def test_doseframe_properties(self): m, n = 100, 50 d = DoseFrame(m, n, None) # frame dimensions immutable once set self.assert_exception( call=d.voxels, args=[m] ) self.assert_exception( call=d.beams, args=[m] ) d.data = rand(m, n) self.assertTrue( isinstance(d.data, ndarray) ) # matrix with wrong size fails self.assert_exception( call=d.data, args=[rand(m + 1, n)] ) d.data = sprand(m, n, 0.2, 'csr') self.assertTrue( isspmatrix(d.data) ) d.data = sprand(m, n, 0.2, 'csc') # coo matrix fails self.assert_exception( call=d.data, args=[sprand(m, n, 0.2)] ) # voxel labels vl = (10 * rand(m)).astype(int) d.voxel_labels = vl self.assertTrue( sum(vl - d.voxel_labels) == 0 ) vl_missized = (10 * rand(m + 1)).astype(int) self.assert_exception( call=d.voxel_labels, args=[vl_missized] ) # beam labels bl = (4 * rand(n)).astype(int) d.beam_labels = bl self.assertTrue( sum(bl - d.beam_labels) == 0 ) bl_missized = (10 * rand(n + 1)).astype(int) self.assert_exception( call=d.beam_labels, args=[bl_missized] ) # voxel weights self.assertTrue( isinstance(d.voxel_weights, ndarray) ) self.assertTrue( d.voxel_weights.size == m ) self.assertTrue( sum(d.voxel_weights != 1) == 0 ) vw = (5 * rand(m)).astype(int).astype(float) d.voxel_weights = vw self.assert_vector_equal( vw, d.voxel_weights ) vw_missized = (5 * rand(m + 1)).astype(int).astype(float) self.assert_exception( call=d.voxel_weights, args=[vw_missized] ) # beam weights self.assertTrue( isinstance(d.beam_weights, ndarray) ) self.assertTrue( d.beam_weights.size == n ) self.assertTrue( sum(d.beam_weights != 1) == 0 ) bw = (5 * rand(n)).astype(int).astype(float) d.beam_weights = bw self.assert_vector_equal( bw, d.beam_weights ) bw_missized = (5 * rand(n + 1)).astype(int).astype(float) self.assert_exception( call=d.beam_weights, args=[bw_missized] )
def generate_H(m,n,d):
''' generate a parity check matrix H (m*n), density d (0<d<1) '''
H = lil_matrix(m,n)
H %= 2
while not (all(sum(H,1)>=2) and all(sum(H,2)>=2)):
H += abs(sprand(m,n,d)) > 0
H %= 2
def load_simulated_data():
n_users = 1000
n_items = 1000
ratings = sprand(n_users, n_items, density=0.01, format='csr')
data = (np.random.randint(1, 5, size=ratings.nnz).astype(np.float64))
ratings.data = data
return ratings
def test_l1logistic_sparse_input_no_center():
"""Test that multiclass L1 Logistic raises an error when asked to center
sparse data.
"""
rs = np.random.RandomState(17)
X = sprand(10, 10, random_state=rs)
classes = ['abc', 'de', 'fgh']
y = np.array(classes)[rs.randint(3, size=10)]
with pytest.raises(ValueError):
UoI_L1Logistic(fit_intercept=True).fit(X, y)
def disabled_test_factorization_sparse():
I, J, K, rank = 10, 20, 75, 5
Tmat = sprand(I, J * K, 0.1).tocoo()
T = unfolded_sptensor((Tmat.data, (Tmat.row, Tmat.col)), None, 0, [], (I, J, K)).fold()
core, U = tucker.hooi(T, rank, maxIter=20)
Tmat = Tmat.toarray()
T = unfolded_dtensor(Tmat, 0, (I, J, K)).fold()
core2, U2 = tucker.hooi(T, rank, maxIter=20)
assert allclose(core2, core)
for i in range(len(U)):
assert allclose(U2[i], U[i])
def test_factorization_sparse():
I, J, K, rank = 10, 20, 75, 5
Tmat = sprand(I, J * K, 0.1).tocoo()
T = unfolded_sptensor((Tmat.data, (Tmat.row, Tmat.col)), None, 0, [], (I, J, K)).fold()
core, U = tucker_hooi.tucker_hooi(T, rank, maxIter=20)
Tmat = Tmat.toarray()
T = unfolded_dtensor(Tmat, 0, (I, J, K)).fold()
core2, U2 = tucker_hooi.tucker_hooi(T, rank, maxIter=20)
assert_true(allclose(core2, core))
for i in range(len(U)):
assert_true(allclose(U2[i], U[i]))
def test_doseframe_init_basic(self): m, n = 100, 50 d = DoseFrame() self.assert_nan( d.voxels ) self.assert_nan( d.beams ) d = DoseFrame(m, n, None) self.assertTrue( d.voxels == m ) self.assertTrue( d.beams == n ) b = BeamSet(n) d = DoseFrame(m, b, None) self.assertTrue( d.voxels == m ) self.assertTrue( d.beams == n ) A = rand(m, n) d = DoseFrame(None, None, A) self.assertTrue( d.voxels == m ) self.assertTrue( d.beams == n ) # size mismatches self.assert_exception( call=DoseFrame, args=[None, n + 1, A] ) self.assert_exception( call=DoseFrame, args=[m + 1, None, A] ) self.assert_exception( call=DoseFrame, args=[m, n + 1, A] ) self.assert_exception( call=DoseFrame, args=[m + 1, n, A] ) A = sprand(m, n, 0.2, 'csr') self.assertTrue( d.voxels == m ) self.assertTrue( d.beams == n ) A = sprand(m, n, 0.2, 'csc') self.assertTrue( d.voxels == m ) self.assertTrue( d.beams == n ) A = sprand(m, n, 0.2) # COO sparse storage, not supported self.assert_exception( call=DoseFrame, args=[None, None, A] )
def create_toy_data(T=1000,
J=9,
K=9,
M=5,
N=2,
dt=0.02,
c1=0.,
c2=1.,
c3=0.,
noisevar=0.):
# Create PRF and CGF
prf_size = (J, K)
cgf_size = (M + 1, 2 * N + 1)
w_prf = np.zeros(prf_size)
nt = int(np.round(J / 3.))
nf = int(np.round(K / 3.))
w_prf[nt:2 * nt, nf:2 * nf] = np.tile(np.linspace(-1, 1, nt), (nf, 1)).T
w_cgf = np.zeros(cgf_size)
nt = int(np.round(M / 2.))
nf = int(np.round(N / 2.))
w_cgf[-nt:-2 * nt:-1, N - nf:N + nf + 1] = -1
w_cgf[-1, 0] = -1
w_cgf[-1, -1] = 1
w_prf = gaussian_filter(w_prf, .75)
w_cgf = gaussian_filter(w_cgf, .75)
w_cgf[0, N] = 0
# DRC stimulus
S = np.asarray(sprand(T, K, density=.165).todense())
# Pad zeros around stimulus to simplify subsequent computations
S_pad = np.zeros((J - 1 + M + T, K + 2 * N))
S_pad[J - 1 + M:, N:-N] = S
# Predict response of full context model
y = np.zeros((T, ))
context_fast.predict_y_context(S_pad, w_prf.ravel(), w_cgf.ravel(), y, c1,
c2, c3, T, J, K, M, N)
# Add some noise
y += np.sqrt(noisevar) * np.random.randn(y.shape[0])
return S, S_pad, y, w_prf, w_cgf
def main():
# generate random Y
Y = sprand(1000, 500, 0.1, 'csc')
# train IMC
mf_results = train_mf(Y,
k=10,
lamb=0.1,
solver_type=0,
maxiter=10,
threads=4)
# get predictions and compute RMSE
mf_results.predict_mf()
rmse = mf_results.rmse(1)
print "RMSE = %.6f" % rmse
def test_l1logistic_sparse_input():
"""Test that multiclass L1 Logistic works when using sparse matrix
inputs"""
rs = np.random.RandomState(17)
X = sprand(100, 100, random_state=rs)
classes = ['abc', 'de', 'fgh']
y = np.array(classes)[rs.randint(3, size=100)]
kwargs = dict(
fit_intercept=False,
random_state=rs,
n_boots_sel=4,
n_boots_est=4,
n_C=7,
)
l1log = UoI_L1Logistic(**kwargs).fit(X, y)
y_hat = l1log.predict(X)
assert set(classes) >= set(y_hat)
from matplotlib.gridspec import GridSpec
fig = plt.figure(constrained_layout=True)
plt.set_cmap('rainbow')
# gs2 = GridSpec(2,3,figure=fig,left=0.05, right=0.48,wspace=0.03)
# ax0 = fig.add_subplot(gs2[0,0])
# ax1 = fig.add_subplot(gs2[0,1])
# c = ax0.pcolor(sin1*Z)
# ax0.set_title(r'$\mathbf{Y}$')
# ax0.set_axis_off()
# c = ax1.pcolor(sin1*np.ones((10,40)))#, edgecolors='k', linewidths=4)
# ax1.set_title(r'$\mathbf{PQ}$')
# ax1.set_axis_off()
from scipy.sparse import rand as sprand
np.savetxt("PQ.csv",sin1*Z,delimiter=',')
D = np.random.rand(10,60)
S = sprand(60,40,density=0.05)
print(np.shape(np.dot(D,S.todense())))
Z_new = sin1*Z + np.dot(D,S.todense())+np.random.normal(0,0.01,(10,40))# np.dot(D,S)# sin1*Z + np.dot(D,S) + np.random.randn(10,40,0.01)
np.savetxt("Z_new.csv",D,delimiter=',')
np.savetxt("D.csv",D,delimiter=',')
np.savetxt("S.csv",S.todense(),delimiter=',')
gs1 = GridSpec(2,3,figure=fig,left=0.55, right=0.98,wspace=0.03)
ax2 = fig.add_subplot(gs1[0,:-2])
ax3 = fig.add_subplot(gs1[:,2])
c = ax2.pcolor(D)#, edgecolors='k', linewidths=4)
ax2.set_title(r'$\mathbf{D}$')
ax2.set_axis_off()
c = ax3.pcolor(S.todense())
ax3.set_title(r'$\mathbf{S}$')
ax3.set_axis_off()
import time
start = time.time()
import numpy as np
from scipy.sparse import rand as sprand
import torch
from torch.autograd import Variable
n_users = 1000
n_items = 1000
ratings = sprand(n_users,n_items,density=0.1,format='csr',random_state=42)
ratings.data = (np.random.randint(1,5,size=ratings.nnz).astype(np.float64))
ratings = ratings.toarray() #this is the densification step!
class MatrixFactorization(torch.nn.Module):
def __init__(self, n_users, n_items, n_factors=20):
super().__init__()
self.user_factors = torch.nn.Embedding(n_users,n_factors,sparse=True)
self.item_factors = torch.nn.Embedding(n_items,n_factors,sparse=True)
def forward(self, user, item):
return(self.user_factors(user)*self.item_factors(item)).sum(1)
model = MatrixFactorization(n_users,n_items,n_factors=20)
loss_func = torch.nn.MSELoss()
optimizer = torch.optim.SGD(model.parameters(),lr=1e-6)
