def GetTrainData():
GetTrainDataTempSave()
features = np.load('./data/dataset/fix_dataset_temp_features.npy')
labels = np.load('./data/dataset/fix_dataset_temp_labels.npy')
train_data = np.append(features, labels, axis=1)
print("train_data: {}".format(train_data.shape))
# raw_input("Enter ...")
print("reorder...")
order = np.argsort(np.random.random(len(train_data)))
train_data = train_data[order]
train_data = train_data[:2000000]
# raw_input("Enter ...")
# sample_train_data = train_data[:10]
print("get feature ...")
train_features = train_data[:, 0:feature.FEATURE_SIZE()].copy()
# raw_input("Enter ...")
print("get label...")
train_labels = train_data[:,
feature.COL_ACTIVE_LABEL(
):feature.COL_ACTIVE_LABEL() + 1].copy()
# raw_input("Enter ...")
print("train_features: {}".format(train_features.shape))
print("train_labels: {}".format(train_labels.shape))
return train_features, train_labels
def Train(self):
date_num = self.train_dataset.shape[0]
code_num = self.train_dataset.shape[1]
print("date_num: %u" % date_num)
print("code_num: %u" % code_num)
print('\n%-8s%-8s%-8s%-12s%-12s%-12s' %
('T', 'step', 'loss', 'increase', 'trade_count', 'max_Q_mean'))
print('-' * 60)
for iloop in range(0, T):
curren_status = STATUS_OFF
for dloop in reversed(range(0, date_num)): # 遍历train_dataset的日期
if curren_status == STATUS_OFF:
code_index = random.randint(0, code_num - 1)
if self.train_dataset[dloop][code_index][
date_col_index] == 0.0:
continue
current_feature = self.train_dataset[dloop][code_index][
0:feature.FEATURE_SIZE()]
Q = self.GetTrainActionQ(current_feature)
if Q > 0:
curren_status = STATUS_ON
else:
curren_status = STATUS_OFF
t1_date_index = self.NextValidDateIndex(
self.train_dataset, code_index, dloop)
if t1_date_index < 0:
break
t2_date_index = self.NextValidDateIndex(
self.train_dataset, code_index, t1_date_index)
if t2_date_index < 0:
break
next_feature = self.train_dataset[t1_date_index][code_index][
0:feature.FEATURE_SIZE()]
reward = (self.train_dataset[t2_date_index][code_index]
[open_col_index] / self.train_dataset[t1_date_index]
[code_index][open_col_index] - 1.0) * 100.0
self.PerceiveAndTrain(current_feature, next_feature, reward)
increase, trade_count, max_Q_mean = self.TestTop1()
print('%-8u%-8u%-8.5f%-12.3f%-12u%-12.6f' %
(iloop, self.step_num, self.LossMean(), increase,
trade_count, max_Q_mean))
self.LossClean()
self.SaveModel_(iloop)
def GetTrainTestDataSplitByDate():
dataset = np.load(FileNameFixDataSet())
print("dataset: {}".format(dataset.shape))
pos = dataset[:, feature.COL_TRADE_DATE(0)] < dataset_train_test_split_date
train_data = dataset[pos]
test_data = dataset[~pos]
print("train: {}".format(train_data.shape))
print("test: {}".format(test_data.shape))
train_features = train_data[:, 0:feature.FEATURE_SIZE()]
train_labels = train_data[:,
feature.COL_ACTIVE_LABEL(
):feature.COL_ACTIVE_LABEL() + 1]
test_features = test_data[:, 0:feature.FEATURE_SIZE()]
test_labels = test_data[:,
feature.COL_ACTIVE_LABEL(
):feature.COL_ACTIVE_LABEL() + 1]
return train_features, train_labels, test_features, test_labels, test_data
def GetTrainTestDataSampleByDate(test_ratio):
sample_num = int(1.0 / test_ratio + 0.0001)
dataset = np.load(FileNameFixDataSet())
print("dataset: {}".format(dataset.shape))
pos = ((dataset[:, feature.COL_TRADE_DATE(0)].astype(int) % 100) %
sample_num) == 0
test_data = dataset[pos]
train_data = dataset[~pos]
print("train: {}".format(train_data.shape))
print("test: {}".format(test_data.shape))
train_features = train_data[:, 0:feature.FEATURE_SIZE()]
train_labels = train_data[:,
feature.COL_ACTIVE_LABEL(
):feature.COL_ACTIVE_LABEL() + 1]
test_features = test_data[:, 0:feature.FEATURE_SIZE()]
test_labels = test_data[:,
feature.COL_ACTIVE_LABEL(
):feature.COL_ACTIVE_LABEL() + 1]
return train_features, train_labels, test_features, test_labels, test_data
def GetTrainTestDataRandom(test_ratio):
sample_num = int(1.0 / test_ratio + 0.0001)
dataset = np.load(FileNameFixDataSet())
print("dataset: {}".format(dataset.shape))
print('sample_num:%u' % sample_num)
# 生成数值范围在 0-（sample_num-1）的随机数组，pos是值为0的位置
pos = (np.random.randint(0, sample_num, size=len(dataset)) == 0)
test_data = dataset[pos]
train_data = dataset[~pos]
print("train: {}".format(train_data.shape))
print("test: {}".format(test_data.shape))
train_features = train_data[:, 0:feature.FEATURE_SIZE()]
train_labels = train_data[:,
feature.COL_ACTIVE_LABEL(
):feature.COL_ACTIVE_LABEL() + 1]
test_features = test_data[:, 0:feature.FEATURE_SIZE()]
test_labels = test_data[:,
feature.COL_ACTIVE_LABEL(
):feature.COL_ACTIVE_LABEL() + 1]
return train_features, train_labels, test_features, test_labels, test_data
def GetTrainDataTempSave():
dataset = np.load(FileNameFixDataSet())
print("dataset: {}".format(dataset.shape))
pos = dataset[:, feature.COL_TRADE_DATE(0)] < dataset_train_test_split_date
train_data = dataset[pos]
print("train_data: {}".format(train_data.shape))
features = train_data[:, 0:feature.FEATURE_SIZE()]
labels = train_data[:,
feature.COL_ACTIVE_LABEL():feature.COL_ACTIVE_LABEL() +
1]
np.save('./data/dataset/fix_dataset_temp_features.npy', features)
np.save('./data/dataset/fix_dataset_temp_labels.npy', labels)
def SortWaveDataset(dataset):
captions = []
for iloop in range(0, feature.FEATURE_SIZE()):
captions.append('f_%u' % iloop)
captions.append('label')
captions.append('ts_code')
captions.append('pre_on_date')
captions.append('on_date')
captions.append('pre_off_date')
captions.append('off_date')
captions.append('holding_days')
data_df = pd.DataFrame(dataset, columns=captions)
data_df = data_df.sort_values(by=['pre_on_date', 'ts_code'],
ascending=(True, True))
return data_df.values
def Predict(test_data, model, mean, std):
predict_features = test_data[:,
feature.COL_FEATURE_OFFSET(
):feature.COL_FEATURE_OFFSET() +
feature.FEATURE_SIZE()]
predict_features = train_rnn.FeaturesPretreat(predict_features, mean, std)
predictions = model.predict(predict_features)
print("Predict out")
predictions_df = pd.DataFrame(predictions, columns=['pred'])
temp_index = feature.COL_ACTURE_OFFSET(0)
acture_unit_size = feature.ACTURE_UNIT_SIZE()
t0_acture_data = test_data[:, temp_index:temp_index + acture_unit_size]
acture_data_df = pd.DataFrame(t0_acture_data, columns=[ \
'T0_open_increse', \
'T0_low_increase', \
'T0_open', \
'T0_low', \
'T0_close', \
'stock_code', \
'T0_trade_date'])
temp_index = feature.COL_ACTURE_OFFSET(feature.ACTIVE_LABEL_DAY())
td_acture_data = test_data[:, temp_index:temp_index + acture_unit_size]
temp_df=pd.DataFrame(td_acture_data, columns=[ \
'Td_open_increse', \
'Td_low_increase', \
'Td_open', \
'Td_low', \
'Td_close', \
'Td_stock_code', \
'Td_trade_date'])
acture_data_df = pd.merge(acture_data_df,
temp_df,
left_index=True,
right_index=True)
result_all = predictions_df
result_all = pd.merge(result_all,
acture_data_df,
left_index=True,
right_index=True)
return result_all
def Test(self):
date_num = self.test_dataset.shape[0]
code_num = self.test_dataset.shape[1]
curren_status = STATUS_OFF
trade_count = 0
increase_sum = 0.0
for dloop in reversed(range(0, date_num)): # 遍历dataset的日期
if curren_status == STATUS_OFF:
code_index = random.randint(0, code_num - 1)
if self.test_dataset[dloop][code_index][date_col_index] == 0.0:
continue
current_feature = self.test_dataset[dloop][code_index][
0:feature.FEATURE_SIZE()]
Q = self.GetTestActionQ(current_feature)
if Q > 0.01:
# print(Q)
next_status = STATUS_ON
else:
next_status = STATUS_OFF
if curren_status == STATUS_OFF:
if next_status == STATUS_ON:
curren_status = STATUS_ON
t1_date_index = self.NextValidDateIndex(
self.test_dataset, code_index, dloop)
in_price = self.test_dataset[t1_date_index][code_index][
open_col_index]
else:
if next_status == STATUS_OFF:
curren_status = STATUS_OFF
t1_date_index = self.NextValidDateIndex(
self.test_dataset, code_index, dloop)
out_price = self.test_dataset[t1_date_index][code_index][
open_col_index]
increase = out_price / in_price - 1.0
increase_sum += increase
trade_count += 1
print("Test increase_sum:%.3f, trade_count:%u" %
(increase_sum, trade_count))
return increase_sum
def __init__(self):
self.replay_buffer = deque()
self.train_dataset, self.test_dataset = dqn_dataset.GetDataSet()
self.test_dataset = self.train_dataset
self.step_num = 0
self.LossClean()
if ModelExist():
self.epsilon = FINAL_EPSILON
self.model, self.mean, self.std = LoadModel()
else:
self.epsilon = INITIAL_EPSILON
self.model = BuildModel()
temp_shape = self.train_dataset.shape
train_data_2d = self.train_dataset.reshape(
(temp_shape[0] * temp_shape[1], temp_shape[2]))
row_rand_array = np.arange(train_data_2d.shape[0])
np.random.shuffle(row_rand_array)
train_data_2d_sample = train_data_2d[row_rand_array[0:10000]]
train_features = train_data_2d_sample[:, 0:feature.FEATURE_SIZE()]
self.mean = train_features.mean(axis=0)
self.std = train_features.std(axis=0)
def GetTrainTestDataSampleByDate(test_ratio):
sample_num = int(1.0 / test_ratio + 0.0001)
dataset = GetTrainTestDataMerge()
print("dataset: {}".format(dataset.shape))
pos = ((dataset[:, wave_kernel.COL_ON_PRETRADE_DATE()].astype(int) % 100) %
sample_num) == 0
test_data = dataset[pos]
train_data = dataset[~pos]
print("train: {}".format(train_data.shape))
print("test: {}".format(test_data.shape))
train_features = train_data[:, 0:feature.FEATURE_SIZE()]
train_labels = train_data[:,
feature.FEATURE_SIZE():feature.FEATURE_SIZE() +
1]
test_features = test_data[:, 0:feature.FEATURE_SIZE()]
test_labels = test_data[:,
feature.FEATURE_SIZE():feature.FEATURE_SIZE() + 1]
return train_features, train_labels, test_features, test_labels, test_data
def GetTrainTestDataSplitByDate():
dataset = GetTrainTestDataMerge()
print("dataset: {}".format(dataset.shape))
pos = dataset[:, wave_kernel.COL_ON_PRETRADE_DATE(
)] < dataset_train_test_split_date
train_data = dataset[pos]
test_data = dataset[~pos]
test_data = SortWaveDataset(test_data)
print("train: {}".format(train_data.shape))
print("test: {}".format(test_data.shape))
train_features = train_data[:, 0:feature.FEATURE_SIZE()]
train_labels = train_data[:,
feature.FEATURE_SIZE():feature.FEATURE_SIZE() +
1]
test_features = test_data[:, 0:feature.FEATURE_SIZE()]
test_labels = test_data[:,
feature.FEATURE_SIZE():feature.FEATURE_SIZE() + 1]
return train_features, train_labels, test_features, test_labels, test_data
def SettingName():
temp_name = '%s_%s_%u_%u_%f' % (dqn_dataset.SettingName(),
tushare_data.train_test_date,
dqn_dataset.dataset_train_test_split_date,
label_days,
decay_ratio)
return temp_name
if import_tushare_data and import_dqn_dataset:
setting_name_ = SettingName()
dqn_dataset_file_name_ = dqn_dataset.FileNameDataSet(False)
dqn_dataset_ACTURE_DATA_INDEX_DATE_ = dqn_dataset.ACTURE_DATA_INDEX_DATE()
dqn_dataset_ACTURE_DATA_INDEX_OPEN_ = dqn_dataset.ACTURE_DATA_INDEX_OPEN()
dqn_dataset_ACTURE_DATA_INDEX_TSCODE_ = dqn_dataset.ACTURE_DATA_INDEX_TSCODE()
feature_FEATURE_SIZE_ = feature.FEATURE_SIZE()
dqn_dataset_dataset_train_test_split_date_ = dqn_dataset.dataset_train_test_split_date
else:
dqn_dataset_ACTURE_DATA_INDEX_DATE_ = 156
dqn_dataset_ACTURE_DATA_INDEX_OPEN_ = 152
dqn_dataset_ACTURE_DATA_INDEX_TSCODE_ = 155
feature_FEATURE_SIZE_ = 150
dqn_dataset_dataset_train_test_split_date_ = 20170101
setting_name_ = '0_30_1_10_0_0___20000101_10_20020101_20190414_%u_%u_%.6f' % (
dqn_dataset_dataset_train_test_split_date_,
label_days,
decay_ratio)
dqn_dataset_file_name_ = './data/dataset/dqn_0_30_1_10_0_0___20000101_10_20020101_20190414.npy'
def FileNameDataSet():
def TestTop1(self, print_trade_detail=False):
test_features = self.test_dataset[:, :, 0:feature.FEATURE_SIZE()]
predictions = self.GetTestActionQ(test_features)
# print("predictions:{}".format(predictions.shape))
max_Q_codes_index = np.argmax(predictions, axis=1)
# print("max_Q_codes_index:{}".format(max_Q_codes_index.shape))
max_Q_codes_value = np.amax(predictions, axis=1)
max_Q_mean = np.mean(max_Q_codes_value)
# print("max_Q_mean:{}".format(max_Q_mean))
date_num = self.test_dataset.shape[0]
code_num = self.test_dataset.shape[1]
curren_status = STATUS_OFF
trade_count = 0
increase_sum = 0.0
hold_days_sum = 0
if print_trade_detail:
print('%-8s%-10s%-10s%-10s%-10s%-10s%-10s%-10s%-10s' %
('index', 'in_date', 'out_date', 'ts_code', 'pred', 'in',
'out', 'increase', 'hold_days'))
print('-' * 80)
dloop = date_num - 1
while dloop >= 0: # 遍历dataset的日期
if curren_status == STATUS_OFF:
code_index = max_Q_codes_index[dloop]
if self.test_dataset[dloop][code_index][date_col_index] == 0.0:
dloop -= 1
continue
Q = predictions[dloop][code_index]
if Q > 0:
# print(Q)
next_status = STATUS_ON
else:
next_status = STATUS_OFF
if curren_status == STATUS_OFF:
if next_status == STATUS_ON:
curren_status = STATUS_ON
t1_date_index = self.NextValidDateIndex(
self.test_dataset, code_index, dloop)
if t1_date_index < 0:
break
in_price = self.test_dataset[t1_date_index][code_index][
open_col_index]
in_pred = Q
dloop = t1_date_index
else:
dloop -= 1
else:
if next_status == STATUS_OFF:
curren_status = STATUS_OFF
t2_date_index = self.NextValidDateIndex(
self.test_dataset, code_index, dloop)
if t2_date_index < 0:
break
out_price = self.test_dataset[t2_date_index][code_index][
open_col_index]
increase = out_price / in_price - 1.0
hold_days = t1_date_index - t2_date_index
if print_trade_detail:
print(
'%-8u%-10.0f%-10.0f%-10s%-10.2f%-10.2f%-10.2f%-10.4f%-10u'
% (trade_count, self.test_dataset[t1_date_index]
[code_index][date_col_index],
self.test_dataset[t2_date_index][code_index]
[date_col_index],
'%06u' % self.test_dataset[t1_date_index]
[code_index][tscode_col_index], in_pred,
in_price, out_price, increase, hold_days))
increase_sum += increase
trade_count += 1
hold_days_sum += hold_days
dloop = t2_date_index
else:
dloop -= 1
if print_trade_detail:
print('%-8s%-10s%-10s%-10s%-10s%-10s%-10s%-10.4f%-10u' %
('sum', '--', '--', '--', '--', '--', '--', increase_sum,
hold_days_sum))
return increase_sum, trade_count, max_Q_mean