
self.H = H
return H
def cp_decomposition(weights, bias, rank):
# Define parameters
(F, C, X, Y) = weights.shape
# Initialize shared variable for weights.
W_tensor = dtensor(weights)
# Apply CP-Decomposition on the clustered weight tensor
P, fit, itr, exectimes = cp_als(W_tensor, rank, init='random')
output = []
for k in range(rank):
T_F = np.reshape(P.U[0][:,k], (F, 1, 1, 1))
T_C = np.reshape(P.U[1][:,k], (1, C, 1, 1))
T_X = np.reshape(P.U[2][:,k], (1, 1, X, 1))
T_Y = np.reshape(P.U[3][:,k], (1, 1, 1, Y))
output.append([T_C, T_X, T_Y, T_F, bias])
return output
def get_basis(labels, data, n_components):
# Generate decompositions
print("Performing tensor decomposition of training data.")
all_basis = []
for n in np.unique(labels):
idx = np.where(labels == n)[0]
X = data[idx]
grad1, grad2, mag1 = separate_grad_mag(X)
grad = np.concatenate((grad1, grad2))
# Magnetometers look real rough
for idx in [grad, mag1]:
Xi = X[:, idx, :]
r = cp_als(dtensor(Xi), n_components, init="nvecs")
r_good_idx = drop_fifty_and_ten_hz(r[0].U[2])
basis = r[0].U[2][:, r_good_idx]
all_basis.append(basis)
basis = np.hstack(all_basis)
del all_basis
return basis
def get_UDV_decomposition(W, method='svd'):
# current implementation is svd
c_out, khkw, c_in = W.shape
method = method.lower()
with torch.no_grad():
if method == 'svd':
m_W = W.mean(1)
U, _, V = torch.svd(m_W)
D = []
for r in range(khkw):
W_r = W[:, r, :] # c_out * c_in ->
c = min(c_out, c_in)
D_w = torch.diag(U.t() @ W_r @ V).view(c, 1)
D.append(D_w)
S = torch.stack(D, dim=1)
elif method == 'svd_avg':
pass
elif method == 'als':
# m = min(c_out, c_in)
# U: c_out * m
# S: k^2 * m
# V: c_in * m
rank = min(c_out, c_in)
tic = time.clock()
T = dtensor(W.data.cpu().numpy())
P, fit, itr, exectimes = cp_als(T, rank, init='random')
U = np.array(P.U[0]) # c_out * rank
S = np.array(P.U[1]).T # k^2 * rank --> rank * k^2
V = np.array(P.U[2] * P.lmbda) # c_in * rank
print('CP decomposition done. It cost %.5f secs. fit: %f' % (time.clock() - tic, fit[0]))
V = torch.FloatTensor(V).cuda()
S = torch.FloatTensor(S).cuda()
U = torch.FloatTensor(U).cuda()
else:
raise NotImplementedError("Method {} not supported!".format(method))
return U, S, V
def valid_3ord():
print('\nCorrectness benchmark for 3 order tensor CP decomposition.\n')
shape = (2, 3, 4)
max_iter = 30
print('----------TensorNP----------')
norm_errors = 0
for _ in range(max_iter):
tensor = tnp.randn(2, 3, 4)
factors, lamda = tnp.cp(tensor, r=3, stop_iter=500, tol=1e-5, normalize_factor=True)
rec_tensor = tnp.reconstruct_cp(factors, lamda, shape)
norm_error = tnp.linalg.norm(rec_tensor - tensor) / tnp.linalg.norm(tensor)
norm_errors += norm_error
print(f'error ({norm_errors/max_iter})')
print('----------scikit-tensor----------')
norm_errors = 0
for _ in range(max_iter):
tensor = tnp.randn(2, 3, 4)
skt_tensor = skt.dtensor(tensor)
P, _, _ = skt.cp_als(skt_tensor, rank=3, init='random')
rec_tensor = P.toarray()
norm_error = tnp.linalg.norm(rec_tensor - skt_tensor) / tnp.linalg.norm(tensor)
norm_errors += norm_error
print(f'error ({norm_errors/max_iter})')
print('----------Tensorly----------')
norm_errors = 0
for _ in range(max_iter):
tensor = tnp.randn(2, 3, 4)
tl_tensor = tl.tensor(tensor)
cp_tensor = tensorly.decomposition.parafac(
tl_tensor, rank=3, n_iter_max=500, tol=1e-6, normalize_factors=True, init='random')
rec_tensor = tnp.reconstruct_cp(cp_tensor.factors, cp_tensor.weights, shape)
norm_error = tnp.linalg.norm(rec_tensor - tensor) / tnp.linalg.norm(tensor)
norm_errors += norm_error
print(f'error ({norm_errors/max_iter})')
#!/usr/bin/env python
import logging
from scipy.io.matlab import loadmat
from sktensor import dtensor, cp_als
# Set logging to DEBUG to see CP-ALS information
logging.basicConfig(level=logging.DEBUG)
# Load Matlab data and convert it to dense tensor format
mat = loadmat('../data/sensory-bread/brod.mat')
T = dtensor(mat['X'])
# Decompose tensor using CP-ALS
P, fit, itr, exectimes = cp_als(T, 3, init='random')
all_scores = []
for subject in train_subject_names:
print subject
f = loadmat(data_dir / subject)
X = f['X'][:, 160:, 125:250]
y = f['y'].ravel() * 2 - 1
cv = StratifiedShuffleSplit(y, n_iter=50, test_size=.1)
pipeline = Pipeline([('scaler', StandardScaler()),
('estimator', LassoLarsCV(cv=cv))])
T = dtensor(X)
r = cp_als(T, rank=10)
sample_axes = r[0].U[0]
pipeline.fit(sample_axes, y)
weights = pipeline.steps[1][1].coef_
all_weights.append(weights)
selected = np.where(weights != 0)[0]
all_selected_timecourses.append(r[0].U[2][:, selected])
pipeline_global_eval = Pipeline([
('scaler', StandardScaler()),
('estimator', LogisticRegression(C=1., penalty="l1"))])
global_scores = cross_val_score(pipeline_global_eval,
sample_axes, y, cv=cv,
def task1cFunc(userid):
# read mltags, mlrating, mlmovies, movie-actor
mltagsFile = pd.read_csv('mltags.csv')
mlratingsFile = pd.read_csv('mlratings.csv')
genomeFile = pd.read_csv('genome-tags.csv')
movieFile = pd.read_csv('smallmlmovies.csv')
# Exgtract tag from tagid
genomeFile['tagid'] = genomeFile['tagId']
del genomeFile['tagId']
mltagsFile = pd.merge(mltagsFile, genomeFile, on='tagid')
s = movieFile["genres"].str.split('|', expand=True).stack()
i = s.index.get_level_values(0)
movieFile = movieFile.loc[i].copy()
movieFile["genres"] = s.values
# Extract movie from movieid
del movieFile['year']
mlratingsFile = pd.merge(mlratingsFile, movieFile, on='movieid')
mltagsFile = pd.merge(mltagsFile, movieFile, on='movieid')
mltagsFileUser = mltagsFile.loc[mltagsFile['userid'] == userid]
tagUserMovies = mltagsFileUser['moviename'].values
mlratingsFileUser = mlratingsFile.loc[mlratingsFile['userid'] == userid]
ratingUserMovies = mlratingsFileUser['moviename'].values
tagRatingUserMovies = list(set(tagUserMovies) | set(ratingUserMovies))
mltagsFileUser['timestamp'] = pd.to_datetime(mltagsFileUser['timestamp'])
mltagsFileUser['timestamp'] = (mltagsFileUser['timestamp'] -
dt.datetime(1970, 1, 1)).dt.total_seconds()
mltagsFileUser['timestamp'] = \
((mltagsFileUser['timestamp'] - mltagsFileUser['timestamp'].min()) / (mltagsFileUser['timestamp'].max() - mltagsFileUser['timestamp'].min()+1))+1
mlratingsFileUser['timestamp'] = pd.to_datetime(
mlratingsFileUser['timestamp'])
mlratingsFileUser['timestamp'] = (
mlratingsFileUser['timestamp'] -
dt.datetime(1970, 1, 1)).dt.total_seconds()
mlratingsFileUser['timestamp'] = \
((mlratingsFileUser['timestamp'] - mlratingsFileUser['timestamp'].min()) / (mlratingsFileUser['timestamp'].max() - mlratingsFileUser['timestamp'].min()+1))+1
commonTagRating = list(set(tagUserMovies) & set(ratingUserMovies))
uncommonTag = list(set(tagUserMovies) ^ set(commonTagRating))
uncommonRating = list(set(ratingUserMovies) ^ set(commonTagRating))
timeWeights = {}
for i in range(len(commonTagRating)):
tag = mltagsFileUser.loc[mltagsFileUser['moviename'] ==
commonTagRating[i]]['timestamp'].values[0]
rating = mlratingsFileUser.loc[
mlratingsFileUser['moviename'] ==
commonTagRating[i]]['timestamp'].values[0]
if tag > rating:
timeWeights[commonTagRating[i]] = tag
else:
timeWeights[commonTagRating[i]] = rating
for i in range(len(uncommonRating)):
rating = mlratingsFileUser.loc[
mlratingsFileUser['moviename'] ==
uncommonRating[i]]['timestamp'].values[0]
timeWeights[uncommonRating[i]] = rating
for i in range(len(uncommonTag)):
tag = mltagsFileUser.loc[mltagsFileUser['moviename'] ==
uncommonTag[i]]['timestamp'].values[0]
timeWeights[uncommonTag[i]] = tag
#
# deleting columns that are not required
del mlratingsFile['timestamp']
del mlratingsFile['imdbid']
del mlratingsFile['userid']
del mltagsFile['timestamp']
del mltagsFile['userid']
del mltagsFile['tagid']
# creating a dictionary with movieid as key and a list of all tags associated with tha movie and removing duplicates
movieGenreDict = {
k: g['genres'].tolist()
for k, g in movieFile.groupby('moviename')
}
movieGenreDict = {k: list(set(j)) for k, j in movieGenreDict.items()}
# creating a dictionary with movieid as key and a list of all ratings given by a user for that particular movie and removing duplicates
movieRatingDict = {
k: g['rating'].tolist()
for k, g in mlratingsFile.groupby('moviename')
}
movieRatingDict = {k: list(set(j)) for k, j in movieRatingDict.items()}
# computing the average rating for all movies and storing in a dictionary
avgRating = mlratingsFile.groupby('moviename').mean().reset_index()
avgRatingDict = {
k: g['rating'].tolist()
for k, g in avgRating.groupby('moviename')
}
# List of unique movies, genres and ratings
movieList = mlratingsFile.moviename.unique()
movieList = np.asarray(movieList)
movieListDict = dict(enumerate(movieList))
genreList = movieFile.genres.unique()
genreList = np.asarray(genreList)
genreListDict = dict(enumerate(genreList))
ratingList = mlratingsFile.rating.unique()
ratingList = np.asarray(ratingList)
ratingListDict = dict(enumerate(ratingList))
movieListDictInverse = invertDictionary(movieListDict)
genreListDictInverse = invertDictionary(genreListDict)
ratingListDictInverse = invertDictionary(ratingListDict)
movieNotWatched = list(set(movieList) ^ set(tagRatingUserMovies))
# declaring a tensor with three modes - with movie, tags and ratings
T = np.zeros((movieList.shape[0], genreList.shape[0], ratingList.shape[0]))
arrayofvalues = []
for i in movieList:
if i in movieRatingDict:
if i in movieGenreDict:
movieTags = movieGenreDict[i]
rList = movieRatingDict[i]
for j in movieTags:
for k in rList:
mIndex = movieListDictInverse[i]
gIndex = genreListDictInverse[j]
rIndex = ratingListDictInverse[k]
avgRatingValue = avgRatingDict[i][0]
if k >= avgRatingValue:
T[mIndex, gIndex, rIndex] = 1
arrayofvalues.append([mIndex, gIndex, rIndex])
else:
T[mIndex, gIndex, rIndex] = 0
# building the tensor using sktensor
tensor = dtensor(T)
# applying CP-decomposition with ALS(Alternating Least Squares)
U, fit, itr, exectimes, P = cp_als(tensor, 5, init='random')
latent_semantics_movie = pd.DataFrame(
columns=['movie', 'ls1', 'ls2', 'ls3', 'ls4', 'ls5'])
latent_semantics_movie['movie'] = movieList
latent_semantics_movie['ls1'] = U[0][:, 0]
latent_semantics_movie['ls2'] = U[0][:, 1]
latent_semantics_movie['ls3'] = U[0][:, 2]
latent_semantics_movie['ls4'] = U[0][:, 3]
latent_semantics_movie['ls5'] = U[0][:, 4]
x = latent_semantics_movie.loc[latent_semantics_movie['movie'].isin(
tagRatingUserMovies)].values
for i in range(len(x)):
for j in range(1, len(x[0])):
x[i][j] = x[i][j] * timeWeights.get(x[i][0])
y = latent_semantics_movie.loc[latent_semantics_movie['movie'].isin(
movieNotWatched)].values
cossim = cosine_similarity(x[:, 1:], y[:, 1:])
simDF = pd.DataFrame(cossim,
index=tagRatingUserMovies,
columns=movieNotWatched)
simDF.to_csv('cos.csv')
temp = simDF.values.tolist()
sorted_movies_for_each_watched_movieDict = []
for i in range(len(temp)):
sorted_movies_for_each_watched_movie = np.argsort(temp[i])
sorted_movies_for_each_watched_movieDict.append(
sorted_movies_for_each_watched_movie.tolist()[:10])
sortedMoviesRavel = [
item for sublist in sorted_movies_for_each_watched_movieDict
for item in sublist
]
freq = {}
for i in range(len(sorted_movies_for_each_watched_movieDict)):
for j in range(len(sorted_movies_for_each_watched_movieDict[0])):
freq[sorted_movies_for_each_watched_movieDict[i][j]] = 0
for i in range(len(sorted_movies_for_each_watched_movieDict)):
for j in range(len(sorted_movies_for_each_watched_movieDict[0])):
freq[sorted_movies_for_each_watched_movieDict[i][j]] += (10 - j)
freq = OrderedDict(sorted(freq.items(), reverse=True, key=lambda t: t[1]))
freq = freq.items()
recommendedMovies = []
for i in range(5):
index = freq[i][0]
recommendedMovies.append(y[index][0])
relevant = []
notRelevant = []
choice = 'y'
while choice != 'n':
rel_dict = {}
selected_dict = {}
N = 5
R = 0
for i in range(len(recommendedMovies)):
print "If ", recommendedMovies[
i], " is relevant, enter 1. If it is not relevant, enter 0"
relevant.append(int(raw_input()))
rel_dict[recommendedMovies[i]] = relevant[i]
if relevant[i] == 1:
R = R + 1
else:
notRelevant.append(recommendedMovies[i])
genreset = set()
for movie in recommendedMovies:
genres_list = movieGenreDict[movie]
selected_dict[movie] = genres_list
genreset = genreset.union(set(genres_list))
genreTop5 = list(genreset)
ri = []
ni = []
for i in range(0, len(genreTop5)):
ri.append(0)
ni.append(0)
for m in recommendedMovies:
for i in range(0, len(genreTop5)):
l1 = selected_dict[m]
rval = rel_dict[m]
if genreTop5[i] in l1:
ni[i] = ni[i] + 1
if rval == 1:
ri[i] = ri[i] + 1
pr_feedback = {}
for i in range(0, len(genreTop5)):
try:
numerator = ri[i] / (R - ri[i])
denominator = (ni[i] - ri[i]) / (N - R - ni[i] + ri[i])
pr = math.log((numerator / denominator), 2)
except:
numerator = (ri[i] + 0.5) / (R - ri[i] + 1)
denominator = (ni[i] - ri[i] + 0.5) / (N - R - ni[i] + ri[i] +
1)
pr = math.log((numerator / denominator), 2)
pr_feedback[genreTop5[i]] = pr
for key, value in pr_feedback.iteritems():
pr_feedback[key] = (pr_feedback[key] - min(
pr_feedback.values())) / max(pr_feedback.values())
pr_dict = {}
for i in movieList:
if i in movieRatingDict:
if i in movieGenreDict:
movieTags = movieGenreDict[i]
rList = movieRatingDict[i]
for j in movieTags:
for k in rList:
mIndex = movieListDictInverse[i]
tIndex = genreListDictInverse[j]
rIndex = ratingListDictInverse[k]
avgRatingValue = avgRatingDict[i][0]
if k >= avgRatingValue:
if j in genreTop5:
T[mIndex, tIndex, rIndex] *= pr_feedback[j]
tensor = dtensor(T)
# applying CP-decomposition with ALS(Alternating Least Squares)
U, fit, itr, exectimes, P = cp_als(tensor, 5, init='random')
latent_semantics_movie = pd.DataFrame(
columns=['movie', 'ls1', 'ls2', 'ls3', 'ls4', 'ls5'])
latent_semantics_movie['movie'] = movieList
latent_semantics_movie['ls1'] = U[0][:, 0]
latent_semantics_movie['ls2'] = U[0][:, 1]
latent_semantics_movie['ls3'] = U[0][:, 2]
latent_semantics_movie['ls4'] = U[0][:, 3]
latent_semantics_movie['ls5'] = U[0][:, 4]
x = latent_semantics_movie.loc[latent_semantics_movie['movie'].isin(
tagRatingUserMovies)].values
for i in range(len(x)):
for j in range(1, len(x[0])):
x[i][j] = x[i][j] * timeWeights.get(x[i][0])
y = latent_semantics_movie.loc[latent_semantics_movie['movie'].isin(
movieNotWatched)].values
cossim = cosine_similarity(x[:, 1:], y[:, 1:])
simDF = pd.DataFrame(cossim,
index=tagRatingUserMovies,
columns=movieNotWatched)
temp = simDF.values.tolist()
sorted_movies_for_each_watched_movieDict = []
for i in range(len(temp)):
sorted_movies_for_each_watched_movie = np.argsort(temp[i])
sorted_movies_for_each_watched_movieDict.append(
sorted_movies_for_each_watched_movie.tolist()[:10])
sortedMoviesRavel = [
item for sublist in sorted_movies_for_each_watched_movieDict
for item in sublist
]
freq = {}
for i in range(len(sorted_movies_for_each_watched_movieDict)):
for j in range(len(sorted_movies_for_each_watched_movieDict[0])):
freq[sorted_movies_for_each_watched_movieDict[i][j]] = 0
for i in range(len(sorted_movies_for_each_watched_movieDict)):
for j in range(len(sorted_movies_for_each_watched_movieDict[0])):
freq[sorted_movies_for_each_watched_movieDict[i][j]] += (10 -
j)
freq = OrderedDict(
sorted(freq.items(), reverse=True, key=lambda t: t[1]))
freq = freq.items()
recommendedMovies = []
for i in range(5):
index = freq[i][0]
recommendedMovies.append(y[index][0])
print recommendedMovies
relevant = []
print('Do you want to continue? Enter Y for yes and N for No')
choice = raw_input()
while choice not in ['y', 'n']:
print('invalid input')
choice = input()
def task2c(recommendedMovies, T):
relevant = []
notRelevant = []
choice = 'y'
while choice != 'n':
rel_dict = {}
selected_dict = {}
N = 5
R = 0
for i in range(len(recommendedMovies)):
print "If ", recommendedMovies[
i], " is relevant, enter 1. If it is not relevant, enter 0"
relevant.append(int(raw_input()))
rel_dict[recommendedMovies[i]] = relevant[i]
if relevant[i] == 1:
R = R + 1
else:
notRelevant.append(recommendedMovies[i])
genreset = set()
for movie in recommendedMovies:
genres_list = movieGenreDict[movie]
selected_dict[movie] = genres_list
genreset = genreset.union(set(genres_list))
genreTop5 = list(genreset)
ri = []
ni = []
for i in range(0, len(genreTop5)):
ri.append(0)
ni.append(0)
for m in recommendedMovies:
for i in range(0, len(genreTop5)):
l1 = selected_dict[m]
rval = rel_dict[m]
if genreTop5[i] in l1:
ni[i] = ni[i] + 1
if rval == 1:
ri[i] = ri[i] + 1
pr_feedback = {}
for i in range(0, len(genreTop5)):
try:
numerator = ri[i] / (R - ri[i])
denominator = (ni[i] - ri[i]) / (N - R - ni[i] + ri[i])
pr = math.log((numerator / denominator), 2)
except:
numerator = (ri[i] + 0.5) / (R - ri[i] + 1)
denominator = (ni[i] - ri[i] + 0.5) / (N - R - ni[i] + ri[i] +
1)
pr = math.log((numerator / denominator), 2)
pr_feedback[genreTop5[i]] = pr
for key, value in pr_feedback.iteritems():
pr_feedback[key] = (pr_feedback[key] - min(
pr_feedback.values())) / max(pr_feedback.values())
pr_dict = {}
for i in movieList:
if i in movieRatingDict:
if i in movieGenreDict:
movieTags = movieGenreDict[i]
rList = movieRatingDict[i]
for j in movieTags:
for k in rList:
mIndex = movieListDictInverse[i]
tIndex = genreListDictInverse[j]
rIndex = ratingListDictInverse[k]
avgRatingValue = avgRatingDict[i][0]
if k >= avgRatingValue:
if j in genreTop5:
T[mIndex, tIndex, rIndex] *= pr_feedback[j]
tensor = dtensor(T)
# applying CP-decomposition with ALS(Alternating Least Squares)
U, fit, itr, exectimes, P = cp_als(tensor, 5, init='random')
latent_semantics_movie = pd.DataFrame(
columns=['movie', 'ls1', 'ls2', 'ls3', 'ls4', 'ls5'])
latent_semantics_movie['movie'] = movieList
latent_semantics_movie['ls1'] = U[0][:, 0]
latent_semantics_movie['ls2'] = U[0][:, 1]
latent_semantics_movie['ls3'] = U[0][:, 2]
latent_semantics_movie['ls4'] = U[0][:, 3]
latent_semantics_movie['ls5'] = U[0][:, 4]
x = latent_semantics_movie.loc[latent_semantics_movie['movie'].isin(
tagRatingUserMovies)].values
for i in range(len(x)):
for j in range(1, len(x[0])):
x[i][j] = x[i][j] * timeWeights.get(x[i][0])
y = latent_semantics_movie.loc[latent_semantics_movie['movie'].isin(
movieNotWatched)].values
cossim = cosine_similarity(x[:, 1:], y[:, 1:])
simDF = pd.DataFrame(cossim,
index=tagRatingUserMovies,
columns=movieNotWatched)
temp = simDF.values.tolist()
sorted_movies_for_each_watched_movieDict = []
for i in range(len(temp)):
sorted_movies_for_each_watched_movie = np.argsort(temp[i])
sorted_movies_for_each_watched_movieDict.append(
sorted_movies_for_each_watched_movie.tolist()[:10])
sortedMoviesRavel = [
item for sublist in sorted_movies_for_each_watched_movieDict
for item in sublist
]
freq = {}
for i in range(len(sorted_movies_for_each_watched_movieDict)):
for j in range(len(sorted_movies_for_each_watched_movieDict[0])):
freq[sorted_movies_for_each_watched_movieDict[i][j]] = 0
for i in range(len(sorted_movies_for_each_watched_movieDict)):
for j in range(len(sorted_movies_for_each_watched_movieDict[0])):
freq[sorted_movies_for_each_watched_movieDict[i][j]] += (10 -
j)
freq = OrderedDict(
sorted(freq.items(), reverse=True, key=lambda t: t[1]))
freq = freq.items()
movieFreq = {}
for i in range(len(freq)):
index = freq[i][0]
movieFreq[y[index][0]] = freq[i][1]
recommendedMovies = []
for i in range(5):
index = freq[i][0]
recommendedMovies.append(y[index][0])
print recommendedMovies
relevant = []
print('Do you want to continue? Enter Y for yes and N for No')
choice = raw_input()
while choice not in ['y', 'n']:
print('invalid input')
choice = input()
def create_features(XX, tmin, tmax,
sfreq, tmin_original=-0.5,
perform_baseline_correction=True,
plot_name=""):
"""
Creation of the feature space.
- restricting the time window of MEG data to [tmin, tmax]sec.
- Concatenating the 306 timeseries of each trial in one long
vector.
- Normalizing each feature independently (z-scoring).
- optional: "baseline correction", a data centering concept often
used in M/EEG, will calculate a mean value per sensor
from pre-stimulus measurements, and subtract this from
the relevant measurement. Replaces centering based on
post-stimulus data
Returns a feature vector XX,
"""
print("Applying the desired time window and dropping sensors.")
lower_limit = 240
XX = XX[:, lower_limit:, :]
# instead of post-stimulus centering
baseline = XX[..., :125].mean(-1)
beginning = np.round((tmin - tmin_original) * sfreq).astype(np.int)
end = np.round((tmax - tmin_original) * sfreq).astype(np.int)
XX = XX[:, :, beginning:end].copy()
XX /= np.linalg.norm(XX, axis=2)[..., np.newaxis]
#Assuming 250Hz == fs, 125Hz == fs/2, 50Hz = 50/125 = .4
#5 Hz bw = 5/125 = .04
print("Applying notch filter for powerline.")
bw = .04
freq = .4
b, a = notch(freq, bw)
XX = sig.lfilter(b, a, XX)
#Assuming 250Hz == fs, 125Hz == fs/2, 50Hz = 10/125 = .08
#5 Hz bw = 5/125 = .04
print("Applying filter for alpha wave.")
bw = .04
freq = .08
b, a = notch(freq, bw)
XX = sig.lfilter(b, a, XX)
XX -= baseline[..., np.newaxis]
print("CP-ALS Decomposition.")
T = dtensor(XX)
P, fit, itr, exectimes = cp_als(T, 2, init='nvecs')
#P, fit, itr, exectimes = cp_als(T, 8, init='random')
proj = P.U[2]
fproj = np.abs(np.fft.fft(proj, axis=0))[:XX.shape[-1] // 2, :]
plt.figure()
plt.plot(proj)
plt.title(plot_name)
print("Projecting.")
XX = np.dot(XX, proj)
print("New shape is %sx%sx%s" % XX.shape)
print("2D Reshaping: concatenating all 306 timeseries.")
XX = XX.reshape(XX.shape[0], XX.shape[1] * XX.shape[2])
print("Features Normalization.")
XX -= XX.mean(0)
XX = np.nan_to_num(XX / XX.std(0))
return XX
def CPDecomposition(tensor, rank):
T = dtensor(tensor)
# Decompose tensor using CP-ALS
P, fit, itr, exectimes = cp_als(T, rank, init='random')
u = P.U
return u
def task1cFunc(userid):
arrayofvalues = []
for i in movieList:
if i in movieRatingDict:
if i in movieGenreDict:
movieTags = movieGenreDict[i]
rList = movieRatingDict[i]
for j in movieTags:
for k in rList:
mIndex = movieListDictInverse[i]
gIndex = genreListDictInverse[j]
rIndex = ratingListDictInverse[k]
avgRatingValue = avgRatingDict[i][0]
if k >= avgRatingValue:
T[mIndex, gIndex, rIndex] = 1
arrayofvalues.append([mIndex, gIndex, rIndex])
else:
T[mIndex, gIndex, rIndex] = 0
# building the tensor using sktensor
tensor = dtensor(T)
# applying CP-decomposition with ALS(Alternating Least Squares)
U, fit, itr, exectimes, P = cp_als(tensor, 5, init='random')
latent_semantics_movie = pd.DataFrame(
columns=['movie', 'ls1', 'ls2', 'ls3', 'ls4', 'ls5'])
latent_semantics_movie['movie'] = movieList
latent_semantics_movie['ls1'] = U[0][:, 0]
latent_semantics_movie['ls2'] = U[0][:, 1]
latent_semantics_movie['ls3'] = U[0][:, 2]
latent_semantics_movie['ls4'] = U[0][:, 3]
latent_semantics_movie['ls5'] = U[0][:, 4]
x = latent_semantics_movie.loc[latent_semantics_movie['movie'].isin(
tagRatingUserMovies)].values
for i in range(len(x)):
for j in range(1, len(x[0])):
x[i][j] = x[i][j] * timeWeights.get(x[i][0])
y = latent_semantics_movie.loc[latent_semantics_movie['movie'].isin(
movieNotWatched)].values
cossim = cosine_similarity(x[:, 1:], y[:, 1:])
simDF = pd.DataFrame(cossim,
index=tagRatingUserMovies,
columns=movieNotWatched)
simDF.to_csv('cos.csv')
temp = simDF.values.tolist()
sorted_movies_for_each_watched_movieDict = []
for i in range(len(temp)):
sorted_movies_for_each_watched_movie = np.argsort(temp[i])
sorted_movies_for_each_watched_movieDict.append(
sorted_movies_for_each_watched_movie.tolist()[:10])
sortedMoviesRavel = [
item for sublist in sorted_movies_for_each_watched_movieDict
for item in sublist
]
freq = {}
for i in range(len(sorted_movies_for_each_watched_movieDict)):
for j in range(len(sorted_movies_for_each_watched_movieDict[0])):
freq[sorted_movies_for_each_watched_movieDict[i][j]] = 0
for i in range(len(sorted_movies_for_each_watched_movieDict)):
for j in range(len(sorted_movies_for_each_watched_movieDict[0])):
freq[sorted_movies_for_each_watched_movieDict[i][j]] += (10 - j)
freq = OrderedDict(sorted(freq.items(), reverse=True, key=lambda t: t[1]))
freq = freq.items()
movieFreq = {}
for i in range(len(freq)):
index = freq[i][0]
movieFreq[y[index][0]] = freq[i][1]
recommendedMovies = []
for i in range(5):
index = freq[i][0]
recommendedMovies.append(y[index][0])
task2c(recommendedMovies, T)
def decompose_model(model_def_path, model_weights_path, layer_ranks):
""" CREATE DECOMPOSED MODEL DEFINITION """
with open(model_def_path) as f:
model_def = caffe.proto.caffe_pb2.NetParameter()
google.protobuf.text_format.Merge(f.read(), model_def)
new_model_def = caffe.proto.caffe_pb2.NetParameter()
new_model_def.name = model_def.name + '_decomposed'
if model_def.input:
new_model_def.input.extend(['data'])
new_model_def.input_dim.extend(model_def.input_dim)
new_layers = [
] #Keeping track of new layers helps renaming nodes in the future
for layer in model_def.layer:
if layer.name not in layer_ranks.keys() or layer.type != 'Convolution':
new_model_def.layer.extend([layer])
else:
decomposed_layer = decompose_layer(layer, layer_ranks[layer.name])
for i in range(4):
new_layers.append(decomposed_layer[i].name)
new_model_def.layer.extend(decomposed_layer)
#Rename bottom/top nodes for some layers !!!
layer_index = len(new_model_def.layer)
for i in range(layer_index):
#Label Decomposed layers nodes
if new_model_def.layer[i].name in new_layers:
if i == 0:
new_model_def.layer[i].bottom.extend(['data'])
elif new_model_def.layer[i - 1].type == 'ReLU':
new_model_def.layer[i].bottom.extend(
[new_model_def.layer[i - 2].name])
elif new_model_def.layer[i - 1].type in ['Convolution', 'Pooling']:
new_model_def.layer[i].bottom.extend(
[new_model_def.layer[i - 1].name])
new_model_def.layer[i].top.extend([new_model_def.layer[i].name])
#Rename Convolution layers nodes
elif new_model_def.layer[i].type == 'Convolution':
if new_model_def.layer[i - 2].name in new_layers:
new_model_def.layer[i].bottom[0] = new_model_def.layer[i -
2].name
#Rename ReLU layers nodes
elif new_model_def.layer[i].type == 'ReLU':
if new_model_def.layer[i - 1].name in new_layers:
new_model_def.layer[i].bottom[0] = new_model_def.layer[i -
1].name
new_model_def.layer[i].top[0] = new_model_def.layer[i - 1].name
#Rename Pooling layers nodes
elif new_model_def.layer[i].type == 'Pooling':
if new_model_def.layer[i - 2].name in new_layers:
new_model_def.layer[i].bottom[0] = new_model_def.layer[i -
2].name
new_model_def_path = model_def_path[:-9] + '_decomposed.prototxt'
with open(new_model_def_path, 'w') as f:
google.protobuf.text_format.PrintMessage(new_model_def, f)
""" CREATE DECOMPOSED MODEL WEIGHTS """
net = caffe.Net(model_def_path, model_weights_path, caffe.TEST)
decomposed_net = caffe.Net(new_model_def_path, model_weights_path,
caffe.TEST)
layers = net.params.items()
net_dict = {}
for name, params in layers:
net_dict[name] = {
'weights': params[0].data,
'bias': params[1].data,
}
log_dict = {} #Log containing fit, n_itr and exectimes for each layer
for conv_layer in layer_ranks.keys():
log_dict[conv_layer] = {}
rank = layer_ranks[conv_layer]
T = dtensor(net_dict[conv_layer]['weights'])
print('\nDecomposing %s...' % conv_layer)
P, fit, n_itr, exectimes = cp_als(T, rank, init='random')
print('Reconstruction: %f%%' % (fit * 100))
print('Elapsed time: %.2fs' % sum(exectimes))
log_dict[conv_layer]['fit'] = fit[0]
log_dict[conv_layer]['n_itr'] = n_itr
log_dict[conv_layer]['exectime'] = sum(exectimes)
num_output = net_dict[conv_layer]['weights'].shape[0]
channels = net_dict[conv_layer]['weights'].shape[1]
kernel_size = net_dict[conv_layer]['weights'].shape[2]
bias = net_dict[conv_layer]['bias']
P_x = (P.U[3] * P.lmbda).T
P_y = P.U[2].T
P_c = P.U[1].T
P_n = P.U[0]
P_x = np.reshape(P_x, [rank, 1, 1, kernel_size]).astype(np.float32)
P_y = np.reshape(P_y, [rank, 1, kernel_size, 1]).astype(np.float32)
P_c = np.reshape(P_c, [rank, channels, 1, 1]).astype(np.float32)
P_n = np.reshape(P_n, [num_output, rank, 1, 1]).astype(np.float32)
np.copyto(decomposed_net.params[conv_layer + '_x'][0].data, P_x)
np.copyto(decomposed_net.params[conv_layer + '_y'][0].data, P_y)
np.copyto(decomposed_net.params[conv_layer + '_c'][0].data, P_c)
np.copyto(decomposed_net.params[conv_layer + '_n'][0].data, P_n)
np.copyto(decomposed_net.params[conv_layer + '_n'][1].data, bias)
new_model_weights_path = model_weights_path[:-11] + '_decomposed.caffemodel'
decomposed_net.save(new_model_weights_path)
return log_dict, [new_model_def_path, new_model_weights_path]
import logging
import numpy as np
from scipy.io.matlab import loadmat
from sktensor import dtensor, cp_als
# Set logging to DEBUG to see CP-ALS information
logging.basicConfig(level=logging.DEBUG)
# Load Matlab data and convert it to dense tensor format
# mat = loadmat('/Users/monty/Julia/dNTF.py/brod.mat')
X = np.zeros((2, 10, 5))
X[:, 0:2, 0] = 1
X[:, 2:4, 1] = 1
X[:, 4:6, 2] = 1
X[:, 6:8, 3] = 1
X[:, 8:10, 4] = 1
T = dtensor(X)
# Decompose tensor using CP-ALS
P, fit, itr, exectimes = cp_als(T, 4, init='random')
for i in range(len(year)):
for j in range(len(movie)):
if ((merged['year'] == year[i]) &
(merged['moviename'] == movie[j])).any():
for k in range(len(actor)):
if ((merged['moviename'] == movie[j]) &
(merged['name'] == actor[k])).any():
T[i, j, k] = 1
else:
T[i, j, k] = 0
tensor = dtensor(T)
# Decompose tensor using CP-ALS
U, fit, itr, exectimes = cp_als(tensor, 5, init='random')
print U
# Latent Semantics for Year
latent_semantics_year = pd.DataFrame(
columns=['year', 'ls1', 'ls2', 'ls3', 'ls4', 'ls5'])
latent_semantics_year['year'] = year
latent_semantics_year['ls1'] = U[0][:, 0]
latent_semantics_year['ls2'] = U[0][:, 1]
latent_semantics_year['ls3'] = U[0][:, 2]
latent_semantics_year['ls4'] = U[0][:, 3]
latent_semantics_year['ls5'] = U[0][:, 4]
print 'Latent Semantic for Year sorted by LS1'
ls1 = latent_semantics_year.sort_values(by='ls1', ascending=False)
print ls1
for j in range(0, 192, 2):
X[i][j][:] = faketnsr[i][k][:]
X[i][j + 1][:] = realtnsr[i][k][:]
k = k + 1
print('K is:', k)
print(X[0].shape)
print('Densifying X tensor..')
T1 = dtensor(X)
print('Shape of tensor:', tf.shape(T1))
rnk = 45
print('Rank is:', rnk)
print(T1.shape[0], T1.shape[1], T1.shape[2])
print('CP decomposition for tensor..')
P1, fit1, itr1, exectimes1 = cp_als(T1, rnk, init='random')
print('End of 1st Decomposition')
X_train = P1.U[1]
print((P1.U[0]).shape, (P1.U[1]).shape, (P1.U[2]).shape)
print('Shape of decomposed array:', X_train.shape)
print('Creating label array, 1 means fake, 0 means real..')
y_train = []
for i in range(96):
y_train.append(1)
y_train.append(0)
y_test = []
def fast_conv(weights, img, iter, w_shp, size, N_i, N_j, rank):
# Define parameters
(F, C, X, Y) = w_shp
(N, C, H, W) = size
# Instantiate 4D tensor for input
input = T.tensor4(name='input')
# Initialize shared variable for weights.
weights = weights.eval().transpose(1, 2, 3, 0)
# Rows Clustering
kmeans_rows = KMeans(init='k-means++', n_clusters=N_i , n_init=10)
W_C = np.reshape(weights, (C, X*Y*F))
kmeans_rows.fit(W_C)
[clusters_rows, pos_c] = map(list, zip(*sorted(zip(kmeans_rows.labels_, range(C)), key=lambda item:item[0])))
new_W_C = W_C[pos_c, :]
# Cols Clustering
kmeans_cols = KMeans(init='k-means++', n_clusters=N_j , n_init=10)
W_F = np.reshape(new_W_C, (C*X*Y, F))
kmeans_cols.fit(W_F.T)
[clusters_cols, pos_f] = map(list, zip(*sorted(zip(kmeans_cols.labels_, range(F)), key=lambda item:item[0])))
new_W_F = np.reshape(W_F[:, pos_f], (C, X, Y, F))
# Breakdown into cubes
counter_cols = Counter(clusters_cols)
counter_rows = Counter(clusters_rows)
c_sum = 0
D_CF = []
for i, c in enumerate(list(counter_rows.values())):
f_sum = 0
D_F = []
for j, f in enumerate(list(counter_cols.values())):
# Crop cluster
W_tensor = dtensor(new_W_F[c_sum:c+c_sum, :, :, f_sum:f+f_sum])
# Apply CP-Decomposition on the clustered weight tensor
P, fit, itr, exectimes = cp_als(W_tensor, rank, init='random')
D_F.append(P)
f_sum += f
D_CF.append(D_F)
c_sum += c
# Compute Convolution
conv_out = theano.shared(np.zeros((N, F, H-X+1, W-Y+1)), name='out')
c_sum = 0
for i, c in enumerate(list(counter_rows.values())):
f_sum = 0
for j, f in enumerate(list(counter_cols.values())):
# Crop cluster
f_out = pos_f[f_sum:f+f_sum]
c_in = pos_c[c_sum:c+c_sum]
# Apply CP-Decomposition on the clustered weight tensor
P = D_CF[i][j]
for k in range(rank):
# Create shared variables
T_C = theano.shared(np.reshape(P.U[0][:,k], (1, c, 1, 1)), name='C_{0}_{1}_{2}'.format(k, i, j))
T_X = theano.shared(np.reshape(P.U[1][:,k], (1, 1, X, 1)), name='X_{0}_{1}_{2}'.format(k, i, j))
T_Y = theano.shared(np.reshape(P.U[2][:,k], (1, 1, 1, Y)), name='Y_{0}_{1}_{2}'.format(k, i, j))
T_F = theano.shared(np.reshape(P.U[3][:,k], (f, 1, 1, 1)), name='F_{0}_{1}_{2}'.format(k, i, j))
# Apply convolution on each dimension individually
conv_C = conv.conv2d(input[:, c_in, :, :], T_C)
conv_X = conv.conv2d(conv_C, T_X)
conv_Y = conv.conv2d(conv_X, T_Y)
conv_F = conv.conv2d(conv_Y, T_F)
if f == 1:
conv_out = T.set_subtensor(conv_out[:, f_out[0], :, :], np.add(conv_out[:, f_out[0], :, :], conv_F[:, 0, :, :]))
else:
conv_out = T.set_subtensor(conv_out[:, f_out, :, :], np.add(conv_out[:, f_out, :, :], conv_F))
f_sum += f
c_sum += c
# Map Theano function
f = theano.function([input], conv_out, profile=False)
# Execute Theano function
times = []
for i in range(iter):
start = time.time()
filtered_img = f(img)
done = time.time()
times.append(done-start)
avg1 = np.mean(times)
return filtered_img, avg1