さくら VPS LSTM 豪ドル/円予測
さくら VPS LSTM 豪ドル/円４時間足予測ファイル
さくら VPS LSTM 豪ドル/円４時間足予測ファイル(サイレント) keras-go-lstm-4h.py
をまとめます。
RNN 豪ドル/円４時間足予測最新表示サイレント化ファイル keras-go-4h-5.py
から変更した部分は赤字にしています。
# -*- coding: utf-8 -*-
#!/usr/bin/python3
import os
os.environ['TF_CPP_MIN_LOG_LEVEL']='1'
import datetime
import numpy as np
import pandas as pd
from matplotlib import pyplot as plt
from tensorflow import keras
from mplfinance.original_flavor import candlestick_ohlc
def addBusinessDays_4h(current_date, business_days_to_add_h):
# current_date : 処理する初日の日時 例) 2022-02-03 00:00:00
# business_days_to_add_h : 4時間ずつ延長する回数
cur_date8 = [] # 21/03/30 12～
cur_date4 = [] # 3/30
man_f = 1 #月曜フラグ 0 or 1
doyo_f = 0 #土曜フラグ 0 or 1
current_date0 = current_date
weekday0 = current_date0.weekday()
if(weekday0 == 0): #月
man_f = 0
while business_days_to_add_h > 0:
current_date += datetime.timedelta(hours=4)
weekday = current_date.weekday()
hour_doyo = current_date.hour
if weekday == 6: # sunday = 6
continue
if((weekday == 5) and (hour_doyo == 8)): #土
doyo_f = 1
if((weekday == 5) and (doyo_f == 1)): #土
continue
if((weekday == 0) and (man_f == 1)): #月
man_f = 0
continue
# 年/月/日 時刻として文字列にする処理
# 例 22/03/30 12～
e11 = current_date.strftime("%y/%m/%d %H")
e22 = e11.replace("/0", "/")
e22 = e22 + "～"
e33 = e22[3:] # 22/ 削除
cur_date8.append(e22)
cur_date4.append(e33)
business_days_to_add_h -= 1
return cur_date8, cur_date4
xl_df = pd.read_csv(
"/home/yamada/public_html/manep_4h/mane_chart_go_4h.csv",
encoding="cp932")
raw_data0 = xl_df["終値(売り)"].values
raw_data = raw_data0.copy()
# 行列の平均、標準偏差を求めます。
mean = np.mean(raw_data)
# print("Mean", mean)
raw_data -= mean
std = np.std(raw_data)
# print("Std ",std)
# 標準偏差値に変換
raw_data /= std
# print("各標準偏差値")
# print(raw_data)
# 連続デ－タとする。一つおきは、2
sampling_rate = 1
# 過去20間隔デ－タをひとまとまりとして時系列予測する
sequence_length = 20
delay = sampling_rate * sequence_length
# print("delay:", delay)
batch_size = 32 # 適当
# 検証デ－タのスタ－ト値
num_half_samples = int(0.5 * len(raw_data))
train_dataset = keras.utils.timeseries_dataset_from_array(
raw_data,
targets=raw_data[delay:],
sampling_rate=sampling_rate,
sequence_length=sequence_length,
batch_size=batch_size,
)
val_dataset = keras.utils.timeseries_dataset_from_array(
raw_data[:-1],
targets=raw_data[delay:],
sampling_rate=sampling_rate,
sequence_length=sequence_length,
batch_size=batch_size,
start_index=num_half_samples,
)
test_dataset = keras.utils.timeseries_dataset_from_array(
raw_data,
targets=None,
sampling_rate=sampling_rate,
sequence_length=sequence_length,
batch_size=batch_size,
)
# テストデ－タセット表示
i = 0
for inputs_t in test_dataset:
inputs_n = inputs_t.numpy()
if i == 0:
# print("Start test:", inputs_n[0])
i = i + 1
# print("End test:", inputs_n[-1])
# from tensorflow import keras
from keras import layers
from keras import initializers
"""
inputs = keras.Input(shape=(sequence_length,))
x = layers.Flatten()(inputs)
x = layers.Dense(
20,
activation="tanh",
kernel_initializer='zeros'
)(x)
"""
inputs = keras.Input(shape=(sequence_length, 1))
x = layers.LSTM(20, kernel_initializer='zeros')(inputs)
outputs = layers.Dense(1)(x)
model = keras.Model(inputs, outputs)
# print("モデルア－キテクチャ")
# print(model.summary())
callbacks_list = [
keras.callbacks.EarlyStopping(
monitor="val_loss",
patience=6,
),
keras.callbacks.ModelCheckpoint(
"/home/yamada/public_html/colab/jena_dense_lstm_4h.keras",
save_best_only=True,
)
]
model.compile(optimizer="rmsprop", loss="mse", metrics=["mae"])
history = model.fit(
train_dataset,
epochs=40,
verbose=0,
validation_data=val_dataset,
callbacks=callbacks_list)
loss = history.history["mae"] # 平均絶対誤差(MAE)
val_loss = history.history["val_mae"]
epochs = range(1, len(loss) + 1)
model = keras.models.load_model(
"/home/yamada/public_html/colab/jena_dense_lstm_4h.keras")
# pre = model.predict(test_dataset)
pre = model.predict(test_dataset, verbose=0)
pre1 = np.reshape(pre, (-1))
# print(f"Test 予測値:")
# print(pre1)
future_test = inputs_t[-1:]
# print("future_test 最初の配列値")
# print(future_test)
future_result = []
for i in range(6):
test_data_f = np.reshape(future_test, (1, 20, 1))
# batch_predict = model.predict(test_data_f)
batch_predict = model.predict(test_data_f, verbose=0)
future_test = np.delete(future_test, 0)
future_test = np.append(future_test, batch_predict)
future_result = np.append(future_result, batch_predict)
# print("future_result :")
# print(future_result)
# ここまで colab-21 まとめ
len_raw_data = len(raw_data)
xx1 = np.arange(sequence_length, len_raw_data + 1)
xx3 = np.arange(len_raw_data, len_raw_data + 6)
pre_chg = pre.copy()
pre_chg *= std
pre_chg += mean
pre_chg1 = np.reshape(pre_chg, (-1))
# print("pre_chg1:")
# print(pre_chg1)
f_result = future_result.copy()
f_result *= std
f_result += mean
# ここから４時間ロ－ソク足表示
Open = xl_df["始値(売り)"].values
High = xl_df["高値(売り)"].values
Low = xl_df["安値(売り)"].values
Close = xl_df["終値(売り)"].values
Date = xl_df["日付"].values
Idx = xl_df.index
# X 軸日付処理
xDate = []
xD = []
dayf = 0
for i, key in enumerate(Date):
day00 = str(key)[11:19]
if((dayf == 1) and (day00 == '00:00:00')):
dayf = 0
else:
if((dayf == 0) and (day00 == '00:00:00')):
e4 = str(key)[4:10]
e6 = e4.replace("/0", "/")
e8 = e6.lstrip("/")
xDate.append(e8)
xD.append(i)
dayf = 1
tstr = Date[-1]
if(dayf == 0):
e4l = tstr[7:14]
e6l = e4l.replace("/0", "/")
e8l = e6l.lstrip("/")
xDate.append(e8l)
xD.append(i)
# ロ－ソク４時間足表示
ohlc = zip(
Idx, Open, High, Low, Close)
fig = plt.figure(
figsize=(8.34, 5.56))
ax = fig.add_subplot(1,1,1) #④
ax.grid()
plt.plot(xx1, pre_chg)
plt.plot(xx3, f_result, 'bo')
candlestick_ohlc(
ax, ohlc, width=0.5, alpha = 1,
colorup='r', colordown='g')
plt.xticks(xD, xDate)
plt.title('AUS$ / JPY chart')
plt.xlabel('Date')
plt.ylabel('Yen')
# plt.show
plt.savefig(
'/home/yamada/public_html/manep-img/mane_chart_go_lstm_4h.png')
# ここから４時間足予測最終日時取得
tstr = Date[-1]
tdatetime = datetime.datetime.strptime(tstr, '%Y/%m/%d %H:%M:%S')
lastday = addBusinessDays_4h(tdatetime, 6)
lastday_pd = pd.DataFrame(lastday[0])
f_result_pd = pd.DataFrame(np.round(f_result, 3))
# 有効数字3桁に変更
df_concat = pd.concat([lastday_pd, f_result_pd], axis = 1)
df_concat.to_csv(
'/home/yamada/public_html/manep-img/mane_chart_go_lstm_4h.csv',
header=False, index=False)
さくら VPS LSTM 豪ドル/円４時間足予測ファイル(サイレント) keras-go-lstm-4h.py
がまとまりました。
引き続き、このファイルを保存します。