In [1]:
# import modules as usual
import os
import glob
import numpy as np
import pandas as pd
from pandas import DataFrame, Series
import matplotlib.pyplot as plt
%matplotlib inline
import seaborn as sns
import cv2
from sklearn.model_selection import StratifiedKFold
from sklearn.metrics import roc_auc_score
from xgboost import XGBClassifier
In [2]:
# path of files
path_positive_1 = '../eskin_data/yamamoto/throw/'
path_negative_1 = '../eskin_data/yamamoto/others/'
path_positive_2 = '../eskin_data/zennra/throw/'
path_negative_2 = '../eskin_data/zennra/others/'
In [3]:
# extract the moment of throwing based on accel values
def extract_action(df):
df = df.reset_index()
mom_action = int((np.argmax(abs(df.accelX))+ np.argmax(abs(df.accelY))+ np.argmax(abs(df.accelZ)))/3)
df = df.ix[mom_action-90:mom_action+90]
df.index = df.time
df.drop(["time"], axis=1, inplace=True)
return df.as_matrix()
In [4]:
def load_positive_data(path):
path = os.path.join(path, '*.csv')
files = glob.glob(path)
X_positives = []
for file_path in files:
df = pd.read_csv(file_path, index_col=0)
df = extract_action(df)
X_positives.append(df)
X_positives = np.array(X_positives)
y_positives = np.ones(len(X_positives))
return X_positives, y_positives
In [5]:
def load_negative_data(path, num_clip=100, random_state=71):
np.random.seed(random_state)
path = os.path.join(path, '*.csv')
files = glob.glob(path)
X_negatives = []
for file_path in files:
df = pd.read_csv(file_path, index_col=0)
for i in range(num_clip):
start = np.random.choice(range(len(df)-180))
df_extracted = df.iloc[start:start+180].as_matrix()
X_negatives.append(df_extracted)
X_negatives = np.array(X_negatives)
y_negatives = np.zeros(len(X_negatives))
return X_negatives, y_negatives
In [6]:
def resize_matrix(X, size = (20, 20), flatten=False):
X_resized = []
for i in range(len(X)):
X_ = X[i] /1.
X_ = cv2.resize(X_, size, interpolation = cv2.INTER_LINEAR)
if flatten == True: # True for XGBoost etc., False for CNN (Convolutional Newral Networks)
X_ = X_.ravel()
X_resized.append(X_)
X_resized = np.array(X_resized)
return X_resized
In [7]:
def add_baseline(X, y, num_augment=1, max_abs_tilt=0.1, max_abs_intercept=20, random_state=71):
np.random.seed(random_state)
X_augmented = []
y_augmented = []
for j in range(num_augment):
for i in range(X.shape[0]):
X_ = X[i].copy()
y_ = y[i].copy()
for col in range(X_.shape[1]-6):
tilt = np.random.uniform(-max_abs_tilt, max_abs_tilt)
intercept = np.random.uniform(-max_abs_intercept, max_abs_intercept)
baseline = np.array(range(X_.shape[0])) * tilt + intercept
X_[:,col] = X_[:,col] + baseline
X_augmented.append(X_)
y_augmented.append(y_)
X_augmented = np.array(X_augmented)
y_augmented = np.array(y_augmented)
X = np.concatenate((X, X_augmented), axis=0)
y = np.concatenate((y, y_augmented), axis=0)
return X, y
In [8]:
def prepare_positive_data(path, size = (20, 20), flatten=True, is_augment=True, num_augment=10, max_abs_tilt=5, max_abs_intercept=20, random_state=71):
X_positives, y_positives = load_positive_data(path)
if is_augment == True:
X_positives, y_positives = add_baseline(X_positives, y_positives, num_augment=num_augment, max_abs_tilt=max_abs_tilt, max_abs_intercept=max_abs_intercept, random_state=random_state)
X_positives = resize_matrix(X_positives, flatten=flatten)
return X_positives, y_positives
def prepare_negative_data(path, num_clip=500, size = (20, 20), flatten=True, is_augment=True, num_augment=10, max_abs_tilt=5, max_abs_intercept=20, random_state=71):
X_negatives, y_negatives = load_negative_data(path, num_clip=num_clip, random_state=random_state)
if is_augment == True:
X_negatives, y_negatives = add_baseline(X_negatives, y_negatives, num_augment=num_augment, max_abs_tilt=max_abs_tilt, max_abs_intercept=max_abs_intercept, random_state=random_state)
X_negatives = resize_matrix(X_negatives, flatten=flatten)
return X_negatives, y_negatives
In [9]:
def scan_evaluate(clf, path, size = (20, 20), flatten=True):
df = pd.read_csv(path, index_col=0)
X_windows = []
for i in range(df.shape[0]-180):
X_window = df.iloc[i:i+180].as_matrix()
X_windows.append(X_window)
X_windows = np.array(X_windows)
X_windows = resize_matrix(X_windows, size = size, flatten=flatten)
probabilities = clf.predict_proba(X_windows)[:,1]
plt.plot([x/60 for x in range(df.shape[0]-180)], probabilities)
plt.xlabel("sec")
plt.ylabel("probability")
plt.ylim(0,1)
plt.show()
In [10]:
# 1人目（山本）のモーションデータをData Augmentationありで読み込む
X_positives_1, y_positives_1 = prepare_positive_data(path_positive_1, is_augment=True, max_abs_tilt=5.0, num_augment=10)
X_negatives_1, y_negatives_1 = prepare_negative_data(path_negative_1, is_augment=False)
X_1_aug = np.concatenate((X_positives_1, X_negatives_1), axis=0)
y_1_aug = np.concatenate((y_positives_1, y_negatives_1), axis=0)
# 1人目（山本）のモーションデータをData Augmentationなしで読み込む
X_positives_1, y_positives_1 = prepare_positive_data(path_positive_1, is_augment=False)
X_negatives_1, y_negatives_1 = prepare_negative_data(path_negative_1, is_augment=False)
X_1 = np.concatenate((X_positives_1, X_negatives_1), axis=0)
y_1 = np.concatenate((y_positives_1, y_negatives_1), axis=0)
# 2人目（武井氏）のモーションデータをData Augmentationありで読み込む
X_positives_2, y_positives_2 = prepare_positive_data(path_positive_2, is_augment=True, max_abs_tilt=5.0, num_augment=10)
X_negatives_2, y_negatives_2 = prepare_negative_data(path_negative_2, is_augment=False)
X_2_aug = np.concatenate((X_positives_2, X_negatives_2), axis=0)
y_2_aug = np.concatenate((y_positives_2, y_negatives_2), axis=0)
# 2人目（武井氏）のモーションデータをData Augmentationなしで読み込む
X_positives_2, y_positives_2 = prepare_positive_data(path_positive_2, is_augment=False)
X_negatives_2, y_negatives_2 = prepare_negative_data(path_negative_2, is_augment=False)
X_2 = np.concatenate((X_positives_2, X_negatives_2), axis=0)
y_2 = np.concatenate((y_positives_2, y_negatives_2), axis=0)
In [11]:
# それぞれのサイズを確認する
print(X_1_aug.shape, y_1_aug.shape)
print(X_1.shape, y_1.shape)
print(X_2_aug.shape, y_2_aug.shape)
print(X_2.shape, y_2.shape)
In [12]:
# 山本データで学習して、武井データに対して予測を行う。やや低下が見られたが依然として良好。
# もう一方のfoldで大きなゲインが得られたことを考えるとトータルでは分類性能向上しているか。
clf_xgb = XGBClassifier(base_score=0.5, colsample_bylevel=1, colsample_bytree=0.7,
gamma=0, learning_rate=0.1, max_delta_step=0, max_depth=3,
min_child_weight=1, missing=None, n_estimators=100, nthread=-1,
objective='binary:logistic', reg_alpha=0, reg_lambda=1,
scale_pos_weight=1, seed=0, silent=True, subsample=0.7)
clf_xgb.fit(X_1, y_1)
probs = clf_xgb.predict_proba(X_2)[:,1]
score = roc_auc_score(y_2, probs)
print("Without Data Augmentation of train data, ROC-AUC: ", score)
# 時系列データをスキャンしてprobabilityの時間変化を可視化する
scan_evaluate(clf_xgb, "../eskin_data/zennra_throw.csv", flatten=True)
scan_evaluate(clf_xgb, "../eskin_data/zennra/others/zennra_radiotaiso.csv", flatten=True)
scan_evaluate(clf_xgb, "../eskin_data/zennra/others/zennra_ebikanix.csv", flatten=True)
clf_xgb.fit(X_1_aug, y_1_aug)
probs = clf_xgb.predict_proba(X_2)[:,1]
score = roc_auc_score(y_2, probs)
print("With Data Augmentation of train data, ROC-AUC: ", score)
In [13]:
# 武井データで学習して、山本データに対して予測を行う。顕著な改善効果が認められた。
clf_xgb = XGBClassifier(base_score=0.5, colsample_bylevel=1, colsample_bytree=0.7,
gamma=0, learning_rate=0.1, max_delta_step=0, max_depth=3,
min_child_weight=1, missing=None, n_estimators=100, nthread=-1,
objective='binary:logistic', reg_alpha=0, reg_lambda=1,
scale_pos_weight=1, seed=0, silent=True, subsample=0.7)
clf_xgb.fit(X_2, y_2)
probs = clf_xgb.predict_proba(X_1)[:,1]
score = roc_auc_score(y_1, probs)
print("Without Data Augmentation of train data, ROC-AUC: ", score)
clf_xgb.fit(X_2_aug, y_2_aug)
probs = clf_xgb.predict_proba(X_1)[:,1]
score = roc_auc_score(y_1, probs)
print("With Data Augmentation of train data, ROC-AUC: ", score)