Hfrxdjt

I have different results in a function approximation problem. I am trying to approximate a sine wave using MLPRegressor and Keras (um dense layer)
Here is the code for the MLPRegressor:

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from sklearn.neural_network import MLPRegressor
from sklearn.metrics import mean_squared_error

#Cria um dataset
X_train = np.arange(0.0, 1, 0.01).reshape(-1, 1)
noise= np.random.normal(0,0.1,100).reshape(-1,1)



y_train = np.sin(2*np.pi*X_train) 
y_train=y_train + noise
y_train=y_train.ravel() # transfoprma em 1D array

#X_train = np.arange(0.0, 1, 0.01).reshape(-1, 1)
#y_train = np.sin(2 * np.pi * X_train).ravel()


# Experimentos 
#hidden_layer sizes : 1,3, 100
#max_iter=10,100,1000
#
nn = MLPRegressor(
 hidden_layer_sizes=(3,), activation='tanh', solver='lbfgs', alpha=0.000, batch_size='auto',
 learning_rate='constant', learning_rate_init=0.01, power_t=0.5, max_iter=80, shuffle=True,
 random_state=0, tol=0.0001, verbose=True, warm_start=False, momentum=0.0, nesterovs_momentum=False,
 early_stopping=False, validation_fraction=0.0, beta_1=0.9, beta_2=0.999, epsilon=1e-08)

#Treina a Rede
n = nn.fit(X_train, y_train)

#previsoes na rede no conjunto de treinamento
predict_train =nn.predict(X_train)

#Plota o treinamento
fig = plt.figure()
ax1 = fig.add_subplot(111)
ax1.scatter(X_train, y_train, s=5, c='b', marker="o", label='real')
ax1.plot(X_train,predict_train, c='r', label='NN Prediction')


#Conjunto de Teste
X_test = np.arange(0.0, 1, 0.01).reshape(-1, 1)
y_test = np.sin(2*np.pi*X_test) + np.random.normal(0,0.2,100).reshape(-1,1)
y_test=y_test.ravel()


#Calcula as previsoes no conjunto de teste

predict_test= nn.predict(X_test)

fig = plt.figure()
ax1 = fig.add_subplot(111)
ax1.scatter(X_test, y_test, s=5, c='b', marker="o", label='real')
ax1.plot(X_test,predict_test, c='r', label='NN Prediction')

plt.legend()
plt.show()

print('MSE training : :.3f'.format(mean_squared_error(y_train, predict_train)))
print('MSE testing : :.3f'.format(mean_squared_error(y_test, predict_test)))

Using MLPRegressor, I found satisfactory results with just 3 neurons. However, when I try to use Keras, I can not get reasonably results. The code is very similar with the exception of the optmizer and the activation function. Here is the code for Keras:

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt

import keras
from keras.models import Sequential
from keras.layers import Dense
from sklearn.metrics import mean_squared_error





#
#Cria um dataset

#Cria um dataset
X_train = np.arange(0.0, 1, 0.01).reshape(-1, 1)
noise= np.random.normal(0,0.1,100).reshape(-1,1)



y_train = np.sin(2*np.pi*X_train) 
y_train=y_train + noise
y_train=y_train.ravel() # transfoprma em 1D array


#Construir a Rede
nn = Sequential() # sequencia de camada
#activation
# sigmoid, tanh, relu, linear
# units: numero de neuronios na camada
#primeira camada escondida tem input_dim
nn.add(Dense(units = 100, activation = 'relu',
 kernel_initializer = 'random_uniform', input_dim = 1))
nn.add(Dense(units = 1, activation = 'linear'))

# Algorritmo de aprendizado
#sgd = keras.optimizers.SGD(lr=0.1, decay=0, momentum=0, nesterov=False)
adam=keras.optimizers.Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0, amsgrad=False)

#determina a funcao de custo e a metrica utilizada
nn.compile(loss = 'mean_squared_error', optimizer = adam,
 metrics = ['mean_squared_error'])
history= nn.fit(X_train, y_train, batch_size = 1, epochs = 1000)

#previsoes na rede no conjunto de treinamento
predict_train =nn.predict(X_train)



#Plota o treinamento
fig = plt.figure()
ax1 = fig.add_subplot(111)
ax1.scatter(X_train, y_train, s=5, c='b', marker="o", label='real')
ax1.plot(X_train,predict_train, c='r', label='NN Prediction')


#Conjunto de Teste
X_test = np.arange(0.0, 1, 0.01).reshape(-1, 1)
y_test = np.sin(2*np.pi*X_test) + np.random.normal(0,0.2,100).reshape(-1,1)
y_test=y_test.ravel()


#Calcula as previsoes no conjunto de teste

predict_test= nn.predict(X_test)

fig = plt.figure()
ax1 = fig.add_subplot(111)
ax1.scatter(X_test, y_test, s=5, c='b', marker="o", label='real')
ax1.plot(X_test,predict_test, c='r', label='NN Prediction')

plt.legend()
plt.show()

print('MSE training : :.3f'.format(mean_squared_error(y_train, predict_train)))
print('MSE testing : :.3f'.format(mean_squared_error(y_test, predict_test)))

I already tried sgd as optimizer and also tanh for activation function. I do not undestand what I am missing, that is why I cann make the code for function approximation using Keras work.