TF2-HyperparameterSearch-2

import tensorflow as tf
import matplotlib as mlp
import matplotlib.pyplot as plt
%matplotlib inline
import numpy as np
import pandas as pd
import sklearn
import os
import sys
from tensorflow import keras
print("tf.__version__ = ",tf.__version__)
print(keras.__version__)

tf.__version__ =  2.0.0-alpha0
2.2.4-tf

#Get data
from sklearn.datasets import fetch_california_housing
housing = fetch_california_housing()
data = housing.data
target = housing.target

#split train test valid data
from sklearn.model_selection import train_test_split
x_train_all , x_test , y_train_all , y_test = train_test_split(
    data,target,random_state = 7 , test_size = 0.25
)
print(x_train_all.shape)
x_train , x_valid , y_train , y_valid = train_test_split(
    x_train_all,y_train_all,random_state = 11 , test_size = 0.25
)
print(x_train.shape)

(15480, 8)
(11610, 8)

#Normarlization
from sklearn.preprocessing import StandardScaler
scaler = StandardScaler()
x_train_scaled = scaler.fit_transform(x_train)
x_valid_scaled = scaler.transform(x_valid)
x_test_scaled = scaler.transform(x_test)

#build model
# 1:把tf.keras的model轉化爲sklearn的model
# 2:定義參數集合
# 3:搜索參數


def build_model(hidden_layers = 1,layer_size = 30,learning_rate=3e-3):
    model = keras.models.Sequential()
    #imput
    model.add(keras.layers.Dense(layer_size,activation="relu",input_shape = x_train_scaled.shape[1:]))
    #hidden
    for _ in range(hidden_layers):
        model.add(keras.layers.Dense(layer_size,activation="relu"))
    #out
    model.add(keras.layers.Dense(1))
    
    #compile
    optimizer = keras.optimizers.SGD(learning_rate)
    model.compile(loss='mse',optimizer = optimizer)
    return model
#建立keras模型病轉化爲sklearn模型
sklearn_model = keras.wrappers.scikit_learn.KerasRegressor(build_model)

#train
#
callbacks = [keras.callbacks.EarlyStopping(patience=5,min_delta=1e-2)]
history = sklearn_model.fit(x_train_scaled,
                  y_train,epochs = 100 , 
                  validation_data=(x_valid_scaled,y_valid),
                  callbacks = callbacks
                 )

Train on 11610 samples, validate on 3870 samples
Epoch 1/100
11610/11610 [==============================] - 2s 136us/sample - loss: 1.1758 - val_loss: 0.8212
Epoch 2/100
11610/11610 [==============================] - 1s 67us/sample - loss: 0.6568 - val_loss: 0.6284
Epoch 3/100
11610/11610 [==============================] - 1s 66us/sample - loss: 0.5523 - val_loss: 0.5740
Epoch 4/100
11610/11610 [==============================] - 1s 69us/sample - loss: 0.5134 - val_loss: 0.5395
Epoch 5/100
11610/11610 [==============================] - 1s 68us/sample - loss: 0.4880 - val_loss: 0.5137
Epoch 6/100
11610/11610 [==============================] - 1s 68us/sample - loss: 0.4782 - val_loss: 0.4952
Epoch 7/100
11610/11610 [==============================] - 1s 67us/sample - loss: 0.4615 - val_loss: 0.4847
Epoch 8/100
11610/11610 [==============================] - 1s 67us/sample - loss: 0.4455 - val_loss: 0.4709
Epoch 9/100
11610/11610 [==============================] - 1s 67us/sample - loss: 0.4373 - val_loss: 0.4597
Epoch 10/100
11610/11610 [==============================] - 1s 70us/sample - loss: 0.4289 - val_loss: 0.4528
Epoch 11/100
11610/11610 [==============================] - 1s 67us/sample - loss: 0.4242 - val_loss: 0.4480
Epoch 12/100
11610/11610 [==============================] - 1s 67us/sample - loss: 0.4149 - val_loss: 0.4349
Epoch 13/100
11610/11610 [==============================] - 1s 68us/sample - loss: 0.4108 - val_loss: 0.4287
Epoch 14/100
11610/11610 [==============================] - 1s 67us/sample - loss: 0.4115 - val_loss: 0.4255
Epoch 15/100
11610/11610 [==============================] - 1s 67us/sample - loss: 0.3967 - val_loss: 0.4531
Epoch 16/100
11610/11610 [==============================] - 1s 67us/sample - loss: 0.3913 - val_loss: 0.4107
Epoch 17/100
11610/11610 [==============================] - 1s 67us/sample - loss: 0.3895 - val_loss: 0.4058
Epoch 18/100
11610/11610 [==============================] - 1s 69us/sample - loss: 0.3813 - val_loss: 0.4062
Epoch 19/100
11610/11610 [==============================] - 1s 67us/sample - loss: 0.3838 - val_loss: 0.4045
Epoch 20/100
11610/11610 [==============================] - 1s 68us/sample - loss: 0.3767 - val_loss: 0.3936
Epoch 21/100
11610/11610 [==============================] - 1s 67us/sample - loss: 0.3761 - val_loss: 0.3912
Epoch 22/100
11610/11610 [==============================] - 1s 70us/sample - loss: 0.3687 - val_loss: 0.3986
Epoch 23/100
11610/11610 [==============================] - 1s 70us/sample - loss: 0.3662 - val_loss: 0.3949
Epoch 24/100
11610/11610 [==============================] - 1s 67us/sample - loss: 0.3686 - val_loss: 0.3851
Epoch 25/100
11610/11610 [==============================] - 1s 69us/sample - loss: 0.3618 - val_loss: 0.3795
Epoch 26/100
11610/11610 [==============================] - 1s 69us/sample - loss: 0.3632 - val_loss: 0.3804
Epoch 27/100
11610/11610 [==============================] - 1s 67us/sample - loss: 0.3564 - val_loss: 0.3807
Epoch 28/100
11610/11610 [==============================] - 1s 66us/sample - loss: 0.3600 - val_loss: 0.3772
Epoch 29/100
11610/11610 [==============================] - 1s 69us/sample - loss: 0.3599 - val_loss: 0.3742
Epoch 30/100
11610/11610 [==============================] - 1s 68us/sample - loss: 0.3545 - val_loss: 0.3731

def plot_learning_curves(history):
    pd.DataFrame(history.history).plot(figsize=(8,5))
    plt.grid(True)
    plt.gca().set_ylim(0,1)
    plt.show()
plot_learning_curves(history)