def test_constructor_avoid_collision_level2(self):
# Test that level 2 collisions are avoided
pm = Methods(*[Pipe(SelectKBest(k=2), SVC(kernel="linear", C=C))\
for C in [1, 10]])
leaves_key = [l.get_key() for l in pm.walk_leaves()]
self.assertTrue(len(leaves_key) == len(set(leaves_key)),
u'Collision could not be avoided')
def test_twomethods(self):
key_y_pred = 'y' + conf.SEP + conf.PREDICTION
X, y = datasets.make_classification(n_samples=20, n_features=5,
n_informative=2)
# = With EPAC
wf = Methods(LDA(), SVC(kernel="linear"))
r_epac = wf.run(X=X, y=y)
# = With SKLEARN
lda = LDA()
svm = SVC(kernel="linear")
lda.fit(X, y)
svm.fit(X, y)
r_sklearn = [lda.predict(X), svm.predict(X)]
# Comparison
for i_cls in range(2):
comp = np.all(np.asarray(r_epac[i_cls][key_y_pred]) ==
np.asarray(r_sklearn[i_cls]))
self.assertTrue(comp, u'Diff Methods')
# test reduce
r_epac_reduce = [wf.reduce().values()[0][key_y_pred],
wf.reduce().values()[1][key_y_pred]]
comp = np.all(np.asarray(r_epac_reduce) == np.asarray(r_sklearn))
self.assertTrue(comp, u'Diff Perm / CV: EPAC reduce')
def test_constructor_avoid_collision_level2(self):
# Test that level 2 collisions are avoided
pm = Methods(*[Pipe(SelectKBest(k=2), SVC(kernel="linear", C=C))
for C in [1, 10]])
leaves_key = [l.get_key() for l in pm.walk_leaves()]
self.assertTrue(len(leaves_key) == len(set(leaves_key)),
u'Collision could not be avoided')
def test_twomethods(self):
key_y_pred = 'y' + conf.SEP + conf.PREDICTION
X, y = datasets.make_classification(n_samples=20, n_features=5,
n_informative=2)
# = With EPAC
wf = Methods(LDA(), SVC(kernel="linear"))
r_epac = wf.run(X=X, y=y)
# = With SKLEARN
lda = LDA()
svm = SVC(kernel="linear")
lda.fit(X, y)
svm.fit(X, y)
r_sklearn = [lda.predict(X), svm.predict(X)]
# Comparison
for i_cls in range(2):
comp = np.all(np.asarray(r_epac[i_cls][key_y_pred]) ==
np.asarray(r_sklearn[i_cls]))
self.assertTrue(comp, u'Diff Methods')
# test reduce
r_epac_reduce = [wf.reduce().values()[0][key_y_pred],
wf.reduce().values()[1][key_y_pred]]
comp = np.all(np.asarray(r_epac_reduce) == np.asarray(r_sklearn))
self.assertTrue(comp, u'Diff Perm / CV: EPAC reduce')
def get_workflow(self):
n_folds = 2
n_folds_nested = 3
k_values = [1, 2]
C_values = [1, 2]
pipelines = Methods(*[Pipe(SelectKBest(k=k),
Methods(*[SVC(kernel="linear", C=C)
for C in C_values]))
for k in k_values])
pipeline = CVBestSearchRefitParallel(pipelines,
n_folds=n_folds_nested)
wf = CV(pipeline, n_folds=n_folds)
return wf
def get_workflow(self):
####################################################################
## EPAC WORKFLOW
# -------------------------------------
# Perms Perm (Splitter)
# / | \
# 0 1 2 Samples (Slicer)
# |
# CV CV (Splitter)
# / | \
# 0 1 2 Folds (Slicer)
# | | |
# Pipeline Pipeline Pipeline Sequence
# |
# 2 SelectKBest (Estimator)
# |
# Methods
# | \
# SVM(linear,C=1) SVM(linear,C=10) Classifiers (Estimator)
pipeline = Pipe(SelectKBest(k=2),
Methods(*[SVC(kernel="linear", C=C)
for C in [1, 3]]))
wf = Perms(CV(pipeline, n_folds=3),
n_perms=3,
permute="y",
random_state=1)
return wf
def func_memm_local():
print("memm_local pt1")
## 1) Build a dataset and convert to np.memmap (for big matrix)
## ============================================================
X, y = datasets.make_classification(n_samples=500,
n_features=5000,
n_informative=2,
random_state=1)
print("memm_local pt2")
X = convert2memmap(X)
y = convert2memmap(y)
Xy = dict(X=X, y=y)
## 2) Build two workflows respectively
## =======================================================
print("memm_local pt3")
from sklearn.svm import SVC
from epac import CV, Methods
cv_svm_local = CV(Methods(*[SVC(
kernel="linear"), SVC(kernel="rbf")]),
n_folds=3)
print("memm_local pt4")
# from epac import LocalEngine
# local_engine = LocalEngine(cv_svm_local, num_processes=2)
# cv_svm = local_engine.run(**Xy)
cv_svm_local.run(**Xy)
print(cv_svm_local.reduce())
print("memm_local pt5")
def get_workflow(self, n_features=int(1E03)):
random_state = 0
C_values = [1, 10]
k_values = 0
k_max = "auto"
n_folds_nested = 5
n_folds = 10
n_perms = 10
if k_max != "auto":
k_values = range_log2(np.minimum(int(k_max), n_features),
add_n=True)
else:
k_values = range_log2(n_features, add_n=True)
cls = Methods(*[
Pipe(SelectKBest(k=k), SVC(C=C, kernel="linear")) for C in C_values
for k in k_values
])
pipeline = CVBestSearchRefit(cls,
n_folds=n_folds_nested,
random_state=random_state)
wf = Perms(CV(pipeline, n_folds=n_folds),
n_perms=n_perms,
permute="y",
random_state=random_state)
return wf
def test_cvbestsearchrefit(self):
X, y = datasets.make_classification(n_samples=12,
n_features=10,
n_informative=2)
n_folds_nested = 2
#random_state = 0
C_values = [.1, 0.5, 1, 2, 5]
kernels = ["linear", "rbf"]
key_y_pred = 'y' + conf.SEP + conf.PREDICTION
# With EPAC
methods = Methods(*[SVC(C=C, kernel=kernel)
for C in C_values for kernel in kernels])
wf = CVBestSearchRefitParallel(methods, n_folds=n_folds_nested)
wf.run(X=X, y=y)
r_epac = wf.reduce().values()[0]
# - Without EPAC
r_sklearn = dict()
clf = SVC(kernel="linear")
parameters = {'C': C_values, 'kernel': kernels}
cv_nested = StratifiedKFold(y=y, n_folds=n_folds_nested)
gscv = grid_search.GridSearchCV(clf, parameters, cv=cv_nested)
gscv.fit(X, y)
r_sklearn[key_y_pred] = gscv.predict(X)
r_sklearn[conf.BEST_PARAMS] = gscv.best_params_
# - Comparisons
comp = np.all(r_epac[key_y_pred] == r_sklearn[key_y_pred])
self.assertTrue(comp, u'Diff CVBestSearchRefitParallel: prediction')
for key_param in r_epac[conf.BEST_PARAMS][0]:
if key_param in r_sklearn[conf.BEST_PARAMS]:
comp = r_sklearn[conf.BEST_PARAMS][key_param] == \
r_epac[conf.BEST_PARAMS][0][key_param]
self.assertTrue(
comp,
u'Diff CVBestSearchRefitParallel: best parameters')
def test_prev_state_methods(self):
## 1) Build dataset
## ================================================
X, y = datasets.make_classification(n_samples=5,
n_features=20,
n_informative=2)
Xy = {"X": X, "y": y}
methods = Methods(*[TOY_CLF(v_lambda=v_lambda)
for v_lambda in [2, 1]])
methods.run(**Xy)
ps_methods = WarmStartMethods(*[TOY_CLF(v_lambda=v_lambda)
for v_lambda in [2, 1]])
ps_methods.run(**Xy)
self.assertTrue(compare_two_node(methods, ps_methods))
self.assertTrue(comp_2wf_reduce_res(methods, ps_methods))
def test_engine_info(self):
n_samples = 20
n_features = 100
n_proc = 2
X, y = datasets.make_classification(n_samples=n_samples,
n_features=n_features,
n_informative=2,
random_state=1)
Xy = dict(X=X, y=y)
cv_svm_local = CV(Methods(*[SVC(
kernel="linear"), SVC(kernel="rbf")]),
n_folds=3)
swf_engine = SomaWorkflowEngine(cv_svm_local,
num_processes=n_proc,
resource_id="jl237561@gabriel",
login="******",
remove_finished_wf=False,
remove_local_tree=False,
queue="Global_long")
swf_engine.run(**Xy)
print("engine_info ================")
for job_info in swf_engine.engine_info:
print(" job_info=================")
print(" mem_cost= ", job_info.mem_cost)
print(" vmem_cost= ", job_info.vmem_cost)
print(" time_cost= ", job_info.time_cost)
self.assertTrue(job_info.time_cost > 0)
def test_prev_state_methods(self):
## 1) Build dataset
## ================================================
X, y = datasets.make_classification(n_samples=5,
n_features=20,
n_informative=2)
Xy = {"X": X, "y": y}
methods = Methods(*[TOY_CLF(v_lambda=v_lambda)
for v_lambda in [2, 1]])
methods.run(**Xy)
ps_methods = WarmStartMethods(*[TOY_CLF(v_lambda=v_lambda)
for v_lambda in [2, 1]])
ps_methods.run(**Xy)
self.assertTrue(compare_two_node(methods, ps_methods))
self.assertTrue(comp_2wf_reduce_res(methods, ps_methods))
def do_all(options):
if options.k_max != "auto":
k_values = range_log2(np.minimum(int(options.k_max),
options.n_features),
add_n=True)
else:
k_values = range_log2(options.n_features, add_n=True)
C_values = [1, 10]
random_state = 0
#print options
#sys.exit(0)
if options.trace:
from epac import conf
conf.TRACE_TOPDOWN = True
## 1) Build dataset
## ================
X, y = datasets.make_classification(n_samples=options.n_samples,
n_features=options.n_features,
n_informative=options.n_informative)
## 2) Build Workflow
## =================
time_start = time.time()
## CV + Grid search of a pipeline with a nested grid search
cls = Methods(*[
Pipe(SelectKBest(k=k), SVC(kernel="linear", C=C)) for C in C_values
for k in k_values
])
pipeline = CVBestSearchRefit(cls,
n_folds=options.n_folds_nested,
random_state=random_state)
wf = Perms(CV(pipeline, n_folds=options.n_folds),
n_perms=options.n_perms,
permute="y",
random_state=random_state)
print "Time ellapsed, tree construction:", time.time() - time_start
## 3) Run Workflow
## ===============
time_fit_predict = time.time()
## Run on local machine
sfw_engine = SomaWorkflowEngine(tree_root=wf,
num_processes=options.n_cores)
## Run on cluster
# sfw_engine = SomaWorkflowEngine(
# tree_root=wf,
# num_processes=options.n_cores,
# resource_id="jl237561@gabriel",
# login="******")
wf = sfw_engine.run(X=X, y=y)
print "Time ellapsed, fit predict:", time.time() - time_fit_predict
time_reduce = time.time()
## 4) Reduce Workflow
## ==================
print wf.reduce()
print "Time ellapsed, reduce:", time.time() - time_reduce
def test_peristence_load_and_fit_predict(self):
X, y = datasets.make_classification(n_samples=20,
n_features=10,
n_informative=2)
n_folds = 2
n_folds_nested = 3
k_values = [1, 2]
C_values = [1, 2]
pipelines = Methods(*[
Pipe(SelectKBest(
k=k), Methods(*[SVC(kernel="linear", C=C) for C in C_values]))
for k in k_values
])
pipeline = CVBestSearchRefitParallel(pipelines, n_folds=n_folds_nested)
tree_mem = CV(pipeline,
n_folds=n_folds,
reducer=ClassificationReport(keep=False))
# Save Tree
import tempfile
store = StoreFs(dirpath=tempfile.mkdtemp(), clear=True)
tree_mem.save_tree(store=store)
tree_mem.run(X=X, y=y)
res_mem = tree_mem.reduce().values()[0]
# Reload Tree
tree_fs_noresults = store.load()
tree_fs_noresults.run(X=X, y=y)
res_fs_noresults = tree_fs_noresults.reduce().values()[0]
# Save with results
tree_fs_noresults.save_tree(store=store)
tree_fs_withresults = store.load()
res_fs_withresults = tree_fs_withresults.reduce().values()[0]
# Compare
comp = np.all([
np.all(np.asarray(res_mem[k]) == np.asarray(res_fs_noresults[k]))
and np.all(
np.asarray(res_fs_noresults[k]) == np.asarray(
res_fs_withresults[k])) for k in res_mem
])
self.assertTrue(comp)
def test_cv_best_search_refit_parallel(self):
n_folds = 2
n_folds_nested = 3
k_values = [1, 2]
C_values = [1, 2]
n_samples = 500
n_features = 10000
n_cores = 2
X, y = datasets.make_classification(n_samples=n_samples,
n_features=n_features,
n_informative=5)
# epac workflow for paralle computing
pipelines = Methods(*[
Pipe(SelectKBest(
k=k), Methods(*[SVC(kernel="linear", C=C) for C in C_values]))
for k in k_values
])
pipeline = CVBestSearchRefitParallel(pipelines, n_folds=n_folds_nested)
wf = CV(pipeline, n_folds=n_folds)
sfw_engine = SomaWorkflowEngine(tree_root=wf,
num_processes=n_cores,
remove_finished_wf=False,
remove_local_tree=False)
sfw_engine_wf = sfw_engine.run(X=X, y=y)
# epac workflow for normal node computing
pipelines2 = Methods(*[
Pipe(SelectKBest(
k=k), Methods(*[SVC(kernel="linear", C=C) for C in C_values]))
for k in k_values
])
pipeline2 = CVBestSearchRefitParallel(pipelines2,
n_folds=n_folds_nested)
wf2 = CV(pipeline2, n_folds=n_folds)
wf2.run(X=X, y=y)
self.assertTrue(compare_two_node(sfw_engine_wf, wf2))
self.assertTrue(comp_2wf_reduce_res(sfw_engine_wf, wf2))
def test_peristence_perm_cv_parmethods_pipe_vs_sklearn(self):
key_y_pred = 'y' + conf.SEP + conf.PREDICTION
X, y = datasets.make_classification(n_samples=12,
n_features=10,
n_informative=2)
n_folds_nested = 2
#random_state = 0
C_values = [.1, 0.5, 1, 2, 5]
kernels = ["linear", "rbf"]
# With EPAC
methods = Methods(
*[SVC(C=C, kernel=kernel) for C in C_values for kernel in kernels])
wf = CVBestSearchRefitParallel(methods, n_folds=n_folds_nested)
# Save workflow
# -------------
import tempfile
#store = StoreFs("/tmp/toto", clear=True)
store = StoreFs(tempfile.mktemp())
wf.save_tree(store=store)
wf = store.load()
wf.run(X=X, y=y)
## Save results
wf.save_tree(store=store)
wf = store.load()
r_epac = wf.reduce().values()[0]
# - Without EPAC
r_sklearn = dict()
clf = SVC(kernel="linear")
parameters = {'C': C_values, 'kernel': kernels}
cv_nested = StratifiedKFold(y=y, n_folds=n_folds_nested)
gscv = grid_search.GridSearchCV(clf, parameters, cv=cv_nested)
gscv.fit(X, y)
r_sklearn[key_y_pred] = gscv.predict(X)
r_sklearn[conf.BEST_PARAMS] = gscv.best_params_
r_sklearn[conf.BEST_PARAMS]['name'] = 'SVC'
# - Comparisons
comp = np.all(r_epac[key_y_pred] == r_sklearn[key_y_pred])
self.assertTrue(comp, u'Diff CVBestSearchRefitParallel: prediction')
comp = np.all([
r_epac[conf.BEST_PARAMS][0][p] == r_sklearn[conf.BEST_PARAMS][p]
for p in r_sklearn[conf.BEST_PARAMS]
])
self.assertTrue(comp,
u'Diff CVBestSearchRefitParallel: best parameters')
def test_cvbestsearchrefit_select_k_best(self):
list_C_value = range(2, 10, 1)
# print repr(list_C_value)
for C_value in list_C_value:
# C_value = 2
# print C_value
X, y = datasets.make_classification(n_samples=100,
n_features=500,
n_informative=5)
n_folds_nested = 2
#random_state = 0
k_values = [2, 3, 4, 5, 6]
key_y_pred = 'y' + conf.SEP + conf.PREDICTION
# With EPAC
methods = Methods(*[Pipe(SelectKBest(k=k),
SVC(C=C_value, kernel="linear"))
for k in k_values])
wf = CVBestSearchRefitParallel(methods, n_folds=n_folds_nested)
wf.run(X=X, y=y)
r_epac = wf.reduce().values()[0]
# - Without EPAC
from sklearn.pipeline import Pipeline
r_sklearn = dict()
clf = Pipeline([('anova', SelectKBest(k=3)),
('svm', SVC(C=C_value, kernel="linear"))])
parameters = {'anova__k': k_values}
cv_nested = StratifiedKFold(y=y, n_folds=n_folds_nested)
gscv = grid_search.GridSearchCV(clf, parameters, cv=cv_nested)
gscv.fit(X, y)
r_sklearn[key_y_pred] = gscv.predict(X)
r_sklearn[conf.BEST_PARAMS] = gscv.best_params_
r_sklearn[conf.BEST_PARAMS]['k'] = \
r_sklearn[conf.BEST_PARAMS]['anova__k']
# - Comparisons
comp = np.all(r_epac[key_y_pred] == r_sklearn[key_y_pred])
self.assertTrue(comp,
u'Diff CVBestSearchRefitParallel: prediction')
for key_param in r_epac[conf.BEST_PARAMS][0]:
if key_param in r_sklearn[conf.BEST_PARAMS]:
comp = r_sklearn[conf.BEST_PARAMS][key_param] == \
r_epac[conf.BEST_PARAMS][0][key_param]
self.assertTrue(
comp,
u'Diff CVBestSearchRefitParallel: best parameters')
def get_workflow(self):
####################################################################
## EPAC WORKFLOW
# -------------------------------------
# Perms Perm (Splitter)
# / | \
# 0 1 2 Samples (Slicer)
# |
# CV CV (Splitter)
# / | \
# 0 1 2 Folds (Slicer)
# | | |
# Pipeline Pipeline Pipeline Sequence
# |
# 2 SelectKBest (Estimator)
# |
# Methods
# | \
# SVM(linear,C=1) SVM(linear,C=10) Classifiers (Estimator)
pipeline = Pipe(SelectKBest(k=2),
Methods(*[SVC(kernel="linear", C=C) for C in [1, 3]]))
wf = CV(pipeline, n_folds=3, reducer=ClassificationReport(keep=True))
return wf
# -*- coding: utf-8 -*-
"""
Created on Thu May 23 15:21:35 2013
@author: ed203246
"""
from sklearn import datasets
from sklearn.svm import LinearSVC as SVM
from sklearn.discriminant_analysis import LinearDiscriminantAnalysis as LDA
from sklearn.feature_selection import SelectKBest
from epac.map_reduce.reducers import PvalPerms
import numpy
X, y = datasets.make_classification(n_samples=100,
n_features=200,
n_informative=2)
X = numpy.random.rand(*X.shape)
from epac import Perms, CV, Methods
perms_cv_svm = Perms(CV(Methods(SVM(loss="l1"), SVM(loss="l2"))), n_perms=100)
perms_cv_svm.run(X=X, y=y)
perms_cv_svm.reduce()
self = perms_cv_svm
key = 'LinearSVC(loss=l1)'
self = PvalPerms()
# To save the results of the top-down operation (run) on the disk, it is # possible to convert it to CSV format from epac import export_leaves_csv export_leaves_csv(pipe, 'my_result_run.csv') ## Parallelization ## =============== # Multi-classifiers # ----------------- # Methods Methods (Splitter) # / \ # SVM(C=1) SVM(C=10) Classifiers (Estimator) from epac import Methods multi = Methods(SVM(C=1), SVM(C=10)) multi.run(X=X, y=y) print multi.reduce() # Reduce format outputs into "ResultSet" which is a dict-like structure # which contains the "keys" of the methods that have beeen used. # You can also export the results of the bottom-up operation (reduce) to CSV from epac import export_resultset_csv export_resultset_csv(multi.reduce(), 'my_result_reduce.csv') # Methods Methods (Splitter) # / \ # SVM(l1, C=1) SVM(l1, C=10) ..... SVM(l2, C=10) Classifiers (Estimator)
n_folds = 10
anova_svm = Pipe(SelectKBest(k=5),
preprocessing.StandardScaler(),
SVM(class_weight='auto'))
cv = CV(anova_svm, n_folds=n_folds)
cv.run(X=X, y=y)
#
res_cv_anova_svm = cv.reduce()
res_cv_anova_svm["SelectKBest/StandardScaler/LinearSVC"]['y/test/score_recall']
##############################################################################
## Multimethods, "Methods": SVM l1 vs l2
from epac import Methods, CV
svms = Methods(SVM(penalty="l1", class_weight='auto', dual=False),
SVM(penalty="l2", class_weight='auto', dual=False))
cv = CV(svms, n_folds=n_folds)
cv.run(X=X, y=y)
res_cv_svms = cv.reduce()
#
print res_cv_svms
print res_cv_svms["LinearSVC(penalty=l1)"]['y/test/score_recall']
print res_cv_svms["LinearSVC(penalty=l2)"]['y/test/score_recall']
# !!! BIASED RESULT !!!
# Re-fit on all data to see which mode is choosen. Warning !!! this is biased
# since all data have been used. We use for information only. No score can be
# used from it. We look the weights map.
svms.run(X=X, y=y)
print svms.children[0]
# Predict can return an array. In this case EPAC will
# put the prediction in a Result (a dictionnary). with key = "y/pred". y being the
# difference between input agrument of fit and predict. The true y will also figure
# in the result with key "y/true"
class MySVM:
def __init__(self, C=1.0):
self.C = C
def fit(self, X, y):
from sklearn.svm import SVC
self.svc = SVC(C=self.C)
self.svc.fit(X, y)
def predict(self, X):
return self.svc.predict(X)
svms = Methods(MySVM(C=1.0), MySVM(C=2.0))
cv = CV(svms, cv_key="y", cv_type="stratified", n_folds=2,
reducer=None)
cv.run(X=X, y=y) # top-down process to call transform
cv.reduce() # buttom-up process
from sklearn.decomposition import PCA
class MyPCA(PCA):
"""PCA with predict method"""
def predict(self, X):
"""Project to X PCs then project back to original space
If X is not singular, self.fit(X).predict(X) == X"""
return np.dot(self.transform(X), self.components_) + self.mean_
pcas = Methods(MyPCA(n_components=1), MyPCA(n_components=2))
cv = CV(pcas, n_folds=2, reducer=None)
X, y = datasets.make_classification(n_samples=500,
n_features=10000,
n_informative=2,
random_state=1)
X = convert2memmap(X)
y = convert2memmap(y)
Xy = dict(X=X, y=y)
## 2) Build two workflows respectively
## =======================================================
from sklearn.svm import SVC
from epac import CV, Methods
cv_svm_local = CV(
Methods(*[SVC(kernel="linear"), SVC(kernel="rbf")]), n_folds=3)
cv_svm_swf = CV(Methods(*[SVC(kernel="linear"), SVC(kernel="rbf")]), n_folds=3)
## 3) Run two workflows using local engine and soma-workflow
## =========================================================
from epac import LocalEngine
local_engine = LocalEngine(cv_svm_local, num_processes=2)
cv_svm = local_engine.run(X=X, y=y)
print(cv_svm.reduce())
from epac import SomaWorkflowEngine
swf_engine = SomaWorkflowEngine(
cv_svm_swf,
num_processes=2,
#resource_id="jl237561@gabriel",
# Each node call the "transform" method, that take a dictionnary as input
# and produces a dictionnary as output. The output is passed to the next node.
# The return value of the run is simply agregation of the outputs (dict) of
# the leaf nodes
## Parallelization
## ===============
# Multi-classifiers
# -----------------
# Methods Methods (Splitter)
# / \
# SVM(C=1) SVM(C=10) Classifiers (Estimator)
from epac import Methods
multi = Methods(SVM(C=1), SVM(C=10))
multi.run(X=X, y=y)
print multi.reduce()
# Reduce format outputs into "ResultSet" which is a dict-like structure
# which contains the "keys" of the methods that as beeen used.
# Methods Methods (Splitter)
# / \
# SVM(l1, C=1) SVM(l1, C=10) ..... SVM(l2, C=10) Classifiers (Estimator)
svms = Methods(*[SVM(loss=loss, C=C) for loss in ("l1", "l2") for C in [1, 10]])
svms.run(X=X, y=y)
print svms.reduce()
# Parallelize sequential Pipeline: Anova(k best selection) + SVM.
from epac import Methods
import numpy as np
from sklearn import datasets
from sklearn.svm import SVC
X, y = datasets.make_classification(n_samples=500,
n_features=200000,
n_informative=2,
random_state=1)
methods = Methods(*[SVC(C=1, kernel='linear'), SVC(C=1, kernel='rbf')])
data = {"X":X, 'y':y, "methods": methods}
# X = np.random.random((500, 200000))
def map_func(data):
from sklearn.cross_validation import StratifiedKFold
from sklearn import svm, cross_validation
kfold = StratifiedKFold(y=data['y'], n_folds=3)
# kfold = cross_validation.KFold(n=data.X.shape[0], n_folds=3)
# svc = SVC(C=1, kernel='linear')
for train, test in kfold:
# svc.fit(data['X'][train], data['y'][train])
# svc.predict(data['X'][test])
data['methods'].run(X=data["X"][train], y=data['y'][train])
# To save the results of the top-down operation (run) on the disk, it is # possible to convert it to CSV format from epac import export_leaves_csv export_leaves_csv(pipe, 'my_result_run.csv') ## Parallelization ## =============== # Multi-classifiers # ----------------- # Methods Methods (Splitter) # / \ # SVM(C=1) SVM(C=10) Classifiers (Estimator) from epac import Methods multi = Methods(SVM(C=1), SVM(C=10)) multi.run(X=X, y=y) print(multi.reduce()) # Reduce format outputs into "ResultSet" which is a dict-like structure # which contains the "keys" of the methods that have beeen used. # You can also export the results of the bottom-up operation (reduce) to CSV from epac import export_resultset_csv export_resultset_csv(multi.reduce(), 'my_result_reduce.csv') # Methods Methods (Splitter) # / \ # SVM(l1, C=1) SVM(l1, C=10) ..... SVM(l2, C=10) Classifiers (Estimator) svms = Methods(
def get_workflow(self):
wf = Methods(*[SVC(kernel="linear", C=C) for C in [1, 3]])
return wf
return {"y/pred": pred_y, "y/true": y, "best_beta": self.v_beta}
if __name__ == "__main__":
## 1) Build dataset
## ================================================
X, y = datasets.make_classification(n_samples=10,
n_features=5,
n_informative=2,
random_state=1)
Xy = {"X": X, "y": y}
## 2) Build Methods
## ================================================
print("Methods ===================================")
methods = Methods(*[TOY_CLF(v_lambda=v_lambda) for v_lambda in [2, 1]])
print(methods.run(**Xy))
## 3) Build WarmStartMethods like Methods
## ================================================
## WarmStartMethods
## / \
## TOY_CLF(v_lambda=2) TOY_CLF(v_lambda=1)
##
## 1. WarmStartMethods will look for different argumenets as signature
## For example, here is v_lambda, there are different for each leaf
## 2. And then run TOY_CLF(v_lambda=2).transform
## 3. Except v_lambda, WarmStartMethods copy all the other parameters
## from TOY_CLF(v_lambda=2) to TOY_CLF(v_lambda=1) as initialization
## 4. Finally call TOY_CLF(v_lambda=1).transform
print("WarmStartMethods ==========================")
def test_memmapping(self):
## 1) Building dataset
## ============================================================
if self.memmap:
# If the proc is 1, always generate the matrix
# Otherwise, load it if it exists, or create it if it doesn't
writing_mode = (self.n_proc == 1)
X = create_mmat(self.n_samples,
self.n_features,
dir=self.directory,
writing_mode=writing_mode)
y = create_array(self.n_samples, [0, 1],
dir=self.directory,
writing_mode=writing_mode)
Xy = dict(X=X, y=y)
else:
X, y = datasets.make_classification(n_samples=self.n_samples,
n_features=self.n_features,
n_informative=2,
random_state=1)
Xy = dict(X=X, y=y)
## 2) Building workflow
## =======================================================
from sklearn.svm import SVC
from epac import CV, Methods
cv_svm_local = CV(Methods(*[SVC(
kernel="linear"), SVC(kernel="rbf")]),
n_folds=3)
cv_svm = None
if self.is_swf:
# Running on the cluster
from epac import SomaWorkflowEngine
mmap_mode = None
if self.memmap:
mmap_mode = "r+"
swf_engine = SomaWorkflowEngine(
cv_svm_local,
num_processes=self.n_proc,
resource_id="jl237561@gabriel",
login="******",
# remove_finished_wf=False,
# remove_local_tree=False,
mmap_mode=mmap_mode,
queue="Global_long")
cv_svm = swf_engine.run(**Xy)
# Printing information about the jobs
time.sleep(2)
print('')
sum_memory = 0
max_time_cost = 0
for job_info in swf_engine.engine_info:
print(
"mem_cost = {0}, vmem_cost = {1}, time_cost = {2}".format(
job_info.mem_cost, job_info.vmem_cost,
job_info.time_cost))
sum_memory += job_info.mem_cost
if max_time_cost < job_info.time_cost:
max_time_cost = job_info.time_cost
print("sum_memory = ", sum_memory)
print("max_time_cost = ", max_time_cost)
else:
# Running on the local machine
from epac import LocalEngine
local_engine = LocalEngine(cv_svm_local, num_processes=self.n_proc)
cv_svm = local_engine.run(**Xy)
cv_svm_reduce = cv_svm.reduce()
print("\n -> Reducing results")
print(cv_svm_reduce)
# Creating the directory to save results, if it doesn't exist
dirname = 'tmp_save_tree/'
if self.directory is None:
directory = '/tmp'
else:
directory = self.directory
if not os.path.isdir(directory):
os.mkdir(directory)
dirpath = os.path.join(directory, dirname)
if not os.path.isdir(dirpath):
os.mkdir(dirpath)
if self.n_proc == 1:
## 4.1) Saving results on the disk for one process
## ===================================================
store = StoreFs(dirpath=dirpath, clear=True)
cv_svm.save_tree(store=store)
with open(os.path.join(directory, "tmp_save_results"), 'w+') \
as filename:
print(filename.name)
pickle.dump(cv_svm_reduce, filename)
else:
## 4.2) Loading the results for one process
## ===================================================
try:
store = StoreFs(dirpath=dirpath, clear=False)
cv_svm_one_proc = store.load()
with open(os.path.join(directory, "tmp_save_results"), 'r+') \
as filename:
cv_svm_reduce_one_proc = pickle.load(filename)
## 5.2) Comparing results to the results for one process
## ===================================================
print("\nComparing %i proc with one proc" % self.n_proc)
self.assertTrue(compare_two_node(cv_svm, cv_svm_one_proc))
self.assertTrue(isequal(cv_svm_reduce, cv_svm_reduce_one_proc))
except KeyError:
print("Warning: ")
print("No previous tree detected, no possible "\
"comparison of results")
Xd, yd = IO.read_Xy(WD=WD)
Xd.PAS2gr[Xd.PAS2gr == 1] = -1
Xd.PAS2gr[Xd.PAS2gr == 2] = 1
Xd.CB_EXPO[Xd.CB_EXPO == 0] = -1
X = np.asarray(Xd)
y = np.asarray(yd)
C_values = [0.01, 0.05, .1, .5, 1, 5, 10]
# SVM L1
# ======
svms = Methods(*[
SVM(dual=False, class_weight='auto', penalty="l1", C=C) for C in C_values
])
cv = CV(svms, cv_type="stratified", n_folds=10)
cv.run(X=X, y=y)
cv_results = cv.reduce()
#print cv_results
epac.export_csv(
cv, cv_results,
os.path.join(WD, "results", "cv10_caarms+pas+canabis_svmsl1.csv"))
# SVM L1 with CVBestSearchRefit
# =============================
svms_cv = CVBestSearchRefit(svms, n_folds=10, cv_type="stratified")
def test_constructor_avoid_collision_level1(self):
# Test that level 1 collisions are avoided
pm = Methods(*[SVC(kernel="linear", C=C) for C in [1, 10]])
leaves_key = [l.get_key() for l in pm.walk_leaves()]
self.assertTrue(len(leaves_key) == len(set(leaves_key)), u"Collision could not be avoided")
def test_mysvc_reducer(self):
## 1) Build dataset
## ===================================================================
X, y = datasets.make_classification(n_samples=12,
n_features=10,
n_informative=2,
random_state=1)
## 2) run with Methods
## ===================================================================
my_svc1 = MySVC(C=1.0)
my_svc2 = MySVC(C=2.0)
two_svc_single = Methods(my_svc1, my_svc2)
two_svc_local = Methods(my_svc1, my_svc2)
two_svc_swf = Methods(my_svc1, my_svc2)
two_svc_single.reducer = MyReducer()
two_svc_local.reducer = MyReducer()
two_svc_swf.reducer = MyReducer()
for leaf in two_svc_single.walk_leaves():
print leaf.get_key()
for leaf in two_svc_local.walk_leaves():
print leaf.get_key()
for leaf in two_svc_swf.walk_leaves():
print leaf.get_key()
# top-down process to call transform
two_svc_single.run(X=X, y=y)
# buttom-up process to compute scores
res_single = two_svc_single.reduce()
### You can get below results:
### ==================================================================
### [{'MySVC(C=1.0)': array([ 1., 1.])}, {'MySVC(C=2.0)': array([ 1., 1.])}]
### 3) Run using local multi-processes
### ==================================================================
from epac.map_reduce.engine import LocalEngine
local_engine = LocalEngine(two_svc_local, num_processes=2)
two_svc_local = local_engine.run(**dict(X=X, y=y))
res_local = two_svc_local.reduce()
### 4) Run using soma-workflow
### ==================================================================
from epac.map_reduce.engine import SomaWorkflowEngine
sfw_engine = SomaWorkflowEngine(tree_root=two_svc_swf,
num_processes=2)
two_svc_swf = sfw_engine.run(**dict(X=X, y=y))
res_swf = two_svc_swf.reduce()
if not repr(res_swf) == repr(res_local):
raise ValueError("Cannot dump class definition")
if not repr(res_swf) == repr(res_single):
raise ValueError("Cannot dump class definition")
def todo_perm_cv_grid_vs_sklearn(self):
X, y = datasets.make_classification(n_samples=100,
n_features=500,
n_informative=5)
n_perms = 3
n_folds = 2
n_folds_nested = 2
random_state = 0
k_values = [2, 3]
C_values = [1, 10]
# = With EPAC
pipelines = Methods(*[Pipe(SelectKBest(k=k),
SVC(C=C, kernel="linear"))
for C in C_values
for k in k_values])
#print [n for n in pipelines.walk_leaves()]
pipelines_cv = CVBestSearchRefit(pipelines,
n_folds=n_folds_nested,
random_state=random_state)
wf = Perms(CV(pipelines_cv, n_folds=n_folds,
reducer=ClassificationReport(keep=True)),
n_perms=n_perms, permute="y",
reducer=PvalPerms(keep=True),
random_state=random_state)
wf.fit_predict(X=X, y=y)
r_epac = wf.reduce().values()[0]
for key in r_epac:
print("key=" + repr(key) + ", value=" + repr(r_epac[key]))
# = With SKLEARN
from sklearn.cross_validation import StratifiedKFold
from epac.sklearn_plugins import Permutations
from sklearn.pipeline import Pipeline
from sklearn import grid_search
clf = Pipeline([('anova', SelectKBest(k=3)),
('svm', SVC(kernel="linear"))])
parameters = {'anova__k': k_values, 'svm__C': C_values}
r_sklearn = dict()
r_sklearn['pred_te'] = [[None] * n_folds for i in range(n_perms)]
r_sklearn['true_te'] = [[None] * n_folds for i in range(n_perms)]
r_sklearn['score_tr'] = [[None] * n_folds for i in range(n_perms)]
r_sklearn['score_te'] = [[None] * n_folds for i in range(n_perms)]
r_sklearn['mean_score_te'] = [None] * n_perms
r_sklearn['mean_score_tr'] = [None] * n_perms
perm_nb = 0
perms = Permutations(n=y.shape[0],
n_perms=n_perms,
random_state=random_state)
for idx in perms:
#idx = perms.__iter__().next()
y_p = y[idx]
cv = StratifiedKFold(y=y_p, n_folds=n_folds)
fold_nb = 0
for idx_train, idx_test in cv:
#idx_train, idx_test = cv.__iter__().next()
X_train = X[idx_train, :]
X_test = X[idx_test, :]
y_p_train = y_p[idx_train, :]
y_p_test = y_p[idx_test, :]
# Nested CV
cv_nested = StratifiedKFold(y=y_p_train,
n_folds=n_folds_nested)
gscv = grid_search.GridSearchCV(clf, parameters, cv=cv_nested)
gscv.fit(X_train, y_p_train)
r_sklearn['pred_te'][perm_nb][fold_nb] = gscv.predict(X_test)
r_sklearn['true_te'][perm_nb][fold_nb] = y_p_test
r_sklearn['score_tr'][perm_nb][fold_nb] =\
gscv.score(X_train, y_p_train)
r_sklearn['score_te'][perm_nb][fold_nb] =\
gscv.score(X_test, y_p_test)
fold_nb += 1
# Average over folds
r_sklearn['mean_score_te'][perm_nb] = \
np.mean(np.asarray(r_sklearn['score_te'][perm_nb]), axis=0)
r_sklearn['mean_score_tr'][perm_nb] = \
np.mean(np.asarray(r_sklearn['score_tr'][perm_nb]), axis=0)
#np.mean(R2[key]['score_tr'][perm_nb])
perm_nb += 1
print(repr(r_sklearn))
# - Comparisons
shared_keys = set(r_epac.keys()).intersection(set(r_sklearn.keys()))
comp = {k: np.all(np.asarray(r_epac[k]) == np.asarray(r_sklearn[k]))
for k in shared_keys}
print("comp=" + repr(comp))
#return comp
for key in comp:
self.assertTrue(comp[key], u'Diff for attribute: "%s"' % key)
Xd, yd = IO.read_Xy(WD=WD) X = np.asarray(Xd) y = np.asarray(yd) # DO NOT scale #X -= X.mean(axis=0) #X /= X.std(axis=0) #k_values = [1, 2, 3, 4, 5, 10, 15, 20, 25, 27] C_values = [0.01, 0.05, .1, .5, 1, 5, 10] # SVM L1 # ====== svms = Methods(*[SVM(dual=False, class_weight='auto', penalty="l1", C=C) for C in C_values]) # #anova_svms = Methods(*[Pipe(SelectKBest(k=k), #preprocessing.StandardScaler(), # Methods(*[SVM(C=C, penalty=penalty, class_weight='auto', dual=False) for C in C_values for penalty in ['l1', 'l2']])) for k in k_values]) cv = CV(svms, cv_type="stratified", n_folds=10) cv.run(X=X, y=y) cv_results = cv.reduce() #print cv_results epac.export_csv(cv, cv_results, os.path.join(WD, "results", "cv10_svmsl1.csv")) # SVM L1 with CVBestSearchRefit # =============================
"""
from sklearn import datasets
X, y = datasets.make_classification(n_samples=500,
n_features=200000,
n_informative=2,
random_state=1)
Xy = dict(X=X, y=y)
## 2) Building workflow
## =======================================================
print " -> Pt2 : X and y created, building workflow"
from sklearn import svm, cross_validation
#kfold = cross_validation.KFold(n=len(X), n_folds=3)
#svc = svm.SVC(C=1, kernel='linear')
#print [svc.fit(X[train], y[train]).score(X[test], y[test]) for train, test in kfold]
from epac import CV, Methods
cv_svm_local = CV(Methods(*[svm.SVC(kernel="linear"),
svm.SVC(kernel="rbf")]),
n_folds=3)
print " -> Pt3 : Workflow built, defining local engine"
cv_svm = None
n_proc = 2
# Running on the local machine
from epac import LocalEngine
local_engine = LocalEngine(cv_svm_local, num_processes=n_proc)
print " -> Pt4 : Running"
cv_svm = local_engine.run(**Xy)
print " -> Success with %i procs!" % n_proc
return {"y/pred": pred_y, "y/true": y, "best_beta": self.v_beta}
if __name__ == "__main__":
## 1) Build dataset
## ================================================
X, y = datasets.make_classification(n_samples=10,
n_features=5,
n_informative=2,
random_state=1)
Xy = {"X": X, "y": y}
## 2) Build Methods
## ================================================
print "Methods ==================================="
methods = Methods(*[TOY_CLF(v_lambda=v_lambda)
for v_lambda in [2, 1]])
print methods.run(**Xy)
## 3) Build WarmStartMethods like Methods
## ================================================
## WarmStartMethods
## / \
## TOY_CLF(v_lambda=2) TOY_CLF(v_lambda=1)
##
## 1. WarmStartMethods will look for different argumenets as signature
## For example, here is v_lambda, there are different for each leaf
## 2. And then run TOY_CLF(v_lambda=2).transform
## 3. Except v_lambda, WarmStartMethods copy all the other parameters
## from TOY_CLF(v_lambda=2) to TOY_CLF(v_lambda=1) as initialization
## 4. Finally call TOY_CLF(v_lambda=1).transform
print "WarmStartMethods =========================="
def do_all(options):
if options.k_max != "auto":
k_values = range_log2(np.minimum(int(options.k_max),
options.n_features),
add_n=True)
else:
k_values = range_log2(options.n_features, add_n=True)
C_values = [1, 10]
random_state = 0
#print options
#sys.exit(0)
if options.trace:
from epac import conf
conf.TRACE_TOPDOWN = True
## 1) Build dataset
## ================
X, y = datasets.make_classification(n_samples=options.n_samples,
n_features=options.n_features,
n_informative=options.n_informative)
## 2) Build Workflow
## =================
time_start = time.time()
## CV + Grid search of a pipeline with a nested grid search
cls = Methods(*[
Pipe(SelectKBest(k=k), SVC(kernel="linear", C=C)) for C in C_values
for k in k_values
])
pipeline = CVBestSearchRefit(cls,
n_folds=options.n_folds_nested,
random_state=random_state)
wf = Perms(CV(pipeline, n_folds=options.n_folds),
n_perms=options.n_perms,
permute="y",
random_state=random_state)
print "Time ellapsed, tree construction:", time.time() - time_start
## 3) Export Workflow to soma_workflow_gui
## ===============
time_fit_predict = time.time()
if os.path.isdir(options.soma_workflow_dir):
shutil.rmtree(options.soma_workflow_dir)
sfw_engine = SomaWorkflowEngine(tree_root=wf,
num_processes=options.n_cores)
sfw_engine.export_to_gui(options.soma_workflow_dir, X=X, y=y)
print "Time ellapsed, fit predict:", time.time() - time_fit_predict
# ## 6) Load Epac tree & Reduce
# ## ==========================
reduce_filename = os.path.join(options.soma_workflow_dir, "reduce.py")
f = open(reduce_filename, 'w')
reduce_str = """from epac.map_reduce.engine import SomaWorkflowEngine
wf = SomaWorkflowEngine.load_from_gui("%s")
print wf.reduce()
""" % options.soma_workflow_dir
f.write(reduce_str)
f.close()
print "#First run\n"\
"soma_workflow_gui\n"\
"\t(1)Open %s\n"\
"\t(2)Submit\n"\
"\t(3)Transfer Input Files\n"\
"\t...wait...\n"\
"\t(4)Transfer Output Files\n"\
"#When done run:\npython %s" % (
os.path.join(options.soma_workflow_dir,
sfw_engine.open_me_by_soma_workflow_gui),
reduce_filename)
@author: ed203246
"""
from sklearn import datasets
from sklearn.svm import SVC
from sklearn.lda import LDA
from sklearn.feature_selection import SelectKBest
X, y = datasets.make_classification(n_samples=12,
n_features=10,
n_informative=2)
from epac import Methods, Pipe
self = Methods(*[
Pipe(SelectKBest(k=k), SVC(kernel=kernel, C=C))
for kernel in ("linear", "rbf") for C in [1, 10] for k in [1, 2]
])
self = Methods(
*[Pipe(SelectKBest(k=k), SVC(C=C)) for C in [1, 10] for k in [1, 2]])
import copy
self.fit_predict(X=X, y=y)
self.reduce()
[l.get_key() for l in svms.walk_nodes()]
[l.get_key(2)
for l in svms.walk_nodes()] # intermediary key collisions: trig aggregation
"""
# Model selection using CV: CV + Grid
# -----------------------------------------
from epac import CVBestSearchRefit
# CV + Grid search of a simple classifier