Fonctionnement des LSTM avant entrainement

code/RNN/Model.py

@@ -5,16 +5,22 @@ class MyModel(tf.keras.Model):
 	def __init__(self, HIDDEN):
 		super(MyModel, self).__init__()
 		self.lstm1 = tf.keras.layers.LSTM(HIDDEN, return_sequences=True)
-		self.lstm2 = tf.keras.layers.LSTM(HIDDEN, return_sequences=True)
+		#self.lstm2 = tf.keras.layers.LSTM(HIDDEN, return_sequences=True)
 		#self.lstm3 = tf.keras.layers.LSTM(HIDDEN, return_sequences=True)
 		#self.lstm4 = tf.keras.layers.LSTM(HIDDEN, return_sequences=True)
-		self.lstmlast = tf.keras.layers.LSTM(1, return_sequences=True)
+		self.lstmlast = tf.keras.layers.LSTM(HIDDEN, return_sequences=True)
+		self.dense1 = tf.keras.layers.Dense(HIDDEN, activation='relu')
+		self.dense2 = tf.keras.layers.Dense(HIDDEN//2, activation='relu')
+		self.denselast = tf.keras.layers.Dense(1, activation='sigmoid')
 
 
 	def call(self, inputs):
 		x = self.lstm1(inputs)
-		x = self.lstm2(x)
+		#x = self.lstm2(x)
 		#x = self.lstm3(x)
 		#x = self.lstm4(x)
 		x = self.lstmlast(x)
+		x = self.dense1(x)
+		x = self.dense2(x)
+		x = self.denselast(x)
 		return x

code/RNN/training.py

@@ -3,18 +3,18 @@ from matplotlib import pyplot as plt
 import numpy as np
 from Model import MyModel
 
-LEN_TRAIN = 5000
-LEN_SEQ = 100
-PRED = 0.02
+
+LEN_SEQ = 64
+PRED = 0.05
 HIDDEN = 128
 
 model = MyModel(HIDDEN)
 
 dataset = np.load('dataset.npy')
-datasetp = np.roll(dataset, 1)
-datasetp[0] = dataset[0]
-data = (dataset - datasetp)/datasetp
-data = data*2/(max(data) - min(data))
+scale = (np.max(dataset) - np.min(dataset))
+data = dataset/scale
+shift = np.min(data)
+data = data - shift
 
 annee = np.array(list(range(len(data))))/365
 annee = annee - annee[-1]
@@ -27,11 +27,20 @@ plt.plot(annee[start_pred:],data[start_pred:], label="validation")
 plt.legend()
 plt.show()
 
-X_train = [data[0:start_pred-1]]
-Y_train = [data[1:start_pred]]
+X_train_tot = [data[0:start_pred-1]]
+Y_train_tot = [data[1:start_pred]]
+
+X_train_tot = np.expand_dims(np.array(X_train_tot),2)
+Y_train_tot = np.expand_dims(np.array(Y_train_tot),2)
 
-X_train = np.expand_dims(np.array(X_train),2)
-Y_train = np.expand_dims(np.array(Y_train),2)
+X_train = X_train_tot[:,:LEN_SEQ,:]
+Y_train = Y_train_tot[:,:LEN_SEQ,:]
+
+for i in range(len(X_train_tot[0]) - LEN_SEQ) :
+	X_train = np.concatenate((X_train, X_train_tot[:,i:i+LEN_SEQ,:]),0)
+	Y_train = np.concatenate((Y_train, Y_train_tot[:,i:i+LEN_SEQ,:]),0)
+print(X_train_tot.shape)
+print(X_train.shape)
 
 
 model.compile(optimizer='adam',
@@ -40,17 +49,16 @@ model.compile(optimizer='adam',
 import os 
 os.system("rm -rf log_dir")
 
-model.fit(x=X_train, y=Y_train, epochs=30)
+model.fit(x=X_train, y=Y_train, batch_size=16, epochs=5, shuffle=True)
 
-Pred = X_train.copy()
+Pred = X_train_tot.copy()
 while len(Pred[0]) < len(data) :
 	print(len(data) - len(Pred[0]))
 	Pred = np.concatenate((Pred, np.array([[model.predict(Pred)[0][-1]]])),1)
-Pred = Pred/2*(max(data) - min(data))
-data_Pred = dataset.copy()
-Pred = np.squeeze(Pred)
-for i in range(start_pred,len(data_Pred)) :
-	data_Pred[i] = data_Pred[i-1]*Pred[i] + data_Pred[i-1]
+Pred = Pred + shift
+Pred = Pred * scale
+
+data_Pred = np.squeeze(Pred)
 
 plt.figure(2)
 plt.plot(annee[:start_pred],dataset[:start_pred], label="apprentissage")