def computing_precise_vs_imprecise(in_path=None,
ell_optimal=0.1,
cv_n_fold=10,
seeds=None,
lib_path_server=None,
model_type_precise='lda',
model_type_imprecise='ilda',
scaling=True):
data = export_data_set('iris.data') if in_path is None else pd.read_csv(
in_path)
logger = create_logger("computing_precise_vs_imprecise", True)
logger.info('Training dataset and models (%s, %s, %s, %s)', in_path,
model_type_precise, model_type_imprecise, ell_optimal)
X = data.iloc[:, :-1].values
if scaling: X = normalize_minmax(X)
y = np.array(data.iloc[:, -1].tolist())
seeds = generate_seeds(cv_n_fold) if seeds is None else seeds
model_impr = __factory_model(model_type_imprecise,
init_matlab=True,
add_path_matlab=lib_path_server,
DEBUG=False)
model_prec = __factory_model_precise(model_type_precise,
store_covariance=True)
avg_imprecise, avg_precise, n_real_times = 0, 0, 0
for time in range(cv_n_fold):
kf = KFold(n_splits=cv_n_fold, random_state=seeds[time], shuffle=True)
imprecise_mean, precise_mean, n_real_fold = 0, 0, 0
for idx_train, idx_test in kf.split(y):
X_cv_train, y_cv_train = X[idx_train], y[idx_train]
X_cv_test, y_cv_test = X[idx_test], y[idx_test]
model_impr.learn(X=X_cv_train, y=y_cv_train, ell=ell_optimal)
model_prec.fit(X_cv_train, y_cv_train)
n_real_tests, time_precise, time_imprecise = 0, 0, 0
n_test, _ = X_cv_test.shape
for i, test in enumerate(X_cv_test):
evaluate_imp, _ = model_impr.evaluate(test)
evaluate = model_prec.predict([test])
if len(evaluate_imp) > 1:
n_real_tests += 1
if y_cv_test[i] in evaluate_imp: time_imprecise += 1
if y_cv_test[i] in evaluate: time_precise += 1
logger.debug(
"(time, iTest, ellOptimal, cautious, prediction, ground-truth)(%s, %s, %s, %s, %s, %s)",
time, i, ell_optimal, evaluate_imp, evaluate, y_cv_test[i])
logger.debug(
"(time, ellOptimal, nRealTests, timeImprecise, timePrecise) (%s, %s, %s, %s, %s)",
time, ell_optimal, n_real_tests, time_imprecise, time_precise)
if n_real_tests > 0:
n_real_fold += 1
imprecise_mean += time_imprecise / n_real_tests
precise_mean += time_precise / n_real_tests
logger.debug("(time, nRealFold, imprecise, precise) (%s, %s, %s, %s)",
time, n_real_fold, imprecise_mean, precise_mean)
if n_real_fold > 0:
n_real_times += 1
avg_imprecise += imprecise_mean / n_real_fold
avg_precise += precise_mean / n_real_fold
logger.debug("(dataset, models, imprec, prec) (%s, %s, %s, %s, %s)",
in_path, model_type_imprecise, model_type_precise,
avg_imprecise / n_real_times, avg_precise / n_real_times)
def performance_hold_out(in_path=None,
out_path=None,
model_type='lda',
test_pct=0.4,
n_times=10,
seeds=None,
scaling=False):
assert os.path.exists(in_path), "Without training data, not testing"
assert os.path.exists(out_path), "Without output saving performance"
logger = create_logger("performance_hold_out", True)
logger.info('Training data set %s, test percentage %s, model_type %s',
in_path, test_pct, model_type)
data = pd.read_csv(in_path, header=None)
X = data.iloc[:, :-1].values
if scaling: X = normalize_minmax(X)
y = data.iloc[:, -1].tolist()
seeds = generate_seeds(n_times) if seeds is None else seeds
logger.info('Seeds generated %s', seeds)
file_csv = open(out_path, 'w')
writer = csv.writer(file_csv)
model = __factory_model_precise(model_type, store_covariance=True)
mean_u65, mean_u80 = np.array([]), np.array([])
for i in range(0, n_times):
X_train, X_test, y_train, y_test = train_test_split(
X, y, test_size=test_pct, random_state=seeds[i])
sum_u65, sum_u80 = 0, 0
model.fit(X_train, y_train)
n, _ = X_test.shape
for j, test in enumerate(X_test):
evaluate = model.predict([test])
if y_test[j] in evaluate:
sum_u65 += u65(evaluate)
sum_u80 += u80(evaluate)
logger.info("time, u65, u80 (%s, %s, %s)", i, sum_u65 / n, sum_u80 / n)
mean_u65 = np.append(mean_u65, sum_u65 / n)
mean_u80 = np.append(mean_u80, sum_u80 / n)
writer.writerow([-999, i, mean_u65[i], mean_u80[i]])
file_csv.flush()
file_csv.close()
logger.info("[total:data-set:avgResults] (%s, %s)", np.mean(mean_u65),
np.mean(mean_u80))
def performance_accuracy_hold_out(in_path=None,
model_type="ilda",
ell_optimal=0.1,
lib_path_server=None,
seeds=None,
DEBUG=False,
scaling=False):
assert os.path.exists(
in_path
), "Without training data, cannot performing cross hold-out accuracy"
logger = create_logger("performance_accuracy_hold_out", True)
logger.info('Training dataset (%s, %s, %s)', in_path, model_type,
ell_optimal)
X, y = dataset_to_Xy(in_path, scaling=scaling)
seeds = generate_seeds(cv_n_fold) if seeds is None else seeds
logger.info('Seeds used for accuracy %s', seeds)
n_time = len(seeds)
mean_u65, mean_u80 = 0, 0
model = __factory_model(model_type,
solver_matlab=True,
add_path_matlab=lib_path_server,
DEBUG=DEBUG)
for k in range(0, n_time):
X_train, X_test, y_train, y_test = train_test_split(
X, y, test_size=0.4, random_state=seeds[k])
model.learn(X=X_cv_train, y=y_cv_train, ell=ell_optimal)
sum_u65, sum_u80 = 0, 0
n_test, _ = X_test.shape
for i, test in enumerate(X_test):
evaluate = lqa.evaluate(test)
logger.debug(
"(testing, ell_current, prediction, ground-truth) (%s, %s, %s, %s)",
i, ell_optimal, evaluate, y_test[i])
if y_test[i] in evaluate:
sum_u65 += u65(evaluate)
sum_u80 += u80(evaluate)
logger.debug("Partial-kfold (%s, %s, %s, %s)", ell_current, k,
sum_u65 / n_test, sum_u80 / n_test)
mean_u65 += sum_u65 / n_test
mean_u80 += sum_u80 / n_test
mean_u65 = mean_u65 / n_time
mean_u80 = mean_u80 / n_time
logger.debug("Total-ell (%s, %s, %s, %s)", in_path, ell_optimal, mean_u65,
mean_u80)
def performance_cv_accuracy(in_path=None,
model_type='lda',
cv_n_fold=10,
seeds=None,
scaling=False):
assert os.path.exists(in_path), "Without training data, not testing"
data = pd.read_csv(in_path, header=None)
logger = create_logger("performance_cv_accuracy", True)
logger.info('Training data set %s, cv_n_fold %s, model_type %s', in_path,
cv_n_fold, model_type)
X = data.iloc[:, :-1].values
if scaling: X = normalize_minmax(X)
y = np.array(data.iloc[:, -1].tolist())
avg_u65, avg_u80 = 0, 0
seeds = generate_seeds(cv_n_fold) if seeds is None else seeds
logger.info('Seeds generated %s', seeds)
for time in range(cv_n_fold):
# Generation a random k-fold validation.
kf = KFold(n_splits=cv_n_fold, random_state=seeds[time], shuffle=True)
model = __factory_model_precise(model_type, store_covariance=True)
mean_u65, mean_u80 = 0, 0
for idx_train, idx_test in kf.split(y):
X_cv_train, y_cv_train = X[idx_train], y[idx_train]
X_cv_test, y_cv_test = X[idx_test], y[idx_test]
model.fit(X_cv_train, y_cv_train)
n_test = len(idx_test)
sum_u65, sum_u80 = 0, 0
for i, test in enumerate(X_cv_test):
evaluate = model.predict([test])
logger.debug(
"(testing, prediction, ground-truth) (%s, %s, %s)", i,
evaluate, y_cv_test[i])
if y_cv_test[i] in evaluate:
sum_u65 += u65(evaluate)
sum_u80 += u80(evaluate)
mean_u65 += sum_u65 / n_test
mean_u80 += sum_u80 / n_test
logger.info("Time, seed, u65, u80 (%s, %s, %s, %s)", time, seeds[time],
mean_u65 / cv_n_fold, mean_u80 / cv_n_fold)
avg_u65 += mean_u65 / cv_n_fold
avg_u80 += mean_u80 / cv_n_fold
logger.info("[Total:data-set:avgResults] (%s, %s, %s)", in_path,
avg_u65 / cv_n_fold, avg_u80 / cv_n_fold)
def performance_cv_accuracy_imprecise(in_path=None,
model_type="ilda",
ell_optimal=0.1,
nb_process=2,
lib_path_server=None,
cv_n_fold=10,
seeds=None,
criterion="maximality"):
assert os.path.exists(in_path), "Without training data, not testing"
data = pd.read_csv(in_path)
logger = create_logger("performance_cv_accuracy_imprecise", True)
logger.info('Training dataset (%s, %s, %s, %s)', in_path, model_type,
ell_optimal, criterion)
X = data.iloc[:, :-1].values
y = np.array(data.iloc[:, -1].tolist())
avg_u65, avg_u80 = 0, 0
seeds = generate_seeds(cv_n_fold) if seeds is None else seeds
logger.info('Seeds used for accuracy %s', seeds)
manager = ManagerWorkers(nb_process=nb_process, criterion=criterion)
manager.executeAsync(model_type, lib_path_server)
for time in range(cv_n_fold):
kf = KFold(n_splits=cv_n_fold, random_state=seeds[time], shuffle=True)
mean_u65, mean_u80 = 0, 0
for idx_train, idx_test in kf.split(y):
logger.info("Splits train %s", idx_train)
logger.info("Splits test %s", idx_test)
X_cv_train, y_cv_train = X[idx_train], y[idx_train]
X_cv_test, y_cv_test = X[idx_test], y[idx_test]
mean_u65, mean_u80 = computing_training_testing_step(
X_cv_train, y_cv_train, X_cv_test, y_cv_test, ell_optimal,
manager, mean_u65, mean_u80)
logger.debug("Partial-kfold (%s, %s, %s, %s)", ell_optimal, time,
mean_u65, mean_u80)
logger.info("Time, seed, u65, u80 (%s, %s, %s, %s)", time, seeds[time],
mean_u65 / cv_n_fold, mean_u80 / cv_n_fold)
avg_u65 += mean_u65 / cv_n_fold
avg_u80 += mean_u80 / cv_n_fold
manager.poisonPillTraining()
logger.debug("total-ell (%s, %s, %s, %s)", in_path, ell_optimal,
avg_u65 / cv_n_fold, avg_u80 / cv_n_fold)
def performance_cv_accuracy_imprecise(in_path=None,
model_type="ilda",
ell_optimal=0.1,
scaling=False,
lib_path_server=None,
cv_n_fold=10,
seeds=None,
nb_process=10):
assert os.path.exists(
in_path
), "Without training data, cannot performing cross validation accuracy"
logger = create_logger("performance_cv_accuracy_imprecise", True)
logger.info('Training dataset (%s, %s, %s)', in_path, model_type,
ell_optimal)
X, y = dataset_to_Xy(in_path, scaling=scaling)
avg_u65, avg_u80 = 0, 0
seeds = generate_seeds(cv_n_fold) if seeds is None else seeds
logger.info('Seeds used for accuracy %s', seeds)
manager = ManagerWorkers(nb_process=nb_process)
manager.executeAsync(model_type, lib_path_server)
for time in range(cv_n_fold):
kf = KFold(n_splits=cv_n_fold, random_state=seeds[time], shuffle=True)
mean_u65, mean_u80 = 0, 0
for idx_train, idx_test in kf.split(y):
mean_u65, mean_u80, _ = computing_training_testing_step(
X[idx_train], y[idx_train], X[idx_test], y[idx_test],
ell_optimal, manager, mean_u65, mean_u80)
logger.debug("Partial-kfold (%s, %s, %s, %s)", ell_optimal, time,
mean_u65, mean_u80)
logger.info("Time, seed, u65, u80 (%s, %s, %s, %s)", time, seeds[time],
mean_u65 / cv_n_fold, mean_u80 / cv_n_fold)
avg_u65 += mean_u65 / cv_n_fold
avg_u80 += mean_u80 / cv_n_fold
manager.poisonPillTraining()
logger.debug("Total-ell (%s, %s, %s, %s)", in_path, ell_optimal,
avg_u65 / cv_n_fold, avg_u80 / cv_n_fold)
def computing_time_prediction(in_path=None,
ell_optimal=0.1,
lib_path_server=None,
model_type="ilda",
criterion="maximality",
k_repetition=10,
seeds=None):
assert os.path.exists(in_path), "Without training data, not testing"
data = pd.read_csv(in_path, header=None)
logger = create_logger("computing_time_prediction", True)
X = data.iloc[:, :-1].values
y = data.iloc[:, -1].tolist()
seeds = generate_seeds(k_repetition) if seeds is None else seeds
logger.info(
'Training dataset %s with maximality version (%s) and model (%s), ell_optimal (%s) and seeds %s',
in_path, criterion, model_type, ell_optimal, seeds)
model = __factory_model(model_type,
solver_matlab=True,
add_path_matlab=lib_path_server,
DEBUG=False)
avg = np.array([])
for k in range(k_repetition):
logger.info("%s-fold repetition randomly, seed %s", k, seeds[k])
X_train, X_test, y_train, y_test = train_test_split(
X, y, test_size=0.2, random_state=seeds[k])
model.learn(X=X_train, y=y_train, ell=ell_optimal)
n, _ = X_test.shape
sum_time = 0
for i, test in enumerate(X_test):
start = time.time()
evaluate = model.evaluate(test, criterion=criterion)
end = time.time()
logger.info("Evaluate %s, Ground-truth %s, Time %s ", evaluate,
y_test[i], (end - start))
sum_time += (end - start)
avg = np.append(avg, sum_time / n)
logger.info("Total time (%s, %s) and average %s and sd %s of %s testing",
in_path, avg, np.mean(avg), np.std(avg), n)
def performance_qda_regularized(in_path=None,
out_path=None,
cv_n_fold=10,
seeds=None,
from_alpha=0,
to_alpha=2.0,
by_alpha=0.01,
scaling=False):
assert os.path.exists(in_path), "Without training data, not testing"
assert os.path.exists(out_path), "Without output saving performance"
data = pd.read_csv(in_path, header=None)
logger = create_logger("performance_qda_regularized", True)
logger.info('Training data set %s, cv_n_fold %s, model_type %s', in_path,
cv_n_fold, "qda")
X = data.iloc[:, :-1].values
if scaling: X = normalize_minmax(X)
y = np.array(data.iloc[:, -1].tolist())
seeds = generate_seeds(cv_n_fold) if seeds is None else seeds
logger.info('Seeds generated %s', seeds)
file_csv = open(out_path, 'a')
writer = csv.writer(file_csv)
alphas = np.arange(from_alpha, to_alpha, by_alpha)
writer.writerow(alphas)
qda_regularized = [None] * len(alphas)
for idx, alpha in enumerate(alphas):
qda_regularized[idx] = __factory_model_precise("qda",
store_covariance=True,
reg_param=alpha)
# Generation a random k-fold validation.
kf_second = KFold(n_splits=cv_n_fold, random_state=None, shuffle=True)
ikfold, accuracy, best_alphas = 0, [0] * cv_n_fold, [0] * cv_n_fold
for idx_learning, idx_testing in kf_second.split(y):
X_training, y_training = X[idx_learning], y[idx_learning]
X_testing, y_testing = X[idx_testing], y[idx_testing]
kf = KFold(n_splits=cv_n_fold,
random_state=seeds[ikfold],
shuffle=True)
acc_u80 = [0] * len(qda_regularized)
for idx_train, idx_test in kf.split(y_training):
X_cv_train, y_cv_train = X_training[idx_train], y_training[
idx_train]
X_cv_test, y_cv_test = X_training[idx_test], y_training[idx_test]
for model in qda_regularized:
model.fit(X_cv_train, y_cv_train)
n_test = len(idx_test)
for i, test in enumerate(X_cv_test):
for im, model in enumerate(qda_regularized):
evaluate = model.predict([test])
if y_cv_test[i] in evaluate:
acc_u80[im] += (u80(evaluate) / n_test) / cv_n_fold
idx_best = np.argmax(acc_u80)
logger.info("[1kfold:best_model:seed:u80] (%s, %s, %s, %s)", ikfold,
alphas[idx_best], seeds[ikfold], acc_u80)
writer.writerow(acc_u80)
file_csv.flush()
best_model = __factory_model_precise("qda",
store_covariance=True,
reg_param=alphas[idx_best])
best_model.fit(X_training, y_training)
accuracy[ikfold], bn_test, best_alphas[ikfold] = 0, len(
idx_testing), alphas[idx_best]
for i, test in enumerate(X_testing):
evaluate = best_model.predict([test])
if y_testing[i] in evaluate:
accuracy[ikfold] += u80(evaluate) / bn_test
logger.info("[2kfold:best_model:seed:accuracy] (%s, %s, %s)", ikfold,
alphas[idx_best], accuracy[ikfold])
ikfold += 1
file_csv.close()
logger.info("[total:data-set:avgResults] (%s, %s, %s, %s)", in_path,
np.mean(accuracy), best_alphas, accuracy)
def computing_best_imprecise_mean(in_path=None, out_path=None, lib_path_server=None, model_type="ilda",
from_ell=0.1, to_ell=1.0, by_ell=0.1, seed=None, cv_kfold_first=10,
nb_process=2, skip_nfold=0, cv_kfold_second=10, seed_second=None, scaling=False):
assert os.path.exists(in_path), "Without training data, not testing"
assert os.path.exists(out_path), "File for putting results does not exist"
logger = create_logger("computing_best_imprecise_mean_cv", True)
logger.info('Training dataset (%s, %s, %s)', in_path, out_path, model_type)
logger.info('Parameters (ells, nbProcess, skip_nfold, cv_kfold_second) (%s, %s, %s, %s, %s, %s)', from_ell,
to_ell, by_ell, nb_process, skip_nfold, cv_kfold_second)
data = pd.read_csv(in_path, header=None)
X = data.iloc[:, :-1].values
if scaling: X = normalize_minmax(X)
y = np.array(data.iloc[:, -1].tolist())
# Seeding a random value for k-fold top learning-testing data
seed = random.randrange(pow(2, 30)) if seed is None else seed
logger.debug("[FIRST-STEP-SEED] MODEL: %s, SEED: %s", model_type, seed)
# Create a CSV file for saving results
file_csv = open(out_path, 'a')
writer = csv.writer(file_csv)
manager = ManagerWorkers(nb_process=nb_process)
manager.executeAsync(model_type, lib_path_server)
kfFirst = KFold(n_splits=cv_kfold_first, random_state=seed, shuffle=True)
acc_u80, acc_u65, idx_kfold = dict(), dict(), 0
seed_2step = generate_seeds(cv_kfold_second) if seed_second is None else seed_second
logger.debug("[SECOND-STEP-SEEDS] MODEL: %s, SEED: %s, SECOND-SEED: %s", model_type, seed, seed_2step)
for idx_learning, idx_testing in kfFirst.split(y):
ell_u65, ell_u80 = dict(), dict()
# Generate sampling k-fold (learning, testing) for optimal ell parameters
X_learning, y_learning = X[idx_learning], y[idx_learning]
X_testing, y_testing = X[idx_testing], y[idx_testing]
logger.info("Splits %s learning %s", idx_kfold, idx_learning)
logger.info("Splits %s testing %s", idx_kfold, idx_testing)
# # n-Skipping sampling and reboot parameter from_ell to 0.01 next sampling
if skip_nfold != 0 and idx_kfold > skip_nfold:
from_ell = 0.01
# n-Skipping fold cross-validation (purpose for parallel computing)
if idx_kfold >= skip_nfold:
# Generate same k-fold-second (train, test) for impartially computing accuracy all ell parameters
splits_ell = list([])
logger.debug("[2-STEP-SEED] MODEL: %s, SEED: %s OF FIRST STEP %s", model_type, seed_2step[idx_kfold], seed)
kfSecond = KFold(n_splits=cv_kfold_second, random_state=seed_2step[idx_kfold], shuffle=True)
for idx_learn_train, idx_learn_test in kfSecond.split(y_learning):
splits_ell.append((idx_learn_train, idx_learn_test))
logger.info("Splits %s train %s", len(splits_ell), idx_learn_train)
logger.info("Splits %s test %s", len(splits_ell), idx_learn_test)
for ell_current in np.arange(from_ell, to_ell, by_ell):
ell_u65[ell_current], ell_u80[ell_current] = 0, 0
logger.info("ELL_CURRENT %s", ell_current)
for idx_learn_train, idx_learn_test in splits_ell:
logger.info("Splits step train %s", idx_learn_train)
logger.info("Splits step test %s", idx_learn_test)
X_cv_train, y_cv_train = X_learning[idx_learn_train], y_learning[idx_learn_train]
X_cv_test, y_cv_test = X_learning[idx_learn_test], y_learning[idx_learn_test]
ell_u65[ell_current], ell_u80[ell_current], _ = \
computing_training_testing_step(X_cv_train, y_cv_train, X_cv_test, y_cv_test, ell_current,
manager, ell_u65[ell_current], ell_u80[ell_current])
logger.info("Partial-kfold (%s, %s, %s)", ell_current, ell_u65[ell_current], ell_u80[ell_current])
ell_u65[ell_current] = ell_u65[ell_current] / cv_kfold_first
ell_u80[ell_current] = ell_u80[ell_current] / cv_kfold_first
writer.writerow([ell_current, idx_kfold, ell_u65[ell_current], ell_u80[ell_current]])
file_csv.flush()
logger.debug("Partial-ell-k-step (%s, %s, %s)", idx_kfold, ell_u65[ell_current], ell_u80[ell_current])
logger.debug("Total-ell-k-step (%s, %s, %s, %s)", in_path, idx_kfold, ell_u65, ell_u80)
# Computing optimal ells for using in testing step
acc_ell_u80 = max(ell_u80.values())
acc_ell_u65 = max(ell_u65.values())
ell_u80_opts = [k for k, v in ell_u80.items() if v == acc_ell_u80]
ell_u65_opts = [k for k, v in ell_u65.items() if v == acc_ell_u65]
acc_u65[idx_kfold], acc_u80[idx_kfold] = 0, 0
n_ell80_opts, n_ell65_opts = len(ell_u80_opts), len(ell_u65_opts)
for ell_u80_opt in ell_u80_opts:
logger.info("ELL_OPTIMAL_CV_U80 %s", ell_u80_opt)
_, _acc_u80, _ = \
computing_training_testing_step(X_learning, y_learning, X_testing, y_testing, ell_u80_opt,
manager, 0, 0)
acc_u80[idx_kfold] += _acc_u80
writer.writerow([-999, -8, ell_u80_opt, _acc_u80])
for ell_u65_opt in ell_u65_opts:
logger.info("ELL_OPTIMAL_CV_U65 %s", ell_u65_opt)
_acc_u65, _, _ = \
computing_training_testing_step(X_learning, y_learning, X_testing, y_testing, ell_u65_opt,
manager, 0, 0)
acc_u65[idx_kfold] += _acc_u65
writer.writerow([-999, -7, ell_u65_opt, _acc_u65])
acc_u65[idx_kfold] = acc_u65[idx_kfold] / n_ell65_opts
acc_u80[idx_kfold] = acc_u80[idx_kfold] / n_ell80_opts
writer.writerow([-999, idx_kfold, acc_u65[idx_kfold], acc_u80[idx_kfold]])
file_csv.flush()
logger.debug("Partial-ell-2step (u80, u65, accs) (%s, %s, %s, %s, %s)", -999, ell_u80_opts, ell_u65_opts,
acc_u65[idx_kfold], acc_u80[idx_kfold])
idx_kfold += 1
writer.writerow([-9999, -9, np.mean(list(acc_u65.values())), np.mean(list(acc_u80.values()))])
manager.poisonPillTraining()
file_csv.close()
logger.debug("Total-accuracy (%s, %s, %s)", in_path, acc_u65, acc_u80)
logger.debug("Total-avg-accuracy (%s, %s, %s)", in_path, np.mean(list(acc_u65.values())),
np.mean(list(acc_u80.values())))
def computing_best_imprecise_mean(in_path=None,
out_path=None,
cv_nfold=10,
model_type="ilda",
test_size=0.4,
from_ell=0.1,
to_ell=1.0,
by_ell=0.1,
seeds=None,
lib_path_server=None,
nb_process=2,
n_sampling=10,
skip_n_sample=0,
criterion="maximality",
scaling=False):
assert os.path.exists(in_path), "Without training data, not testing"
assert os.path.exists(out_path), "File for putting results does not exist"
logger = create_logger("computing_best_imprecise_mean_sampling", True)
logger.info('Training dataset (%s, %s, %s)', in_path, model_type,
criterion)
logger.info(
'Parameters (size, ells, nbProcess, sampling, nSkip) (%s, %s, %s, %s, %s, %s, %s)',
test_size, from_ell, to_ell, by_ell, nb_process, n_sampling,
skip_n_sample)
data = pd.read_csv(in_path, header=None)
X = data.iloc[:, :-1].values
if scaling: X = normalize_minmax(X)
y = np.array(data.iloc[:, -1].tolist())
# Seed for get back up if process is killed
seeds = generate_seeds(n_sampling) if seeds is None else seeds
logger.debug("MODEL: %s, SEED: %s", model_type, seeds)
# Create a CSV file for saving results
file_csv = open(out_path, 'a')
writer = csv.writer(file_csv)
manager = ManagerWorkers(nb_process=nb_process, criterion=criterion)
manager.executeAsync(model_type, lib_path_server)
acc_u80, acc_u65 = dict(), dict()
for sampling in range(min(n_sampling, len(seeds))):
X_learning, X_testing, y_learning, y_testing = \
train_test_split(X, y, test_size=test_size, random_state=seeds[sampling])
logger.info("Splits %s learning %s", sampling, y_learning)
logger.info("Splits %s testing %s", sampling, y_testing)
# n-Skipping sampling and reboot parameter from_ell to 0.01 next sampling
if skip_n_sample != 0 and sampling > skip_n_sample: from_ell = 0.01
# n-Skipping sampling testing (purpose for parallel computing)
if sampling >= skip_n_sample:
kf = KFold(n_splits=cv_nfold, random_state=None, shuffle=True)
ell_u65, ell_u80, splits = dict(), dict(), list([])
for idx_train, idx_test in kf.split(y_learning):
splits.append((idx_train, idx_test))
logger.info("Sampling %s Splits %s train %s", sampling,
len(splits), idx_train)
logger.info("Sampling %s Splits %s test %s", sampling,
len(splits), idx_test)
for ell_current in np.arange(from_ell, to_ell, by_ell):
ell_u65[ell_current], ell_u80[ell_current] = 0, 0
logger.info("ELL_CURRENT %s", ell_current)
for idx_train, idx_test in splits:
logger.info("Splits train %s", idx_train)
logger.info("Splits test %s", idx_test)
X_cv_train, y_cv_train = X_learning[idx_train], y_learning[
idx_train]
X_cv_test, y_cv_test = X_learning[idx_test], y_learning[
idx_test]
# Computing accuracy testing for cross-validation step
ell_u65[ell_current], ell_u80[ell_current] = \
computing_training_testing_step(X_cv_train, y_cv_train, X_cv_test, y_cv_test, ell_current,
manager, ell_u65[ell_current], ell_u80[ell_current])
logger.info("Partial-kfold (%s, %s, %s)", ell_current,
ell_u65[ell_current], ell_u80[ell_current])
ell_u65[ell_current] = ell_u65[ell_current] / cv_nfold
ell_u80[ell_current] = ell_u80[ell_current] / cv_nfold
writer.writerow([
ell_current, sampling, ell_u65[ell_current],
ell_u80[ell_current]
])
file_csv.flush()
logger.debug("Partial-ell-sampling (%s, %s, %s, %s)",
ell_current, sampling, ell_u65, ell_u80)
logger.debug("Total-ell-sampling (%s, %s, %s, %s)", in_path,
sampling, ell_u65, ell_u80)
# Computing optimal ells for using in testing step
acc_ellu80 = max(ell_u80.values())
acc_ellu65 = max(ell_u65.values())
ellu80_opts = [k for k, v in ell_u80.items() if v == acc_ellu80]
ellu65_opts = [k for k, v in ell_u65.items() if v == acc_ellu65]
acc_u65[sampling], acc_u80[sampling] = 0, 0
n_ell80_opts, n_ell65_opts = len(ellu80_opts), len(ellu65_opts)
for ellu80_opt in ellu80_opts:
logger.info("ELL_OPTIMAL_SAMPLING_U80 %s", ellu80_opt)
_, acc_u80[sampling] = \
computing_training_testing_step(X_learning, y_learning, X_testing, y_testing, ellu80_opt,
manager, 0, acc_u80[sampling])
for ellu65_opt in ellu65_opts:
logger.info("ELL_OPTIMAL_SAMPLING_U65 %s", ellu65_opt)
acc_u65[sampling], _ = \
computing_training_testing_step(X_learning, y_learning, X_testing, y_testing, ellu65_opt,
manager, acc_u65[sampling], 0)
acc_u65[sampling] = acc_u65[sampling] / n_ell65_opts
acc_u80[sampling] = acc_u80[sampling] / n_ell80_opts
writer.writerow(
[-999, sampling, acc_u65[sampling], acc_u80[sampling]])
file_csv.flush()
logger.debug("Partial-ell-2step (%s, %s, %s, %s)", -999,
ellu80_opts, acc_u65[sampling], acc_u80[sampling])
writer.writerow([
-9999, -9,
np.mean(list(acc_u65.values())),
np.mean(list(acc_u80.values()))
])
manager.poisonPillTraining()
file_csv.close()
logger.debug("Total-accuracy (%s, %s, %s)", in_path, acc_u65, acc_u80)
logger.debug("Total-avg-accuracy (%s, %s, %s)", in_path,
np.mean(list(acc_u65.values())),
np.mean(list(acc_u80.values())))
