def _train(self, df_train, df_test, score):
df_train = df_train.sort_values(["sym", "date"])
df_train_1 = df_train[df_train[score.get_name()] < 0.5]
#df_train_1 = df_train_1.sort_values([self.confer.scores[1].get_name()], ascending=True).head(int(len(df_train_1)/3*2)).tail(int(len(df_train_1)/3))
#print(df_train_1.head()[["sym", "date", "close", self.confer.scores[1].get_name()]])
df_train_2 = df_train[df_train[score.get_name()] > 0.5]
#df_train_2 = df_train_2.sort_values([self.confer.scores[1].get_name()], ascending=False).head(int(len(df_train_2)/3*2)).tail(int(len(df_train_2)/3))
#print(df_train_2.head()[["sym", "date", "close", self.confer.scores[1].get_name()]])
# @ccl
df_train_2 = df_train_2.sample(n = len(df_train_1))
assert(len(df_train_2) == len(df_train_1))
df_train = pd.concat([df_train_1, df_train_2], axis=0)
df_train = df_train.sample(frac=1.0)
assert(len(df_train) == 2*len(df_train_1))
print("train start : %s train end: %s total:%d" % (df_train.sort_values('date').head(1)['date'].values[0],
df_train.sort_values('date').tail(1)['date'].values[0], len(df_train)))
npTrainFeat, npTrainLabel = base.extract_feat_label(df_train, score.get_name())
df_test = df_test.sort_values(["sym", "date"])
df_test_1 = df_test[df_test[score.get_name()] < 0.5]
df_test_2 = df_test[df_test[score.get_name()] > 0.5]
# assert len(df_test_1) + len(df_test_2) == len(df_test)
df_test_2 = df_test_2.sample(n = len(df_test_1))
assert(len(df_test_2) == len(df_test_1))
df_test = pd.concat([df_test_1, df_test_2], axis=0)
assert(len(df_test) == 2*len(df_test_1))
df_test = df_test.sample(frac=1.0, random_state = 1253)
npTestFeat, npTestLabel = base.extract_feat_label(df_test, score.get_name())
#self.confer.classifier.fit(npTrainFeat, npTrainLabel, npTestFeat, npTestLabel, npTestFeat, npTestLabel)
#self.confer.classifier.fit(npTrainFeat, npTrainLabel, npTestFeat, npTestLabel)
self.confer.classifier.fit(npTrainFeat, npTrainLabel, df_test, score.get_name())
def on_epoch_end(self, epoch, logs={}):
if epoch % (self.interval*5) != 0:
return
print("")
print("LONG...")
(thresholds, scores) = self.cal_accuracy(self.npFeatTest, self.npLabelTest)
print("TEST: ", end='')
for i in range(len(thresholds)):
print("score: %.3f(%.3f)" % (scores[i], thresholds[i]), end=" ")
print()
print("VALD: ", end='')
scores = self.cal_accuracy2(self.npFeatVal, self.npLabelVal, thresholds)
for i in range(len(thresholds)):
print("score: %.3f(%.3f)" % (scores[i], thresholds[i]), end=" ")
print()
print()
is_first = True
for df_year in self.df_years:
npFeat, npLabel = base.extract_feat_label(df_year, self.score,drop=True)
year = df_year['yyyy'].unique()[0]
print("%s: " % year, end='')
if is_first:
(thresholds, scores) = self.cal_accuracy(npFeat, npLabel)
for i in range(len(thresholds)):
print("score: %.3f(%.3f)" % (scores[i], thresholds[i]), end=" ")
is_first = False
else:
scores = self.cal_accuracy2(npFeat, npLabel, thresholds)
for i in range(len(thresholds)):
print("score: %.3f(%.3f)" % (scores[i], thresholds[i]), end=" ")
print()
print("SHORT...")
(thresholds, scores) = self.cal_accuracy(self.npFeatTest, self.npLabelTest, is_short=True)
print("TEST: ", end='')
for i in range(len(thresholds)):
print("score: %.3f(%.3f)" % (scores[i], thresholds[i]), end=" ")
print()
print("VALD: ", end='')
scores = self.cal_accuracy2(self.npFeatVal, self.npLabelVal, thresholds, is_short=True)
for i in range(len(thresholds)):
print("score: %.3f(%.3f)" % (scores[i], thresholds[i]), end=" ")
print()
print()
is_first = True
for df_year in self.df_years:
npFeat, npLabel = base.extract_feat_label(df_year, self.score,drop=True)
year = df_year['yyyy'].unique()[0]
print("%s: " % year, end='')
if is_first:
(thresholds, scores) = self.cal_accuracy(npFeat, npLabel, is_short=True)
for i in range(len(thresholds)):
print("score: %.3f(%.3f)" % (scores[i], thresholds[i]), end=" ")
is_first = False
else:
scores = self.cal_accuracy2(npFeat, npLabel, thresholds, is_short=True)
for i in range(len(thresholds)):
print("score: %.3f(%.3f)" % (scores[i], thresholds[i]), end=" ")
print()
def __init__(self, cls, validation_data=(), interval=10):
super(Callback, self).__init__()
self.interval = interval
self.cls = cls
self.df_test_valid, self.score = validation_data
#self.df_test = self.df_test_valid.sample(frac=0.5, random_state=200)
self.df_test_valid = self.df_test_valid.sort_values("date", ascending=True)
self.df_test = self.df_test_valid.head(int(len(self.df_test_valid)/2))
self.df_valid = self.df_test_valid.drop(self.df_test.index)
assert(len(self.df_valid) + len(self.df_test) == len(self.df_test_valid))
self.df_years = [each for each in base.split_by_year(self.df_test_valid)]
self.npFeatTest, self.npLabelTest = base.extract_feat_label(self.df_test, self.score, drop=True)
self.npFeatVal, self.npLabelVal = base.extract_feat_label(self.df_valid, self.score, drop=True)
def __init__(self, cls, validation_data=(), interval=10):
super(Callback, self).__init__()
self.interval = interval
self.cls = cls
self.df_test_valid, self.score = validation_data
#self.df_test = self.df_test_valid.sample(frac=0.5, random_state=200)
self.df_test_valid = self.df_test_valid.sort_values("date", ascending=True)
self.df_test = self.df_test_valid.head(int(len(self.df_test_valid)/2))
self.df_valid = self.df_test_valid.drop(self.df_test.index)
assert(len(self.df_valid) + len(self.df_test) == len(self.df_test_valid))
self.df_years = [each for each in base.split_by_year(self.df_test_valid)]
self.npFeatTest, self.npLabelTest = base.extract_feat_label(self.df_test, self.score, drop=True)
self.npFeatVal, self.npLabelVal = base.extract_feat_label(self.df_valid, self.score, drop=True)
def on_epoch_end(self, epoch, logs={}):
return
if epoch % (self.interval*5) != 0:
return
print("")
print("LONG...")
(thresholds, scores) = self.cal_accuracy(self.npFeatTest, self.npLabelTest)
print("TEST: ", end='')
for i in range(len(thresholds)):
print("score: %.3f(%.3f)" % (scores[i], thresholds[i]), end=" ")
print()
print("VALD: ", end='')
scores = self.cal_accuracy2(self.npFeatVal, self.npLabelVal, thresholds)
for i in range(len(thresholds)):
print("score: %.3f(%.3f)" % (scores[i], thresholds[i]), end=" ")
print()
print()
is_first = True
for df_year in self.df_years:
npFeat, npLabel = base.extract_feat_label(df_year, self.score,drop=True)
year = df_year['yyyy'].unique()[0]
print("%s: " % year, end='')
if is_first:
(thresholds, scores) = self.cal_accuracy(npFeat, npLabel)
for i in range(len(thresholds)):
print("score: %.3f(%.3f)" % (scores[i], thresholds[i]), end=" ")
is_first = False
else:
scores = self.cal_accuracy2(npFeat, npLabel, thresholds)
for i in range(len(thresholds)):
print("score: %.3f(%.3f)" % (scores[i], thresholds[i]), end=" ")
print()
print("SHORT...")
(thresholds, scores) = self.cal_accuracy(self.npFeatTest, self.npLabelTest, is_short=True)
print("TEST: ", end='')
for i in range(len(thresholds)):
print("score: %.3f(%.3f)" % (scores[i], thresholds[i]), end=" ")
print()
print("VALD: ", end='')
scores = self.cal_accuracy2(self.npFeatVal, self.npLabelVal, thresholds, is_short=True)
for i in range(len(thresholds)):
print("score: %.3f(%.3f)" % (scores[i], thresholds[i]), end=" ")
print()
print()
is_first = True
for df_year in self.df_years:
npFeat, npLabel = base.extract_feat_label(df_year, self.score,drop=True)
year = df_year['yyyy'].unique()[0]
print("%s: " % year, end='')
if is_first:
(thresholds, scores) = self.cal_accuracy(npFeat, npLabel, is_short=True)
for i in range(len(thresholds)):
print("score: %.3f(%.3f)" % (scores[i], thresholds[i]), end=" ")
is_first = False
else:
scores = self.cal_accuracy2(npFeat, npLabel, thresholds, is_short=True)
for i in range(len(thresholds)):
print("score: %.3f(%.3f)" % (scores[i], thresholds[i]), end=" ")
print()