darden1

model_myrnn_rsf = RecurrentNeuralNetwork(rnn_units, return_sequences=False)
model_myrnn_rsf.fit(X_train, Y_train_rsf, 
              batch_size=batch_size, 
              epochs=n_epochs, 
              mu=lr, 
              validation_data=(X_val, Y_val_rsf), 
              verbose=1)

plt.plot(indices, history_rsf.history["loss"], label="loss (keras)")
plt.plot(indices, history_rsf.history["val_loss"], label="val_loss (keras)")

darden1

model_myrnn_rst = RecurrentNeuralNetwork(rnn_units, return_sequences=True)
model_myrnn_rst.fit(X_train, Y_train, 
              batch_size=batch_size, 
              epochs=n_epochs, 
              mu=lr, 
              validation_data=(X_val, Y_val), 
              verbose=1)

plt.plot(indices, history_rst.history["loss"], label="loss (keras)")
plt.plot(indices, history_rst.history["val_loss"], label="val_loss (keras)")

darden1

class RecurrentNeuralNetwork():
    def __init__(self, rnn_units=10, rnn_activation="tanh", return_sequences=False, random_state=0):
        self.init_state = True   # 重みの初期化判定フラグ
        self.loss = None         # トレーニングデータのLoss
        self.val_loss = None     # テストデータのLoss
        self.acc = None          # トレーニングデータの正答率
        self.val_acc = None      # テストデータの正答率
        self.n_layers = 3  # 全レイヤーの数 - 1
        self.rnn_units = rnn_units  # 隠れ層のサイズ
        self.rnn_activation = rnn_activation  # 活性化関数の名前

darden1

class TimeSeriesDense(Dense):
    
    def forward_prop(self, Phi):
        """順伝播演算"""
        n_samples, T = Phi.shape[0], Phi.shape[1]
        Z = np.zeros([n_samples, T, self.units])
        
        for t in range(T):
            Z[:,t,:] = np.dot(Phi[:,t,:], self.W)  + self.b
        

darden1

class RNN:
    def __init__(self, units, input_dim, activation="tanh", 
                 kernel_initializer="glorot_normal", 
                 bias_initializer='zeros', 
                 return_sequences=False):
        self.units = units
        self.input_dim = input_dim
        self.activation = activation
        self.kernel_initializer = kernel_initializer
        self.bias_initializer = bias_initializer

darden1

# ターゲットデータを最終時間のみにする
Y_train_rsf, Y_val_rsf = Y_train[:, -1, :], Y_val[:, -1, :]

model_rsf = Sequential()
model_rsf.add(SimpleRNN(rnn_units, input_shape=(n_sequence, n_features), return_sequences=False))
model_rsf.add(Dense(n_classes, activation="linear"))
model_rsf.compile(loss='mean_squared_error', optimizer=SGD(lr))
history_rsf = model_rsf.fit(X_train, Y_train_rsf,
                            batch_size=batch_size,
                            epochs=n_epochs,

darden1

from keras.models import Sequential
from keras.layers import Dense, SimpleRNN
from keras.optimizers import SGD

n_train = 80

X_train, X_val = X[:n_train], X[n_train:]
Y_train, Y_val = Y[:n_train], Y[n_train:]

batch_size = 10

darden1

import numpy as np
import matplotlib.pyplot as plt
np.random.seed(0)

n_sequence = 5
n_data = int(20+1)*n_sequence

a, b = 1., 1.
phase = np.pi / 4
time = np.linspace(0, 1, n_data)

darden1

plt.figure(figsize=(10, 7))
plt.subplot(211)
plt.title("learning log (loss)")
plt.xlabel("epoch")
plt.ylabel("loss")
plt.plot(np.arange(len(clf.loss)), clf.loss, label="my train")
plt.plot(np.arange(len(clf.loss)), clf.val_loss, label="my test")
plt.plot(np.arange(len(history.history["loss"])), history.history["loss"], label="keras train")
plt.plot(np.arange(len(history.history["loss"])), history.history["val_loss"], label="keras test")
plt.plot(np.arange(len(clf_sk.loss_curve_)),clf_sk.loss_curve_, label="sklearn train")

darden1

from keras.models import Sequential
from keras.layers.core import Dense, Activation
from keras.optimizers import SGD
from keras.losses import categorical_crossentropy
from keras.initializers import he_normal

n_features = X.shape[1]
n_classes = Y.shape[1]

batch_size = int(len(X_train)*0.2)  # ミニバッチサイズ