def convert_data():
if train:
files = config.trainfiles
num = 16
else:
files = config.testfiles
num = 7
utils.parmap(convert_train_file, files, num)
def run(X, y, XTest, yTest, foldsNum, path, inds, jobsNum = 6, calcProbs = True):
# y = y.astype(np.int)
# yTest = yTest.astype(np.int)
# heldoutSize = 1./foldsNum
# X, idx = shuffle(X)
# y = y[idx]
# XTest, idx = shuffle(XTest)
# yTest = yTest[idx]
# scaler = preprocessing.StandardScaler().fit(X)
# X = scaler.transform(X)
# XTest = scaler.transform(XTest)
# X = X[:5000, :]
# y = y[:5000]
# ---------------------------------------------------- XTest = XTest[:500, :]
# ------------------------------------------------------- yTest = yTest[:500]
# cv = StratifiedShuffleSplit(y, foldsNum, heldoutSize, random_state=0)
cv = IndicesKFold(inds, 5)#, 4000, 1000)
# -------------------------------- scores = ['roc_auc','precision', 'recall']
# -------------- aucScoreFunc = make_scorer(aucScore, greater_is_better=True)
# scores = ['roc_auc'] #[aucScoreFunc] # 'roc_auc'
# ---------------------------------------------------- scores = ['precision']
calcProbs = True
print('Start the grid search')
t = time.time()
for tuned_param in tuned_parameters:
params = []
for fold_num, (train_index, test_index) in enumerate(cv):
params.append((X, y, train_index, test_index, tuned_param, fold_num, calcProbs))
print(tuned_param)
if (jobsNum == 1):
mapResults = [calc_cv_scores(p) for p in params] # For debugging
else:
mapResults = utils.parmap(calc_cv_scores, params, jobsNum)
cv_scores = np.array([score for (clf, score) in mapResults][:len(cv)])
print(cv_scores)
clf = mapResults[0][0]
elapsed = time.time() - t
print('Request took '+str(elapsed)+' sec.')
print(str(datetime.now()))
scaler = preprocessing.StandardScaler().fit(X)
X = scaler.transform(X)
XTest = scaler.transform(XTest)
printResults(clf, XTest, yTest, calcProbs, path,cv_scores)
def run(x, y, x_test, y_test, folds_num, path, inds, jobs_num=6, calc_probs=True):
# y = y.astype(np.int)
# y_test = y_test.astype(np.int)
# heldoutSize = 1./folds_num
# x, idx = shuffle(x)
# y = y[idx]
# x_test, idx = shuffle(x_test)
# y_test = y_test[idx]
# scaler = preprocessing.StandardScaler().fit(x)
# x = scaler.transform(x)
# x_test = scaler.transform(x_test)
# x = x[:5000, :]
# y = y[:5000]
# ---------------------------------------------------- x_test = x_test[:500, :]
# ------------------------------------------------------- y_test = y_test[:500]
# cv = StratifiedShuffleSplit(y, folds_num, heldoutSize, random_state=0)
cv = IndicesKFold(inds, folds_num) # , 4000, 1000)
# -------------------------------- scores = ['roc_auc','precision', 'recall']
# -------------- auc_scoreFunc = make_scorer(auc_score, greater_is_better=True)
# scores = ['roc_auc'] #[auc_scoreFunc] # 'roc_auc'
# ---------------------------------------------------- scores = ['precision']
# calc_probs = True
parts_of_path = path.split('/')
dump_path = os.path.join(DUMP_FOLDER, parts_of_path[-1])
if not os.path.exists(dump_path):
os.makedirs(dump_path)
current_dump_folder = dump_path
print('Start the grid search')
t = time.time()
for tuned_param in tuned_parameters:
already_exist = check_if_params_were_calculated(dump_path, tuned_param)
if already_exist:
continue
params = []
for fold_num, (train_index, test_index) in enumerate(cv):
params.append((x, y, train_index, test_index, tuned_param, fold_num, calc_probs, inds))
print(tuned_param)
if jobs_num == 1:
map_results = [calc_cv_scores(p) for p in params] # For debugging
else:
map_results = utils.parmap(calc_cv_scores, params, jobs_num)
cv_scores = np.array([(score,scoreTrial) for (clf, score,scoreTrial) in map_results][:len(cv)])
print(cv_scores)
mean_cv_score = sum(cv_scores)/len(cv_scores)
print('==============================mean auc score is '+str(mean_cv_score)+'==============================')
clf = map_results[0][0]
elapsed = time.time() - t
print('Request took '+str(elapsed)+' sec.')
print(str(datetime.now()))
# scaler = preprocessing.StandardScaler().fit(x)
# x = scaler.transform(x)
# x_test = scaler.transform(x_test)
mini_path = path.split('/')[-1]
mini_path = mini_path.replace(' ', '_')
print_results(clf, x_test, y_test, calc_probs, path, None, cv_scores, current_dump_folder, mean_cv_score)
sys.stdout.flush()
return 1 - test_score
def run_single(train_files_glob, i):
print '... reading'
sys.stdout.flush()
x_train, y_train = gen_vw.read_train_data(train_files_glob, i)
x_test, y_test = gen_vw.read_test_data(train_files_glob, i)
x_valid, y_valid = gen_vw.read_valid_data(train_files_glob, i)
print '... reading done'
sys.stdout.flush()
x = T.matrix('x')
y = T.ivector('y')
error = train_batch(x, y, x_train, y_train, x_valid, y_valid, x_test, y_test)
return 1 - error
if __name__ == '__main__':
#tfile = config.train_folder + "train_subject0[1-4].mat"
#print run_single(tfile, 0)
tfile = config.train_folder + "*.mat"
precisions = utils.parmap(lambda i: run_single(tfile, i), len(glob(tfile)) - 1)
print sum(precisions) / len(precisions)