LSTM.py

neural_networks_models/LSTM.py

@@ -0,0 +1,130 @@
+import keras.backend as K
+import numpy as np
+import pandas as pd
+from keras.callbacks import EarlyStopping, ModelCheckpoint
+from keras.layers.core import Dense, Dropout
+from keras.layers.recurrent import LSTM
+from keras.models import Sequential
+from sklearn.preprocessing import MinMaxScaler
+
+
+def embedding(data, p):
+    data_shifted = data.copy()
+    for lag in range(-p + 1, 2):
+        data_shifted['y_t' + '{0:+}'.format(lag)] = data_shifted['y'].shift(-lag, freq='D')
+    data_shifted = data_shifted.dropna(how='any')
+    y = data_shifted['y_t+1'].to_numpy()
+    X = data_shifted[['y_t' + '{0:+}'.format(lag) for lag in range(-p + 1, 1)]].to_numpy()
+    return X, y, data_shifted
+
+
+def smape(y_true, y_pred):
+    denominator = (y_true + np.abs(y_pred)) / 200.0
+    diff = np.abs(y_true - y_pred) / denominator
+    diff[denominator == 0] = 0.0
+    return np.mean(diff)
+
+
+def smapeLSTM(true,predicted):
+    return K.abs(predicted - true) / K.maximum(K.abs(true) + K.abs(predicted) + 0.1, 0.5 + 0.1) * 2.0
+
+
+def keyvalue(df):
+    df["horizon"] = range(1, df.shape[0] + 1)
+    res = pd.melt(df, id_vars=["horizon"])
+    res = res.rename(columns={"variable": "series"})
+    res["Id"] = res.apply(lambda row: "s" + str(row["series"].split("-")[1]) + "h" + str(row["horizon"]), axis=1)
+    res = res.drop(['series', 'horizon'], axis=1)
+    res = res[["Id", "value"]]
+    res = res.rename(columns={"value": "forecasts"})
+    return res
+
+
+def main():
+    data = pd.read_csv("/home/damien/HandsOnAI2019/Challenges/Defi3/train.csv", index_col="Day")
+
+    data.index = pd.to_datetime(data.index, format="%Y-%m-%d")
+    data = data.asfreq('d')
+
+    interval_train = pd.date_range(start='2015-07-01', end='2017-06-30')
+    interval_valid = pd.date_range(start='2017-07-1', end='2017-08-20')
+
+    interval_test = pd.date_range(start='2017-08-21', end='2017-09-10')
+    HORIZON = len(interval_test)
+
+    data_train = data.loc[interval_train]
+    data_valid = data.loc[interval_valid]
+
+    forecasts_nn = pd.DataFrame(index=interval_test)
+
+    for iseries in data_train.columns:
+        print(iseries)
+
+        train = data_train[iseries].to_frame()
+        valid = data_valid[iseries].to_frame()
+
+        BATCH_SIZE = 32  # number of samples per mini-batch
+        EPOCHS = 2500  # maximum number of times the training algorithm will cycle through all samples
+
+        p = 15
+
+        scaler = MinMaxScaler()
+
+        train["y"] = scaler.fit_transform(train)
+        X_train, y_train, train_embedded = embedding(train, p)
+
+        valid['y'] = scaler.transform(valid)
+        X_valid, y_valid, valid_embedded = embedding(valid, p)
+
+        X_train = np.reshape(X_train, (X_train.shape[0], X_train.shape[1], 1))
+        X_valid = np.reshape(X_valid, (X_valid.shape[0], X_valid.shape[1], 1))
+
+        model = Sequential()
+        model.add(LSTM(256, return_sequences=True, input_shape=(X_train.shape[1], 1)))
+        model.add(Dropout(0.2))
+        model.add(LSTM(128, return_sequences=True))
+        model.add(Dropout(0.2))
+        model.add(LSTM(64, return_sequences=True))
+        model.add(Dropout(0.2))
+        model.add(LSTM(32, return_sequences=True))
+        model.add(Dropout(0.1))
+        model.add(LSTM(16, return_sequences=False))
+        model.add(Dropout(0.2))
+        model.add(Dense(1, activation="relu"))
+        model.compile(loss=smapeLSTM, optimizer="adam")
+
+        best_val = ModelCheckpoint('model_{epoch:02d}.h5', save_best_only=True, mode='min', period=1)
+        earlystop = EarlyStopping(monitor='val_loss', min_delta=0, patience=20)
+        model.fit(X_train, y_train, epochs=EPOCHS, batch_size=BATCH_SIZE, validation_data=(X_valid, y_valid), callbacks=[earlystop], verbose=2)
+
+        x = X_valid[-1, 1:]
+        ypred = y_valid[-1]
+
+        f_nn = []
+        for horizon in range(1, HORIZON + 1):
+            x_test = np.expand_dims(np.append(x, ypred), axis=0)
+            x_test = np.reshape(x_test, (x_test.shape[0], x_test.shape[1], 1))
+            ypred = model.predict(x_test)
+            f_nn.append(float(ypred))
+
+        # f_nn = []
+        # tests = []
+        # for horizon in range(1, HORIZON + 1):
+        #     if len(tests) > 7:
+        #         x_test = np.expand_dims(np.append(x, tests[-7]), axis=0)
+        #     else:
+        #         x_test = np.expand_dims(np.append(x, y_valid[horizon-7]), axis=0)
+        #     x_test = np.reshape(x_test, (x_test.shape[0], x_test.shape[1], 1))
+        #     ypred = model.predict(x_test)
+        #     f_nn.append(float(ypred))
+        #     tests.append(ypred)
+
+        f_nn = np.expand_dims(f_nn, axis=0)
+        forecasts_nn[iseries] = scaler.inverse_transform(f_nn).flatten()
+
+    pred_nn = keyvalue(forecasts_nn)
+    pred_nn.to_csv("submissionLSTMRelu2500.csv", index=False)
+
+
+if __name__ == "__main__":
+    main()