model_name, global_pen, tv_ratio, l1_ratio = params
if model_name == 'pca':
model = sklearn.decomposition.PCA(n_components=N_COMP)
model.components_ = components.T
model.mean_ = X.mean(axis=0)
computed_projections = model.transform(X)
if model_name == 'struct_pca':
# We don'really need those parameters here
ltv = global_pen * tv_ratio
ll1 = l1_ratio * global_pen * (1 - ltv)
ll2 = (1 - l1_ratio) * global_pen * (1 - tv_ratio)
model = pca_tv.PCA_SmoothedL1_L2_TV(n_components=N_COMP,
l1=ll1, l2=ll2, ltv=ltv,
Atv=Atv,
Al1=Al1,
criterion="frobenius",
eps=1e-6,
max_iter=100,
inner_max_iter=int(1e4),
output=False)
model.V = components
computed_projections, _ = model.transform(X)
# Projections
projections_file = os.path.join(input_dir, "X_train_transform.npz")
projections = np.load(projections_file)['arr_0']
# Compare both
assert (np.allclose(projections, computed_projections))
# Create df for this model
index = pd.MultiIndex.from_arrays([subjects_id,
np.asarray([model_name] * n),
np.asarray([global_pen] * n),
alpha = 0.1
l1, l2, ltv = alpha * np.array((.4, 1/alpha, .001))
#k =0.5
#"l = 2*k
#g = 4
pca_sklearn = sklearn.decomposition.PCA(n_components=N_COMP)
pca_sklearn.fit(X)
V_pca_sklearn = pca_sklearn.components_.transpose()
del pca_sklearn
# ExcessiveGap
e_eg_sparse = pca_tv.PCA_SmoothedL1_L2_TV(l1, l2, ltv, Atv, Al1,
n_components=N_COMP,
criterion="frobenius",
eps=1e-6,
inner_max_iter=int(1e5),
use_eg=True,
output=True)
t = timeit.timeit(stmt='e_eg_sparse.fit(X)', setup="from __main__ import e_eg_sparse, X", number=1)
print "EG:", t
V_sparse_eg = e_eg_sparse.V
del e_eg_sparse
# CONESTA
e_con_sparse = pca_tv.PCA_SmoothedL1_L2_TV(50*l1, 50*l2, 50*ltv, Atv, Al1,
n_components=N_COMP,
criterion="frobenius",
eps=1e-6,
# inner_eps=1e-8,
inner_max_iter=int(1e5),
def mapper(key, output_collector):
import mapreduce as GLOBAL # access to global variables:
# GLOBAL.DATA
# GLOBAL.DATA ::= {"X":[Xtrain, ytrain], "y":[Xtest, ytest]}
# key: list of parameters
model_name, global_pen, tv_ratio, l1_ratio = key
if model_name == 'pca':
# Force the key
global_pen = tv_ratio = l1_ratio = 0
if model_name == 'sparse_pca':
# Force the key
tv_ratio = 0
l1_ratio = 1
ll1 = global_pen
if model_name == 'struct_pca':
ltv = global_pen * tv_ratio
ll1 = l1_ratio * global_pen * (1 - tv_ratio)
ll2 = (1 - l1_ratio) * global_pen * (1 - tv_ratio)
assert (np.allclose(ll1 + ll2 + ltv, global_pen))
X_train = GLOBAL.DATA_RESAMPLED["X"][0]
n, p = X_train.shape
X_test = GLOBAL.DATA_RESAMPLED["X"][1]
# A matrices
Atv = GLOBAL.Atv
Al1 = scipy.sparse.eye(p, p)
# Fit model
if model_name == 'pca':
model = sklearn.decomposition.PCA(n_components=N_COMP)
if model_name == 'sparse_pca':
model = sklearn.decomposition.SparsePCA(n_components=N_COMP, alpha=ll1)
if model_name == 'struct_pca':
model = pca_tv.PCA_SmoothedL1_L2_TV(n_components=N_COMP,
l1=ll1,
l2=ll2,
ltv=ltv,
Atv=Atv,
Al1=Al1,
criterion="frobenius",
eps=1e-6,
max_iter=100,
inner_max_iter=int(1e4),
output=False)
t0 = time.clock()
model.fit(X_train)
t1 = time.clock()
_time = t1 - t0
#print "X_test", GLOBAL.DATA["X"][1].shape
# Save the projectors
if (model_name == 'pca') or (model_name == 'sparse_pca'):
V = model.components_.T
if model_name == 'struct_pca':
V = model.V
# Project train & test data
if (model_name == 'pca') or (model_name == 'sparse_pca'):
X_train_transform = model.transform(X_train)
X_test_transform = model.transform(X_test)
if (model_name == 'struct_pca'):
X_train_transform, _ = model.transform(X_train)
X_test_transform, _ = model.transform(X_test)
# Reconstruct train & test data
# For SparsePCA or PCA, the formula is: UV^t (U is given by transform)
# For StructPCA this is implemented in the predict method (which uses
# transform)
if (model_name == 'pca') or (model_name == 'sparse_pca'):
X_train_predict = np.dot(X_train_transform, V.T)
X_test_predict = np.dot(X_test_transform, V.T)
if (model_name == 'struct_pca'):
X_train_predict = model.predict(X_train)
X_test_predict = model.predict(X_test)
# Compute Frobenius norm between original and recontructed datasets
frobenius_train = np.linalg.norm(X_train - X_train_predict, 'fro')
frobenius_test = np.linalg.norm(X_test - X_test_predict, 'fro')
# Compute explained variance ratio
evr_train = metrics.adjusted_explained_variance(X_train_transform)
evr_train /= np.var(X_train, axis=0).sum()
evr_test = metrics.adjusted_explained_variance(X_test_transform)
evr_test /= np.var(X_test, axis=0).sum()
# Compute geometric metrics and norms of components
TV = parsimony.functions.nesterov.tv.TotalVariation(1, A=Atv)
l0 = np.zeros((N_COMP, ))
l1 = np.zeros((N_COMP, ))
l2 = np.zeros((N_COMP, ))
tv = np.zeros((N_COMP, ))
for i in range(N_COMP):
# Norms
l0[i] = np.linalg.norm(V[:, i], 0)
l1[i] = np.linalg.norm(V[:, i], 1)
l2[i] = np.linalg.norm(V[:, i], 2)
tv[i] = TV.f(V[:, i])
ret = dict(frobenius_train=frobenius_train,
frobenius_test=frobenius_test,
components=V,
X_train_predict=X_train_predict,
X_test_predict=X_test_predict,
X_train_transform=X_train_transform,
X_test_transform=X_test_transform,
evr_train=evr_train,
evr_test=evr_test,
l0=l0,
l1=l1,
l2=l2,
tv=tv,
time=_time)
output_collector.collect(key, ret)
number=1)
print "Sparse PCA:", t
V_sparsepca_sklearn = sparsepca_sklearn.components_.transpose()
del sparsepca_sklearn
# Struct PCA with few TV
print "Fitting StructPCA"
l2 = 1
ltv = alpha * .001
e1 = pca_tv.PCA_SmoothedL1_L2_TV(
l1,
l2,
ltv,
Atv,
Al1,
n_components=N_COMP,
criterion="frobenius",
eps=1e-6,
# inner_eps=1e-8,
inner_max_iter=int(1e5),
use_eg=False,
output=False)
t = timeit.timeit(stmt='e1.fit(X)',
setup="from __main__ import e1, X",
number=1)
print "Fitting StructPCA:", t
V1 = e1.V
# Struct PCA with more TV
print "Fitting StructPCA with more TV"
ltv = alpha * .01