simple_LSTM_ed

# import essentials
import matplotlib.pyplot as plt
import pandas as pd
import numpy as np

#
from sklearn.preprocessing import MinMaxScaler

# import keras
from keras.models import Sequential
from keras.layers import LSTM, Dense

# read the data
# data = pd.read_csv("")

# create array to train, validate, and test
pure = np.linspace(0, 100, 200)
noise = np.random.normal(-1, 1, 200)
signal = pure + noise + np.sin(pure) + np.sin(2*pure)

plt.plot(signal, '-')

# set splitting functions to train, validation, and test sets
dataset = np.array(signal).reshape(-1, 1)
window = 10

scaler = MinMaxScaler(feature_range=(0, 1))
dataset = scaler.fit_transform(dataset)

def split(_dataset, p):
  train_size = int(len(_dataset) * p)
  test_size = len(_dataset) - train_size
  return _dataset[0:train_size,:], _dataset[train_size:len(_dataset), :]

def prepare(_dataset, _look_back=1):
    data_x, data_y = [], []
    for i in range(len(_dataset) - _look_back):
      a = _dataset[i:(i + _look_back), 0]
      data_x.append(a)
      data_y.append(_dataset[i + _look_back, 0]) # increase to 1 dim
    data_x = np.array(data_x)
    data_x = np.reshape(data_x, (data_x.shape[0], data_x.shape[1], 1))
    return data_x, np.array(data_y)

train, test = split(dataset, 0.8)
train, val = split(train, 0.6)

train_x, train_y = prepare(train, window)
val_x, val_y = prepare(val, window)
test_x, test_y = prepare(test, window)

# LSTM model

model = Sequential()
model.add(LSTM(window, return_sequences=True, input_shape=(window, 1)))
model.add(LSTM(window, return_sequences=True))
model.add(LSTM(window))
model.add(Dense(1))

model.compile(loss='mean_squared_error', optimizer='adam',)

# train the model
model.fit(train_x, train_y, validation_data=(val_x, val_y),
          epochs=10, batch_size=10, verbose=0)

# make predictions on concatenated set - train, validation, test together
prediction = np.concatenate((train_x, val_x), axis=0)
pred = model.predict(np.concatenate((prediction, test_x), axis=0))

# do the scaler inversion
pred = scaler.inverse_transform(pred)
dataset = scaler.inverse_transform(dataset)

# reshape dataset for plotting
dataset = np.array(signal).reshape(-1, 1)

# Commented out IPython magic to ensure Python compatibility.
# %matplotlib inline
def to_plot(dataset, predict, start, end):
  predict_plot = np.empty_like(dataset)
  predict_plot[:, :] = np.nan
  predict_plot[start:end, :] = predict
  return predict_plot

start = window+1
end =  window + len(pred)+1

pred_plot = to_plot(dataset, pred, start, end)

plt.plot(dataset, color='g', label='dataset')
plt.plot(pred_plot, color='m', label='prediction')
plt.legend()
plt.title("prediction for the 200 length dataset")
plt.show()

x = scaler.inverse_transform(model.predict(test_x))
y = dataset[-len(test_x):]
p = x - y
plt.figure()
plt.plot(p[:,0],)
plt.title("error")
plt.xlabel("index")
plt.ylabel("error")

"""Let's repeat all, but for the larger dataset"""

# create array to train, validate, and test with 2000 values
pure = np.linspace(0, 100, 2000)
noise = np.random.normal(-1, 1, 2000)
signal = pure + noise + np.sin(pure) + np.sin(2*pure)

plt.plot(signal, '-')

# set splitting functions to train, validation, and test sets
dataset = np.array(signal).reshape(-1, 1)
window = 10

scaler = MinMaxScaler(feature_range=(0, 1))
dataset = scaler.fit_transform(dataset)

train, test = split(dataset, 0.8)
train, val = split(train, 0.6)

train_x, train_y = prepare(train, window)
val_x, val_y = prepare(val, window)
test_x, test_y = prepare(test, window)

# train the model
model.fit(train_x, train_y, validation_data=(val_x, val_y),
          epochs=10, batch_size=10, verbose=0)

# make predictions on concatenated set - train, validation, test together
prediction = np.concatenate((train_x, val_x), axis=0)
pred = model.predict(np.concatenate((prediction, test_x), axis=0))

# do the scaler inversion
pred = scaler.inverse_transform(pred)
dataset = scaler.inverse_transform(dataset)

# reshape dataset for plotting
dataset = np.array(signal).reshape(-1, 1)

# Commented out IPython magic to ensure Python compatibility.
# %matplotlib inline
start = window+1
end =  window + len(pred)+1

pred_plot = to_plot(dataset, pred, start, end)

plt.plot(dataset, color='g', label='dataset')
plt.plot(pred_plot, color='m', label='prediction')
plt.legend()
plt.title("prediction for the 2000 length dataset")
plt.show()

x = scaler.inverse_transform(model.predict(test_x))
y = dataset[-len(test_x):]
p = x - y
plt.figure()
plt.plot(p[:,0],)
plt.title("error")
plt.xlabel("index")
plt.ylabel("error")