Adding multiple repeats for LSTM network

Author: JBris

python/shampoo_sales/lstm_neural_network.py

@@ -28,9 +28,8 @@ def forecast_lstm(model, batch_size, X):
 	yhat = model.predict(X, batch_size=batch_size)
 	return yhat[0,0]
 
-def prepare_model():
+def prepare_model(series):
     # Convert series into supervised learning problem
-    series = load_date_series()
     X = series.values
     supervised = timeseries_to_supervised(X, 1)
     print("*** Supervised Learning ***")
@@ -62,14 +61,11 @@ def model(series):
     # transform data to be stationary
     raw_values = series.values
     diff_values = difference(raw_values, 1)
-
     # transform data to be supervised learning
     supervised = timeseries_to_supervised(diff_values, 1)
     supervised_values = supervised.values
-
     # split data into train and test-sets
     train, test = supervised_values[0:-12], supervised_values[-12:]
-
     # transform the scale of the data
     scaler, train_scaled, test_scaled = scale(train, test)
 
@@ -80,7 +76,7 @@ def model(series):
     lstm_model.predict(train_reshaped, batch_size=1)
 
     # walk-forward validation on the test data
-    predictions = list()
+    predictions = []
     for i in range(len(test_scaled)):
         # make one-step forecast
         X, y = test_scaled[i, 0:-1], test_scaled[i, -1]
@@ -102,5 +98,52 @@ def model(series):
     plt.plot(predictions)
     plt.show()
 
+def multiple_repeats(series):
+    # transform data to be stationary
+    raw_values = series.values
+    diff_values = difference(raw_values, 1)
+    # transform data to be supervised learning
+    supervised = timeseries_to_supervised(diff_values, 1)
+    supervised_values = supervised.values
+    # split data into train and test-sets
+    train, test = supervised_values[0:-12], supervised_values[-12:]
+    # transform the scale of the data
+    scaler, train_scaled, test_scaled = scale(train, test)
+    
+    # repeat experiment
+    repeats = 30
+    error_scores = []
+    for r in range(repeats):
+        # fit the model
+        lstm_model = fit_lstm(train_scaled, 1, 3000, 4)
+        # forecast the entire training dataset to build up state for forecasting
+        train_reshaped = train_scaled[:, 0].reshape(len(train_scaled), 1, 1)
+        lstm_model.predict(train_reshaped, batch_size=1)
+        # walk-forward validation on the test data
+        predictions = []
+        for i in range(len(test_scaled)):
+            # make one-step forecast
+            X, y = test_scaled[i, 0:-1], test_scaled[i, -1]
+            yhat = forecast_lstm(lstm_model, 1, X)
+            # invert scaling
+            yhat = invert_scale(scaler, X, yhat)
+            # invert differencing
+            yhat = inverse_difference(raw_values, yhat, len(test_scaled)+1-i)
+            # store forecast
+            predictions.append(yhat)
+        # report performance
+        rmse = sqrt(mse(raw_values[-12:], predictions))
+        print('%d) Test RMSE: %.3f' % (r+1, rmse))
+        error_scores.append(rmse)
+    
+    # summarize results
+    results = pd.DataFrame()
+    results['rmse'] = error_scores
+    print(results.describe())
+    results.boxplot()
+    plt.show()
+
 series = load_date_series()
-model(series)
+# prepare_model(series)
+# model(series)
+multiple_repeats(series)