def read_data():
from assets import models
df = pd.read_csv('assets/data.csv')
for index, _df in df.iterrows():
date = datetime.datetime.strptime(_df['date'],'%Y/%m/%d').date()
row = models.Data(date=date, subscribers=_df['subscribers'],reviews=_df['reviews'])
db_session.add(row)
db_session.commit()
def read_data(): #初期データの読込(csv)
from assets import models
#from assets.database import db_session
import pandas as pd
import datetime
df = pd.read_csv('japanese-stock-price.csv')
for index, _df in df.iterrows():
date = datetime.datetime.strptime(_df['date'],'%Y/%m/%d').date()
row = models.Data(date = date, subscribers = _df['subscribers'], reviews = _df['reviews'])
db_session.add(row)
db_session.commit()
def write_data():
from assets import models
# 2019/02/01～の情報を取得（UTCとの時差は9時間）（close_timeは日本時間で午前9時）
res = requests.get('https://api.cryptowat.ch/markets/bitflyer/btcjpy/ohlc?periods=86400&after=1548979200')
data = json.loads(res.text)
# 取得した情報をデータフレームに格納
daily_data = pd.DataFrame(data['result']['86400'], columns=['close_time', 'open_price', 'high_price', 'low_price', 'close_price', 'volume', 'quote_volume'])
daily_data['date'] = pd.to_datetime(daily_data['close_time'], unit='s')
daily_data = daily_data[:-1] # 最終行を削除（取得した時点の情報が次の日の情報となっているため）
# 日本時間に変更
diff_JST_from_UTC = 9
daily_data['date'] = daily_data['date'] + datetime.timedelta(hours=diff_JST_from_UTC)
# 過去相場の予測結果をDBに格納するために以下で学習済みモデルへの入力データを作成
period = 5
data = daily_data['close_price']
# 各データをリストに格納するためにリストを作成
input_tensor = []
label_tensor = []
mean_list = []
std_list = []
data_len = len(data) #総データ数
#変数とラベルの生成
for i in range(0, data_len - period, 1):
window = data.values[i:i + period]
mean = np.mean(window)
std = np.std(window)
mean_list.append(mean)
std_list.append(std)
input_tensor.append((window-mean) / std)
label_tensor.append((data.values[i + period]-mean) / std)
input_tensor = np.array(input_tensor).reshape(len(data)-period, period, 1)
label_tensor = np.array(label_tensor).reshape(len(data)-period, 1)
mean_list = np.array(mean_list)
std_list = np.array(std_list)
# label_tensorの最初はdaily_data['date']の５列目からなので日付データを整形
date = daily_data['date'][5:].reset_index(drop=True)
# 学習済みモデルを取得
with open("lstm_model.pickle", "rb") as f:
model = pickle.load(f)
# 学習済みモデルで予測
pred = model.predict(input_tensor)
pred = pred.reshape(-1)
pred = pred*std_list + mean_list
label_tensor = label_tensor.reshape(-1)
real = label_tensor*std_list + mean_list
past_df = pd.DataFrame({'date':date, 'pred':pred, 'real':real})
# 過去情報をDBに格納
for index, _past_df in past_df.iterrows():
row = models.Data(date = _past_df['date'], real_price = _past_df['real'], pred_price = _past_df['pred'])
db_session.add(row)
db_session.commit()
# デプロイした日の翌日の相場を予測し、DBに格納するコードを記述
# windowを作成し標準化
window = 5
data_five = db_session.query(models.Data).all()[-window:]
input_will = []
for price in data_five:
input_will.append(price.real_price)
input_will = np.array(input_will).reshape(1, window, 1)
input_mean = np.mean(input_will)
input_std = np.std(input_will)
sta_input = (input_will - input_mean) / input_std
# 翌日の価格を予測、標準化してあるので元に戻す
future_price = model.predict(sta_input)
future_price = future_price.reshape(-1)
future_price = future_price*input_std + input_mean
future_price = float(future_price)
# 翌日の日付を取得
today = db_session.query(models.Data).all()[-1].date
next_day = today + datetime.timedelta(days=1)
# 翌日の日付、予測した価格をデータベースに格納
row = models.Data(date = next_day, real_price = None, pred_price = future_price)
db_session.add(row)
db_session.commit()