def outer_cv_loop(Xdata,
Ydata,
clf,
parameters=[],
n_splits=10,
test_size=0.25):
pred = numpy.zeros(len(Ydata))
importances = []
kf = StratifiedShuffleSplit(n_splits=n_splits, test_size=test_size)
rocscores = []
for train, test in kf.split(Xdata, Ydata):
if numpy.var(Ydata[test]) == 0:
print('zero variance', varname)
rocscores.append(numpy.nan)
continue
Ytrain = Ydata[train]
Xtrain = fancyimpute.SoftImpute(verbose=False).complete(
Xdata[train, :])
Xtest = fancyimpute.SoftImpute(verbose=False).complete(Xdata[test, :])
if numpy.abs(numpy.mean(Ytrain) - 0.5) > 0.2:
smt = SMOTETomek()
Xtrain, Ytrain = smt.fit_sample(Xtrain.copy(), Ydata[train])
# filter out bad folds
clf.fit(Xtrain, Ytrain)
pred = clf.predict(Xtest)
if numpy.var(pred) > 0:
rocscores.append(roc_auc_score(Ydata[test], pred))
else:
rocscores.append(numpy.nan)
importances.append(clf.feature_importances_)
return rocscores, importances
bx_df['beh_nback_neut-gt-neg_rt'] = bx_df['beh_nback_neutface_cor_rt'] - bx_df[
'beh_nback_negface_cor_rt']
bx_df['beh_nback_neut-gt-pos_rt'] = bx_df['beh_nback_neutface_cor_rt'] - bx_df[
'bn_posface_cor_rt']
bx_df['sst_cor_stop_pct'] = 1 - bx_df['bs_incor_stop_percent_total']
bx_vars = [
'cash_choice_task', 'sst_cor_stop_pct', 'upps_y_ss_negative_urgency',
'upps_y_ss_lack_of_planning', 'upps_y_ss_sensation_seeking',
'upps_y_ss_positive_urgency', 'upps_y_lack_perseverance',
'bis_y_ss_bis_sum', 'bis_y_ss_bas_rr', 'bis_y_ss_bas_drive',
'bis_y_ss_bas_fs', 'nihtbx_flanker_agecorrected',
'beh_nback_neut-gt-neg_rt', 'beh_nback_neut-gt-pos_rt'
]
impute_pls = fi.SoftImpute(verbose=False)
complete_bx = impute_pls.fit_transform(bx_df[bx_vars])
complete_scaled = scale(complete_bx)
decomp = PCA().fit(complete_scaled)
fig, ax = plt.subplots(figsize=[7, 5])
plt.tight_layout(pad=2)
ax2 = ax.twinx()
g = sns.pointplot(np.arange(1, 15),
decomp.explained_variance_ratio_,
markers='x',
join=True,
ax=ax)
h = sns.lineplot(x=np.arange(1, 15),
def impute(df, method, verbose=False):
"""
Impute missing data using specified imputation method.
Parameters
----------
df: pd.DataFrame
Stat DataFrame with source columns and player/team multi-index.
method: str/bool
Imputation method for missing data.
- False: Do not impute missing data.
- None: Do not impute missing data.
- 'BiScaler'
- 'IterativeImpute'
- 'IterativeSVD'
- 'KNN': Impute with nearest neighbors.
- 'Mean': Impute missing with average of other sources.
- 'NuclearNorm'
- 'SoftImpute'
verbose: bool, default=False
If True, print debugging information.
Returns
-------
df: pd.DataFrame
Imputed DataFrame with no NaNs.
"""
warnings.filterwarnings('ignore', category=RuntimeWarning)
# Subset DataFrame to only include only projection columns.
ignored_cols = ['Player', 'Team', 'Pos', 'Week', 'STATS']
impute_cols = [col for col in list(df) if col not in ignored_cols]
X = df[impute_cols].copy().T
# Impute DataFrame.
v = verbose
if method in [None, False]:
imputed_vals = X.values
elif np.sum(np.sum(X.isnull())) == 0:
# No missing values.
imputed_vals = X.values
elif method == 'BiScaler':
imputed_vals = fi.BiScaler(verbose=v).fit_transform(X)
elif method == 'IterativeImpute':
imputed_vals = fi.IterativeImputer(verbose=v).fit_transform(X)
elif method == 'IterativeSVD':
imputed_vals = fi.IterativeSVD(verbose=v).fit_transform(X)
elif method == 'KNN':
imputed_vals = fi.KNN(k=3, verbose=v).fit_transform(X)
elif method == 'MatrixFactorization':
imputed_vals = fi.MatrixFactorization(verbose=v).fit_transform(X)
elif method == 'Mean':
imputed_vals = fi.SimpleFill('mean').fit_transform(X)
elif method == 'Median':
imputed_vals = fi.SimpleFill('median').fit_transform(X)
elif method == 'NuclearNorm':
imputed_vals = fi.NuclearNormMinimization(verbose=v).fit_transform(X)
elif method == 'SoftImpute':
imputed_vals = fi.SoftImpute(verbose=v).fit_transform(X)
# Recombine ignored columns with imputed data.
imputed_df = pd.DataFrame(imputed_vals.T, columns=X.index)
for col in impute_cols:
if len(imputed_df[col]) != len(df[col]):
print(f'df: {len(df[col])}\nimp: {len(imputed_df[col])}')
df[col] = imputed_df[col].values
return df
def impute(self): return fi.SoftImpute(verbose=False).complete(self.missing_data)
#**********************************
# Load Data
#**********************************
datasets = pickle.load(open('../Data/subset_data.pkl', 'rb'))
data = datasets['all_data']
task_data = datasets['task_data']
survey_data = datasets['survey_data']
verbose_lookup = datasets['verbose_lookup']
# ************************************
# ************ Imputation *******************
# ************************************
data.drop(['ptid', 'gender', 'age'], axis=1, inplace=True)
data_complete = fancyimpute.SoftImpute().complete(data)
data_complete = pd.DataFrame(data_complete,
index=data.index,
columns=data.columns)
# ************************************
# ************ Connectivity Matrix *******************
# ************************************
spearman_connectivity = calc_connectivity_mat(data_complete,
edge_metric='spearman')
distance_connectivity = calc_connectivity_mat(data_complete,
edge_metric='distance')
# ************************************
# ********* Graphs *******************
import numpy as np
import fancyimpute
print("Loading memory")
mem = np.load(
'/home/aocc/code/DL/MDP_learning/save_memory/first_20mill/BipedalWalker-v2CORRUPT0.1.npy'
)
print("Imputing memory")
memory_final = fancyimpute.SoftImpute().complete(mem)
print("Saving imputed memory")
np.save(
'/home/aocc/code/DL/MDP_learning/save_memory/first_20mill/BipedalWalker-v2IMPUTED.1',
memory_final)
import pandas as pd
import seaborn as sns
from sklearn import decomposition
import sys
sys.path.append('../utils')
from utils import get_behav_data
from plot_utils import dendroheatmap
# get dependent variables
DV_df = get_behav_data('Discovery_9-26-16', use_EZ=True)
# ************************************
# ************ Imputation *******************
# ************************************
DV_df_complete = fancyimpute.SoftImpute().complete(DV_df)
DV_df_complete = pd.DataFrame(DV_df_complete,
index=DV_df.index,
columns=DV_df.columns)
# ************************************
# ************ PCA *******************
# ************************************
pca_data = DV_df_complete.corr()
pca = decomposition.PCA()
pca.fit(pca_data)
# plot explained variance vs. components
sns.plt.plot(pca.explained_variance_ratio_.cumsum())
def softimpute(X, y=None):
return (fancyimpute.SoftImpute(verbose=False).complete(X))
def preprocess(trainfile,
testfile,
outputdir,
useless_attr,
miss_threshold,
xstrategy,
ymin,
ymax,
ystrategy,
fill_method="MICE",
normal01=True):
"""对XY进行数据预处理，矩阵补全、正则化标准化等。
:param trainfile: string, 训练集(d_train_20180102.csv)的路径
:param testfile: string, 测试集(d_test_A_20180102.csv)的路径
:param outputdir: string, 预处理后文件保存的路径
:param useless_attr: list, 需要删除的无用属性，比如[0, 1, 2, 3]
:param miss_threshold: float, 属性确实严重忽略的阈值，百分比，比如0.7
:param xstrategy: string, 对x中奇异点的处理方式{"replace", "nothing"}
:param ymin: float, 对Y中点的最小值，小于这个值，即为奇异点
:param ymax: float, 对Y中点的最大值，超过这个值，就是奇异点
:param ystrategy: string, 对y中奇异点的处理方式("delete", "replace", "nothing")
:param fill_method: string, 矩阵补全的策略，{"KNN", "SoftI", "MF", "MICE"}
:param normal01: bool, 如果为真，则对结果进行归一化到01，否则，不归一化
:return: list, 归一化之后的trainX, trainY, testX
"""
# 0. 读入训练集，测试集
train_XY = convert(trainfile)
test_X = convert(testfile)
print("读入数据集，开始数据预处理")
# 1. 删除无用属性列
train_id = train_XY[:, 0:1]
test_id = test_X[:, 0:1]
train_XY = np.delete(train_XY, useless_attr, axis=1)
test_X = np.delete(test_X, useless_attr, axis=1)
n_test = test_X.shape[0]
info1 = "1. 删除train_XY, test_X上的无用属性:%s, train_X.shape=%s, test_X.shape=%s"\
%(str(useless_attr), str(train_XY.shape), str(test_X.shape))
print(info1)
# 2. 删除缺失严重的列
miss_mask = np.isnan(train_XY)
n = miss_mask.shape[0]
column_del = [] # 删除列的list
for i in range(miss_mask.shape[1]):
miss_n = miss_mask[:, i].sum()
if miss_n / n >= miss_threshold:
column_del.append(i)
train_XY = np.delete(train_XY, column_del, axis=1)
test_X = np.delete(test_X, column_del, axis=1)
info2 = "2. 在train_XY, test_X上删除缺失超过%f%%的属性:%s" % (miss_threshold * 100,
str(column_del))
print(info2)
# 3. 对y进行去噪，手动设置阈值
train_Y = train_XY[:, -1:]
upper_mask = train_Y > ymax
lower_mask = train_Y < ymin
if ystrategy == "replace":
train_Y[upper_mask] = ymax
train_Y[lower_mask] = ymin
elif ystrategy == "delete":
index = np.array(np.arange(0, train_Y.shape[0], 1), ndmin=2).T
chsn_mask = upper_mask | lower_mask
train_XY = np.delete(train_XY, index[chsn_mask], axis=0)
train_id = np.delete(train_id, index[chsn_mask], axis=0)
elif ystrategy == "nothing":
pass
else:
raise ValueError(r"'ystrategy'应该是{nothing, replace, delete}中的一个")
train_Y = train_XY[:, -1:]
print("3. 对trainY去噪(%s)，trainXY.shape=%s" % (ystrategy, train_XY.shape))
# 4. 对X进行操作，通过boxplot计算阈值
train_X = train_XY[:, :-1]
all_X = np.concatenate([train_X, test_X], axis=0)
attr_n = train_XY.shape[1] - 1
attr_min_max = np.zeros(
(attr_n, 2), dtype=np.float64) # 存储每个属性经过boxplot之后的最小最大值，即阈值array
if xstrategy == "nothing":
pass
elif xstrategy == "replace":
# 对X中的奇异点 替换为 最值
for i in range(attr_n):
# 对每列进行浅拷贝，对crt_attr操作相当于对train_XY操作
crt_attr = all_X[:, i:i + 1]
miss = np.isnan(crt_attr)
box_dic = plt.boxplot(crt_attr[~miss])
crt_max = box_dic["caps"][0].get_ydata()[0]
crt_min = box_dic["caps"][1].get_ydata()[0]
if crt_max < crt_min:
tmp = crt_max
crt_max = crt_min
crt_min = tmp
attr_min_max[i, 0] = crt_min
attr_min_max[i, 1] = crt_max
crt_attr[miss] = 0
upper_mask = crt_attr > crt_max
lower_mask = crt_attr < crt_min
upper_mask &= ~miss
lower_mask &= ~miss
crt_attr[upper_mask] = crt_max
crt_attr[lower_mask] = crt_min
crt_attr[miss] = np.nan
else:
raise ValueError(r"'xstrategy'应该是{nothing, replace}中的一个")
print(r"4. 对所有的X进行去噪(%s)." % xstrategy)
# 5. 矩阵补全
completer = None
if fill_method == "KNN":
completer = fi.KNN(verbose=False)
elif fill_method == "SoftI":
completer = fi.SoftImpute(verbose=False)
elif fill_method == "MF":
completer = fi.MatrixFactorization(verbose=False)
elif fill_method == "MICE":
completer = fi.MICE(verbose=False)
else:
ValueError(r"'fill_method'应该是{'KNN','SoftI','MF','MICE'}.")
all_X_complete = completer.complete(all_X)
print("5. 在all_X上进行矩阵补全(%s)." % fill_method)
# train_X = all_X_complete[:-1000, :]
# test_X = all_X_complete[-1000:, :]
# 6. 归一化，以及01缩放
if normal01:
X_nmler = StandardScaler()
X_01 = MinMaxScaler()
Y_nmler = StandardScaler()
Y_01 = MinMaxScaler()
X_nmler.fit(all_X_complete)
Y_nmler.fit(train_Y)
all_X_nml = X_nmler.transform(all_X_complete)
train_Y_nml = Y_nmler.transform(train_Y)
X_01.fit(all_X_nml)
Y_01.fit(train_Y_nml)
all_X_nml01 = X_01.transform(all_X_nml)
train_Y_nml01 = Y_01.transform(train_Y_nml)
final_train_X = all_X_nml01[:-n_test, :]
final_test_X = all_X_nml01[-n_test:, :]
final_train_Y = np.concatenate([train_Y_nml01, train_Y], axis=1)
else:
final_train_X = all_X_complete[:-n_test, :]
final_test_X = all_X_complete[-n_test:, :]
final_train_Y = train_Y
print(r"6. 对all_X, train_Y归一化到01(%s)." % normal01)
# 7. 存储数据
print(r"7. 存储数据为: 集合_类别_日期.csv(%s)." % outputdir)
# timestamp = datetime.now().strftime("%Y%m%d%H%M")
timestamp = "0000"
np.savetxt(outputdir + r"\train_X_" + timestamp + ".csv",
final_train_X,
delimiter=",")
np.savetxt(outputdir + r"\test_X_" + timestamp + ".csv",
final_test_X,
delimiter=",")
np.savetxt(outputdir + r"\train_Y_" + timestamp + ".csv",
final_train_Y,
delimiter=",")
np.savetxt(outputdir + r"\train_id_" + timestamp + ".csv",
train_id.astype(np.int64),
delimiter=",")
np.savetxt(outputdir + r"\test_id_" + timestamp + ".csv",
test_id.astype(np.int64),
delimiter=",")
return train_X, train_Y, test_X, train_id
if not os.path.exists(os.path.dirname(datafile)):
os.mkdir(os.path.dirname(datafile))
behavdata = get_behav_data(dataset)
filter_by_icc = False
if filter_by_icc:
icc_threshold = 0.25
icc_boot = pandas.read_csv(
'../Data/Retest_09-27-2017/bootstrap_merged.csv')
icc = icc_boot.groupby('dv').mean().icc
for v in behavdata.columns:
if icc.loc[v] < icc_threshold:
del behavdata[v]
behavdata_imputed = fancyimpute.SoftImpute().fit_transform(
behavdata.values)
df = pandas.DataFrame(behavdata_imputed, columns=behavdata.columns)
for dropvar in [
'kirby_mturk.percent_patient', 'kirby_mturk.exp_discount_rate',
'kirby_mturk.hyp_discount_rate',
'kirby_mturk.exp_discount_rate_medium',
'kirby_mturk.exp_discount_rate_small',
'kirby_mturk.exp_discount_rate_large',
'probabilistic_selection_mturk.positive_learning_bias',
'shift_task_mturk.nonperseverative_errors',
'stop_signal_mturk.omission_errors', 'stop_signal_mturk.SSRT',
'stroop_mturk.incongruent_errors'
]:
del df[dropvar]
