Skip to content
Snippets Groups Projects
Commit 0fae2601 authored by Baptiste Bauvin's avatar Baptiste Bauvin
Browse files

MKL adaptation ongoing

parent 81803a33
Branches
Tags
No related merge requests found
Show whitespace changes
Inline Side-by-side
config_files/config_test.yml
+ 5
− 5
View file @ 0fae2601
......@@ -24,7 +24,7 @@ Classification:
classes:
type: ["multiview"]
algos_monoview: ["all"]
algos_multiview: ["mvml", "lp_norm_mkl",]
algos_multiview: ["lp_norm_mkl",]
stats_iter: 2
metrics: ["accuracy_score", "f1_score"]
metric_princ: "f1_score"
......@@ -210,10 +210,10 @@ easy_mkl:
lp_norm_mkl:
lmbda: [0.1]
max_rounds: [50]
max_diff: [0.0001]
kernel_types: ["rbf_kernel"]
kernel_configs:
n_loops: [50]
precision: [0.0001]
kernel: ["rbf"]
kernel_params:
gamma: [0.1]
mvml:
......
multiview_platform/mono_multi_view_classifiers/multiview_classifiers/additions/kernel_learning.py
+ 29
− 43
View file @ 0fae2601
......@@ -7,11 +7,8 @@ from ...utils.transformations import sign_labels, unsign_labels
class KernelClassifier(BaseMultiviewClassifier):
def __init__(self, random_state=None,
kernel_types=None, kernel_configs=None):
def __init__(self, random_state=None,):
super().__init__(random_state)
self.kernel_configs=kernel_configs
self.kernel_types=kernel_types
def _compute_kernels(self, X, example_indices, view_indices, ):
new_X = {}
......@@ -22,42 +19,37 @@ class KernelClassifier(BaseMultiviewClassifier):
**kernel_config)
return new_X
def _init_fit(self, X, y, train_indices, view_indices):
train_indices, view_indices = get_examples_views_indices(X,
train_indices,
def format_X(self, X, example_indices, view_indices):
example_indices, view_indices = get_examples_views_indices(X,
example_indices,
view_indices)
self.init_kernels(nb_view=len(view_indices), )
new_X = self._compute_kernels(X,
train_indices, view_indices)
new_y = sign_labels(y[train_indices])
return new_X, new_y
formatted_X = dict((index, X.get_v(view_index, example_indices=example_indices))
for index, view_index in enumerate(view_indices))
return formatted_X, example_indices
def extract_labels(self, predicted_labels):
signed_labels = np.sign(predicted_labels)
return unsign_labels(signed_labels)
def init_kernels(self, nb_view=2, ):
if isinstance(self.kernel_types, KernelDistribution):
self.kernel_functions = self.kernel_types.draw(nb_view)
elif isinstance(self.kernel_types, str):
self.kernel_functions = [getattr(pairwise, self.kernel_types)
if isinstance(self.kernel, KernelDistribution):
self.kernel = self.kernel.draw(nb_view)
elif isinstance(self.kernel, str):
self.kernel = [self.kernel
for _ in range(nb_view)]
elif isinstance(self.kernel_types, list):
self.kernel_functions = [getattr(pairwise, kernel_type)
for kernel_type in self.kernel_types]
if isinstance(self.kernel_configs, KernelConfigDistribution):
self.kernel_configs = self.kernel_configs.draw(nb_view)
self.kernel_configs = [kernel_config[kernel_function.__name__]
for kernel_config, kernel_function
in zip(self.kernel_configs,
self.kernel_functions)]
elif isinstance(self.kernel_configs, dict):
self.kernel_configs = [self.kernel_configs for _ in range(nb_view)]
elif isinstance(self.kernel, list):
pass
if isinstance(self.kernel_params, KernelConfigDistribution):
self.kernel_params = self.kernel_params.draw(nb_view)
self.kernel_params = [kernel_config[kernel_name]
for kernel_config, kernel_name
in zip(self.kernel_params,
self.kernel)]
elif isinstance(self.kernel_params, dict):
self.kernel_params = [self.kernel_params for _ in range(nb_view)]
else:
pass
......@@ -76,13 +68,8 @@ class KernelConfigDistribution:
def __init__(self, seed=42):
self.random_state=np.random.RandomState(seed)
self.possible_config = {
# "polynomial_kernel":{"degree": CustomRandint(low=1, high=7),
# "gamma": CustomUniform(),
# "coef0": CustomUniform()
#
# },
"chi2_kernel": {"gamma": CustomUniform()},
"rbf_kernel": {"gamma": CustomUniform()},
"additive_chi2": {"gamma": CustomUniform()},
"rbf": {"gamma": CustomUniform()},
}
def draw(self, nb_view):
......@@ -108,8 +95,7 @@ class KernelDistribution:
def __init__(self, seed=42):
self.random_state=np.random.RandomState(seed)
self.available_kernels = [pairwise.chi2_kernel,
pairwise.rbf_kernel,]
self.available_kernels = ["rbf",]
def draw(self, nb_view):
return self.random_state.choice(self.available_kernels, nb_view)
return list(self.random_state.choice(self.available_kernels, nb_view))
multiview_platform/mono_multi_view_classifiers/multiview_classifiers/lp_norm_mkl.py
+ 19
− 212
View file @ 0fae2601
from sklearn.metrics import pairwise
from metriclearning.lpMKL import MKL
from ..multiview.multiview_utils import BaseMultiviewClassifier, get_examples_views_indices
from .additions.kernel_learning import KernelClassifier, KernelConfigGenerator, KernelGenerator
from ..utils.hyper_parameter_search import CustomUniform, CustomRandint
classifier_class_name = "LPNormMKL"
### The following code is a welcome contribution by Riikka Huusari
# (riikka.huusari@lis-lab.fr) that we adapted te create the classifier
import numpy as np
from sklearn.base import BaseEstimator
from sklearn.base import ClassifierMixin
from sklearn.utils.multiclass import unique_labels
from sklearn.utils.validation import check_X_y
from .additions.data_sample import Metriclearn_array
class MKL(BaseEstimator, ClassifierMixin):
def __init__(self, lmbda, m_param=1.0, use_approx=True, max_rounds=50,
max_diff=0.0001, p=2):
print(lmbda)
# calculate nyström approximation (if used)
self.lmbda = lmbda
self.use_approx = use_approx
self.m_param = m_param
# Non-optimizable Hyper-params
self.max_rounds = max_rounds
self.max_diff = max_diff
self.p = p
def fit(self, X, y= None, views_ind=None):
if isinstance(X, Metriclearn_array):
self.X_ = X
elif isinstance(X, dict):
self.X_ = Metriclearn_array(X)
elif isinstance(X, np.ndarray) :
self.X_ = Metriclearn_array(X, views_ind)
self.classes_ = unique_labels(y)
check_X_y(self.X_, y)
self.y_ = y
n = self.X_.shape[0]
self._calc_nystrom(self.X_, n)
C, weights = self.learn_lpMKL()
self.C = C
self.weights = weights
def learn_lpMKL(self):
views = self.X_.n_views
X = self.X_
# p = 2
n = self.X_.shape[0]
weights = np.ones(views) / (views)
prevalpha = False
max_diff = 1
kernels = np.zeros((views, n, n))
for v in range(0, views):
kernels[v, :, :] = np.dot(self.U_dict[v], np.transpose(self.U_dict[v]))
rounds = 0
stuck = False
while max_diff > self.max_diff and rounds < self.max_rounds and not stuck:
# gammas are fixed upon arrival to the loop
# -> solve for alpha!
if self.m_param < 1 and self.use_approx:
combined_kernel = np.zeros((n, n))
for v in range(0, views):
combined_kernel = combined_kernel + weights[v] * kernels[v]
else:
combined_kernel = np.zeros((n, n))
for v in range(0, views):
combined_kernel = combined_kernel + weights[v]*X.get_view(v)
# combined kernel includes the weights
# alpha = (K-lambda*I)^-1 y
C = np.linalg.solve((combined_kernel + self.lmbda * np.eye(n)), self.y_)
# alpha fixed -> calculate gammas
weights_old = weights.copy()
# first the ||f_t||^2 todo wtf is the formula used here????
ft2 = np.zeros(views)
for v in range(0, views):
if self.m_param < 1 and self.use_approx:
# ft2[v,vv] = weights_old[v,vv] * np.dot(np.transpose(C), np.dot(np.dot(np.dot(data.U_dict[v],
# np.transpose(data.U_dict[v])),
# np.dot(data.U_dict[vv],
# np.transpose(data.U_dict[vv]))), C))
ft2[v] = np.linalg.norm(weights_old[v] * np.dot(kernels[v], C))**2
else:
ft2[v] = np.linalg.norm(weights_old[v] * np.dot(X.get_view(v), C))**2
# ft2[v] = weights_old[v] * np.dot(np.transpose(C), np.dot(data.kernel_dict[v], C))
# calculate the sum for downstairs
# print(weights_old)
# print(ft2)
# print(ft2 ** (p / (p + 1.0)))
downstairs = np.sum(ft2 ** (self.p / (self.p + 1.0))) ** (1.0 / self.p)
# and then the gammas
weights = (ft2 ** (1 / (self.p + 1))) / downstairs
# convergence
if prevalpha == False: # first time in loop we don't have a previous alpha value
prevalpha = True
diff_alpha = 1
else:
diff_alpha = np.linalg.norm(C_old - C) / np.linalg.norm(C_old)
max_diff_gamma_prev = max_diff_gamma
max_diff_gamma = np.max(np.max(np.abs(weights - weights_old)))
# try to see if convergence is as good as it gets: if it is stuck
if max_diff_gamma < 1e-3 and max_diff_gamma_prev < max_diff_gamma:
# if the gamma difference starts to grow we are most definitely stuck!
# (this condition determined empirically by running algo and observing the convergence)
stuck = True
if rounds > 1 and max_diff_gamma - max_diff_gamma_prev > 1e-2:
# If suddenly the difference starts to grow much
stuck = True
max_diff = np.max([max_diff_gamma, diff_alpha])
# print([max_diff_gamma, diff_alpha]) # print if convergence is interesting
C_old = C.copy()
rounds = rounds + 1
# print("\nlearned the weights:")
# np.set_printoptions(precision=3, suppress=True)
# print(weights)
# print("")
# print if resulting convergence is of interest
# print("convergence of ", max_diff, " at step ", rounds, "/500")
if stuck:
return C_old, weights_old
else:
return C, weights
def predict(self, X, views_ind=None):
if isinstance(X, Metriclearn_array):
# self.X_ = X
pass
elif isinstance(X, dict):
X = Metriclearn_array(X)
elif isinstance(X, np.ndarray):
X = Metriclearn_array(X, views_ind)
C = self.C
weights = self.weights
return self.lpMKL_predict(X , C, weights)
def lpMKL_predict(self, X, C, weights, views_ind=None):
if isinstance(X, Metriclearn_array):
# self.X_ = X
pass
elif isinstance(X, dict):
X = Metriclearn_array(X)
elif isinstance(X, np.ndarray):
X = Metriclearn_array(X, views_ind)
views = X.n_views
tt = X.shape[0]
m = self.X_.shape[0] # self.m_param * n
# NO TEST KERNEL APPROXIMATION
# kernel = weights[0] * self.data.test_kernel_dict[0]
# for v in range(1, views):
# kernel = kernel + weights[v] * self.data.test_kernel_dict[v]
# TEST KERNEL APPROXIMATION
kernel = np.zeros((tt, self.X_.shape[0]))
for v in range(0, views):
if self.m_param < 1:
kernel = kernel + weights[v] * np.dot(np.dot(X.get_view(v)[:, 0:m], self.W_sqrootinv_dict[v]),
np.transpose(self.U_dict[v]))
else:
kernel = kernel + weights[v] * X.get_view(v)
return np.dot(kernel, C)
def _calc_nystrom(self, kernels, n_approx):
# calculates the nyström approximation for all the kernels in the given dictionary
self.W_sqrootinv_dict = {}
self.U_dict = {}
for v in range(kernels.n_views):
kernel = kernels.get_view(v)
E = kernel[:, 0:n_approx]
W = E[0:n_approx, :]
Ue, Va, _ = np.linalg.svd(W)
vak = Va[0:n_approx]
inVa = np.diag(vak ** (-0.5))
U_v = np.dot(E, np.dot(Ue[:, 0:n_approx], inVa))
self.U_dict[v] = U_v
self.W_sqrootinv_dict[v] = np.dot(Ue[:, 0:n_approx], inVa)
classifier_class_name = "LPNormMKL"
class LPNormMKL(KernelClassifier, MKL):
def __init__(self, random_state=None, lmbda=0.1, m_param=1, max_rounds=50,
max_diff=0.0001, use_approx=True, kernel_types="rbf_kernel",
kernel_configs=None, p=2, prev_alpha=False):
super().__init__(random_state, kernel_configs=kernel_configs,
kernel_types=kernel_types)
super(BaseMultiviewClassifier, self).__init__(lmbda, m_param,
use_approx, max_rounds,
max_diff, p)
self.param_names = ["lmbda", "kernel_types", "kernel_configs"]
def __init__(self, random_state=None, lmbda=0.1, m_param=1, n_loops=50,
precision=0.0001, use_approx=True, kernel="rbf",
kernel_params=None):
super().__init__(random_state)
super(BaseMultiviewClassifier, self).__init__(lmbda, m_param=m_param,
kernel=kernel,
n_loops=n_loops,
precision=precision,
use_approx=use_approx,
kernel_params=kernel_params)
self.param_names = ["lmbda", "kernel", "kernel_params"]
self.distribs = [CustomUniform(), KernelGenerator(),
KernelConfigGenerator()]
self.prev_alpha = prev_alpha
def fit(self, X, y, train_indices=None, view_indices=None):
new_X, new_y = self._init_fit(X, y, train_indices, view_indices)
return super(LPNormMKL, self).fit(new_X, new_y)
formatted_X, train_indices = self.format_X(X, train_indices, view_indices)
self.init_kernels(nb_view=len(formatted_X))
return super(LPNormMKL, self).fit(formatted_X, y[train_indices])
def predict(self, X, example_indices=None, view_indices=None):
example_indices, view_indices = get_examples_views_indices(X,
example_indices,
view_indices)
new_X = self._compute_kernels(X, example_indices, view_indices)
print(self.C.shape)
new_X, _ = self.format_X(X, example_indices, view_indices)
return self.extract_labels(super(LPNormMKL, self).predict(new_X))
......
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Please register or to comment