mlp_variacoes.py

import os
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"   # see issue #152
os.environ["CUDA_VISIBLE_DEVICES"] = ""
os.environ['PYTHONHASHSEED'] = '0'
import sys
import numpy as np
import pandas
import math
import tensorflow as tf
import random
import matplotlib.pylab as plt
#import talib

seed = [4395,3129,277,9871,5183,6082,810,6979,2654,5765]

def set_seeds(seed):
    np.random.seed(seed)  # for reproducibility
    tf.set_random_seed(seed)
    random.seed(seed)

from processing import *


from keras.models import Sequential
from keras.layers.core import Dense, Activation, Dropout
from keras.optimizers import SGD, Adam
from keras.callbacks import CSVLogger, EarlyStopping, ModelCheckpoint, TensorBoard, ReduceLROnPlateau
from keras import regularizers

session_conf = tf.ConfigProto(intra_op_parallelism_threads=1, inter_op_parallelism_threads=1)


from keras import backend as K

sess = tf.Session(graph=tf.get_default_graph(), config=session_conf)
K.set_session(sess)



start_time = time.time()


#USD-BRL
dataframe = pandas.read_csv('minidolar/wdo.csv', sep = '|',  engine='python', decimal='.',header=0)
dataset_original = dataframe['fechamento']

dataset_original =  [100 * (b - a) / a for a, b in zip(dataset_original[::1], dataset_original[1::1])]


dataset_original = pandas.DataFrame(dataset_original)
dataset_original = dataset_original[0]



batch_size = 64
nb_epoch = 1000
patience = 1000
TARGET_TIME = 1
LAG_SIZE = 1
EMB_SIZE = 1




def evaluate_model(model, name, n_layers, ep, normalization, TRAIN_SIZE, dataset, ewm_dolar, type):
    #X_train, X_test, Y_train, Y_test = dataset
    #X_trainp, X_testp, Y_trainp, Y_testp = dadosp
    if (normalization == 'AN'):
        X_train, X_test, Y_train, Y_test, scaler, shift_train, shift_test, X_trainp, X_testp, Y_trainp, Y_testp = nn_an_type(dataset, ewm_dolar, TRAIN_SIZE,TARGET_TIME, LAG_SIZE, type)
    if (normalization == 'SW'):
        X_train, X_test, Y_train, Y_test, scaler_train, scaler_test, X_trainp, X_testp, Y_trainp, Y_testp = nn_sw(dataset,TRAIN_SIZE,TARGET_TIME, LAG_SIZE)
    if (normalization == 'MM'):
        X_train, X_test, Y_train, Y_test, scaler, X_trainp, X_testp, Y_trainp, Y_testp = nn_mm(dataset, TRAIN_SIZE, TARGET_TIME, LAG_SIZE)
    if (normalization == 'ZS'):
        X_train, X_test, Y_train, Y_test, scaler, X_trainp, X_testp, Y_trainp, Y_testp = nn_zs(dataset, TRAIN_SIZE, TARGET_TIME, LAG_SIZE)
    if (normalization == 'DS'):
        X_train, X_test, Y_train, Y_test, maximum, X_trainp, X_testp, Y_trainp, Y_testp = nn_ds(dataset, TRAIN_SIZE, TARGET_TIME, LAG_SIZE)



    csv_logger = CSVLogger('output/%d_layers/variacoes_%s_%s.csv' % (n_layers, name, normalization))
    reduce_lr = ReduceLROnPlateau(monitor='val_loss')
    es = EarlyStopping(monitor='val_loss', patience=patience)
    #mcp = ModelCheckpoint('output/mnist_adaptative_%dx800/%s.checkpoint' % (n_layers, name), save_weights_only=True)
    #tb = TensorBoard(log_dir='output/mnist_adaptative_%dx800' % n_layers, histogram_freq=1, write_graph=False, write_images=False)

    
    #sgd = SGD(lr=0.01, momentum=0.9, nesterov=True)
    #sgd = SGD(lr=0.64, momentum=0.8, nesterov=False)

    #optimizer = sgd
    #optimizer = Adam(lr=0.002)
    optimizer = 'adam'
    #optimizer = "adadelta"
    #optimizer = "nadam"

    model.compile(loss='mean_squared_error', optimizer=optimizer)
    
    history = model.fit(X_train, Y_train, batch_size=batch_size, epochs=ep, verbose=0, validation_split=0.1, callbacks=[csv_logger, es])

    # plt.plot(history.history['loss'])
    # plt.plot(history.history['val_loss'])
    # plt.title('model loss')
    # plt.ylabel('loss')
    # plt.xlabel('epoch')
    # plt.legend(['train', 'test'], loc='upper left')
    # plt.show()

    #trainScore = model.evaluate(X_train, Y_train, verbose=0)
    #print('Train Score: %f MSE (%f RMSE)' % (trainScore, math.sqrt(trainScore)))
    #testScore = model.evaluate(X_test, Y_test, verbose=0)
    #print('Test Score: %f MSE (%f RMSE)' % (testScore, math.sqrt(testScore)))

    # make predictions (scaled)
    trainPredict = model.predict(X_train)
    testPredict = model.predict(X_test)

    # invert predictions (back to original)
    if (normalization == 'AN'):
        X_trainp3, X_testp3, new_train_predicted, new_predicted = nn_an_den_type(X_train, X_test, trainPredict, testPredict, scaler, shift_train, shift_test, type)

    if (normalization == 'SW'):
        X_trainp3, X_testp3, new_train_predicted, new_predicted = nn_sw_den(X_train, X_test, trainPredict, testPredict, scaler_train, scaler_test)

    if (normalization == 'MM'):
        X_trainp3, X_testp3, new_train_predicted, new_predicted = nn_mm_den(X_train, X_test, trainPredict, testPredict, scaler)

    if (normalization == 'ZS'):
        X_trainp3, X_testp3, new_train_predicted, new_predicted = nn_zs_den(X_train, X_test, trainPredict, testPredict, scaler)

    if (normalization == 'DS'):
        X_trainp3, X_testp3, new_train_predicted, new_predicted = nn_ds_den(X_train, X_test, trainPredict, testPredict, maximum)

    # np.savetxt("output/previsto.csv", new_predicted)
    # np.savetxt("output/real.csv", Y_testp)
    # np.savetxt("output/previsto_treino.csv", new_train_predicted)
    # np.savetxt("output/real_treino.csv", Y_trainp)

    # np.savetxt("output/x_test-meu.csv", X_testp)
    # np.savetxt("output/y_test-meu.csv", Y_testp)
    # np.savetxt("output/x_treino-meu.csv", X_trainp)
    # np.savetxt("output/y_treino-meu.csv", Y_trainp)

    # calculate root mean squared error
    trainScore = math.sqrt(mean_squared_error(new_train_predicted, Y_trainp))
    #print('Train Score: %f RMSE' % (trainScore))
    testScore = math.sqrt(mean_squared_error(new_predicted, Y_testp))
    #print('Test Score: %f RMSE' % (testScore))
    epochs = len(history.epoch)

    # fig = plt.figure()
    # plt.plot(Y_test[:150], color='black') # BLUE - trained RESULT
    # plt.plot(testPredict[:150], color='blue') # RED - trained PREDICTION
    #plt.plot(Y_testp[:150], color='green') # GREEN - actual RESULT
    #plt.plot(new_predicted[:150], color='red') # ORANGE - restored PREDICTION
    #plt.show()

    return trainScore, testScore, epochs, optimizer

def __main__(argv):
    n_layers = int(argv[0])
    print(n_layers,'layers')

    #nonlinearities = ['aabh', 'abh', 'ah', 'sigmoid', 'relu', 'tanh']
    nonlinearities = ['sigmoid', 'relu', 'tanh']
    #nonlinearities = ['relu']

    # normalizations = ['AN', 'SW', 'MM', 'ZS', 'DS']
    #normalizations = ['DS']
    normalizations = ['ANo']
    # normalizations = ['ANo', 'ANc', 'ANd', 'SW', 'MM', 'ZS', 'DS']
    type = 'c'
    with open("output/%d_layers/compare.csv" % n_layers, "a") as fp:
        fp.write("-VARIACOES MINIDOLAR/MLP NN %s\n" % type)

    hals = []

    # best parameters without outlier removal: TRAIN_SIZE= 7 k=25
    # with outlier removal: TRAIN_SIZE=4 k=3

    for o in range(12, 30, 30):
        for p in seed:
            TRAIN_SIZE = o

            set_seeds(p)

            k = 3

            ewm_dolar = dataset_original.ewm(span=k, min_periods=k).mean()

            # removendo NaN
            dataset = np.array(dataset_original.iloc[k - 1:])
            ewm_dolar = np.array(ewm_dolar.iloc[k - 1:])

            #for name in nonlinearities:
            for normalization in normalizations:
                if (normalization[:-1] == 'AN'):
                    type = normalization[-1]
                    norm = 'AN'
                else:
                    norm = normalization
                name='tanh'
                model = Sequential()

                model.add(Dense(3, input_shape = (TRAIN_SIZE, ) , kernel_initializer='glorot_uniform', kernel_regularizer=regularizers.l2(0.01)))
                model.add(Activation(name))
                model.add(Dropout(0.25))

                for l in range(n_layers):
                    model.add(Dense(3, input_shape = (TRAIN_SIZE, ), kernel_initializer='glorot_uniform',
                                kernel_regularizer=regularizers.l2(0.01)))
                    model.add(Activation(name))
                    model.add(Dropout(0.25))
                   # model.add(Dropout(0.25))


                model.add(Dense(1))
                model.add(Activation(name))
                #model.summary()

                trainScore, testScore, epochs, optimizer = evaluate_model(model, name, n_layers,nb_epoch, norm, TRAIN_SIZE, dataset, ewm_dolar, type)
                # if(testScore_aux > testScore):
                #     testScore_aux=testScore
                #     f_aux = f

                elapsed_time = (time.time() - start_time)
                with open("output/%d_layers/compare.csv" % n_layers, "a") as fp:
                    #fp.write("%i,%s,%f,%f,%d,%s --%s seconds\n" % (f, name, trainScore, testScore, epochs, optimizer, elapsed_time))
                    fp.write("w=%i,k=%i,%s,%s,%f,%f,%d,%s --%s seconds\n" % (o,p, name, normalization, trainScore, testScore, epochs, optimizer, elapsed_time))
                model = None

if __name__ == "__main__":
   __main__(sys.argv[1:])