人工智慧5

前言

系列文章簡介

大家好，我們是 AI . FREE Team - 人工智慧自由團隊，這一次的鐵人賽，自由團隊將從0到1 手把手教各位讀者學會 (1)Python基礎語法 (2)Python Web 網頁開發框架 – Django (3)Python網頁爬蟲 – 周易解夢網 (4)Tensorflow AI語言模型基礎與訓練 – LSTM (5)實際部屬AI解夢模型到Web框架上。

為什麼技術要從零開始寫起

自由團隊的成立宗旨為開發AI/新科技的學習資源，提供各領域的學習者能夠跨域學習資料科學，並透過自主學習發展協槓職涯，結合智能應用到各式領域，無論是文、法、商、管、醫領域的朋友，都可以自由的學習AI技術。

資源

AI . FREE Team 讀者專屬福利 → Python Basics 免費學習資源

這次我們教學大家使用keras來實作上篇文章的題目

導入套件

%matplotlib inline
%config InlineBackend.figure_format = 'retina'
import matplotlib.pyplot as plt
import pandas as pd

import numpy as np
import seaborn as sns
import warnings
from datetime import datetime
from matplotlib.colors import ListedColormap
from sklearn.datasets import make_classification, make_moons, make_circles
from sklearn.metrics import confusion_matrix, classification_report, mean_squared_error, mean_absolute_error, r2_score
from sklearn.linear_model import LogisticRegression
from sklearn.utils import shuffle
from keras.models import Sequential
from keras.layers import Dense, Dropout, BatchNormalization, Activation
from keras.optimizers import Adam
from keras.utils.np_utils import to_categorical

import keras.backend as K
from keras.wrappers.scikit_learn import KerasClassifier

其實除了深度學習的model之外，其他套件都一樣，這意味著待會前處理與視覺化的部分都差不多

資料集(Dataset)

def plot_data(X, y, figsize=None):
    if not figsize:
        figsize = (8, 6)
    plt.figure(figsize=figsize)
    plt.plot(X[y==0, 0], X[y==0, 1], 'y*', alpha=0.5, label=0)
    plt.plot(X[y==1, 0], X[y==1, 1], 'g*', alpha=0.5, label=1)
    plt.xlim((min(X[:, 0])-0.1, max(X[:, 0])+0.1))
    plt.ylim((min(X[:, 1])-0.1, max(X[:, 1])+0.1))
    plt.legend()

X, y = make_classification(n_samples=1000, n_features=2, n_redundant=0, 
                           n_informative=2, random_state=50, n_clusters_per_class=1)
plot_data(X, y)

資料集與上次也是一樣的

找出線性切割線

lr = LogisticRegression()
lr.fit(X, y)
print('LR coefficients:', lr.coef_)
print('LR intercept:', lr.intercept_)

plot_data(X, y)

limits = np.array([-2, 2])
boundary = -(lr.coef_[0][0] * limits + lr.intercept_[0]) / lr.coef_[0][1]
plt.plot(limits, boundary, "r-", linewidth=2)
plt.show()

模型Model

model = Sequential()
model.add(Dense(units=1, input_shape=(2,), activation='sigmoid'))
model.summary()

輸出會長這樣

Model: "sequential_1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense_1 (Dense)              (None, 1)                 3         
=================================================================
Total params: 3
Trainable params: 3
Non-trainable params: 0
_________________________________________________________________

定義loss和optimizer

model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])

大家可以看到，這裡我們除了定義loss和optimizer，還有一個metrics，這個主要是在訓練過程中監看的指標

訓練Training

history =model.fit(x=X, y=y, verbose=1, epochs=25)

輸出

Epoch 1/30
32/32 [==============================] - 0s 572us/step - loss: 0.1842 - accuracy: 0.9630
Epoch 2/30
32/32 [==============================] - 0s 613us/step - loss: 0.1783 - accuracy: 0.9670
Epoch 3/30
32/32 [==============================] - 0s 574us/step - loss: 0.1728 - accuracy: 0.9670
Epoch 4/30
32/32 [==============================] - 0s 592us/step - loss: 0.1677 - accuracy: 0.9690
Epoch 5/30
32/32 [==============================] - 0s 588us/step - loss: 0.1629 - accuracy: 0.9700
Epoch 6/30
32/32 [==============================] - 0s 562us/step - loss: 0.1585 - accuracy: 0.9700
Epoch 7/30
32/32 [==============================] - 0s 697us/step - loss: 0.1544 - accuracy: 0.9720
Epoch 8/30
32/32 [==============================] - 0s 588us/step - loss: 0.1506 - accuracy: 0.9720
Epoch 9/30
32/32 [==============================] - 0s 564us/step - loss: 0.1470 - accuracy: 0.9720
Epoch 10/30
32/32 [==============================] - 0s 552us/step - loss: 0.1437 - accuracy: 0.9720
Epoch 11/30
32/32 [==============================] - 0s 657us/step - loss: 0.1406 - accuracy: 0.9730
Epoch 12/30
32/32 [==============================] - 0s 594us/step - loss: 0.1378 - accuracy: 0.9730
Epoch 13/30
32/32 [==============================] - 0s 583us/step - loss: 0.1351 - accuracy: 0.9730
Epoch 14/30
32/32 [==============================] - 0s 650us/step - loss: 0.1324 - accuracy: 0.9730
Epoch 15/30
32/32 [==============================] - 0s 586us/step - loss: 0.1300 - accuracy: 0.9730
Epoch 16/30
32/32 [==============================] - 0s 588us/step - loss: 0.1276 - accuracy: 0.9730
Epoch 17/30
32/32 [==============================] - 0s 657us/step - loss: 0.1254 - accuracy: 0.9730
Epoch 18/30
32/32 [==============================] - 0s 581us/step - loss: 0.1232 - accuracy: 0.9730
Epoch 19/30
32/32 [==============================] - 0s 588us/step - loss: 0.1212 - accuracy: 0.9730
Epoch 20/30
32/32 [==============================] - 0s 572us/step - loss: 0.1193 - accuracy: 0.9730
Epoch 21/30
32/32 [==============================] - 0s 598us/step - loss: 0.1175 - accuracy: 0.9730
Epoch 22/30
32/32 [==============================] - 0s 582us/step - loss: 0.1157 - accuracy: 0.9730
Epoch 23/30
32/32 [==============================] - 0s 569us/step - loss: 0.1141 - accuracy: 0.9730
Epoch 24/30
32/32 [==============================] - 0s 587us/step - loss: 0.1125 - accuracy: 0.9730
Epoch 25/30
32/32 [==============================] - 0s 581us/step - loss: 0.1110 - accuracy: 0.9730
Epoch 26/30
32/32 [==============================] - 0s 570us/step - loss: 0.1095 - accuracy: 0.9730
Epoch 27/30
32/32 [==============================] - 0s 633us/step - loss: 0.1081 - accuracy: 0.9730
Epoch 28/30
32/32 [==============================] - 0s 593us/step - loss: 0.1068 - accuracy: 0.9740
Epoch 29/30
32/32 [==============================] - 0s 595us/step - loss: 0.1055 - accuracy: 0.9740
Epoch 30/30
32/32 [==============================] - 0s 597us/step - loss: 0.1042 - accuracy: 0.9750

以上就是使用keras的訓練過程，從定義到訓練其實可以發現都比pytorch精簡很多，僅僅用了一行fit就把全部向前向後傳播做完了

再來看一下訓練完的history，一些功能性的API

pd.DataFrame(history.history)

這裡應該是會到30，因為我們的epoch設成了30次，可以看到loss與accuracy會儲存在這個叫history的方法裡面

視覺化看一下預測的如何吧~

def plot_loss_accuracy(history):
    historydf = pd.DataFrame(history.history, index=history.epoch)
    plt.figure(figsize=(8, 6))
    historydf.plot(ylim=(0, max(1, historydf.values.max())))
    loss = history.history['loss'][-1]
    acc = history.history['accuracy'][-1]
    plt.title('Loss: %.3f, Accuracy: %.3f' % (loss, acc))
    
plot_loss_accuracy(history)

訓練的結果還不錯，loss往下、accuracy往上正是我們所樂見的

def plot_decision_boundary(func, X, y, figsize=(9, 6)):
    amin, bmin = X.min(axis=0) - 0.1
    amax, bmax = X.max(axis=0) + 0.1
    hticks = np.linspace(amin, amax, 101)
    vticks = np.linspace(bmin, bmax, 101)
    
    aa, bb = np.meshgrid(hticks, vticks)
    ab = np.c_[aa.ravel(), bb.ravel()]
    c = func(ab)
    cc = c.reshape(aa.shape)

    cm = plt.cm.RdBu
    cm_bright = ListedColormap(['#deff17', '#5be044'])
    
    fig, ax = plt.subplots(figsize=figsize)
    contour = plt.contourf(aa, bb, cc, cmap=cm, alpha=0.8)
    
    ax_c = fig.colorbar(contour)
    ax_c.set_label("$P(y = 1)$")
    ax_c.set_ticks([0, 0.25, 0.5, 0.75, 1])
    
    plt.scatter(X[:, 0], X[:, 1], c=y, cmap=cm_bright)
    plt.xlim(amin, amax)
    plt.ylim(bmin, bmax)

plot_decision_boundary(lambda x: model.predict(x), X, y)

這些都是上一篇的東西，就不贅述啦~

y_pred = model.predict_classes(X, verbose=0)
print(classification_report(y, y_pred))

上面程式碼的輸出

              precision    recall  f1-score   support

           0       1.00      0.96      0.98       499
           1       0.96      1.00      0.98       501

    accuracy                           0.98      1000
   macro avg       0.98      0.98      0.98      1000
weighted avg       0.98      0.98      0.98      1000

月亮資料集

X, y = make_moons(n_samples=1000, noise=0.05, random_state=0)
plot_data(X, y)

和剛才一樣建立一個一層的網路

model = Sequential()
model.add(Dense(1, input_shape=(2,), activation='sigmoid'))

model.compile('adam', 'binary_crossentropy', metrics=['accuracy'])

history = model.fit(X, y, verbose=0, epochs=100)
plot_loss_accuracy(history)

可以看到，其實沒有到非常好，與之前pytorch的實作差不多，我們視覺化就知道了

y_pred = model.predict_classes(X, verbose=0)
print(classification_report(y, y_pred))

上面程式碼的輸出

              precision    recall  f1-score   support

           0       0.79      0.82      0.80       500
           1       0.81      0.78      0.80       500

    accuracy                           0.80      1000
   macro avg       0.80      0.80      0.80      1000
weighted avg       0.80      0.80      0.80      1000

加深模型

model = Sequential()
model.add(Dense(4, input_shape=(2,), activation='tanh'))
model.add(Dense(2, activation='tanh'))
model.add(Dense(1, activation='sigmoid'))

model.compile(Adam(lr=0.01), 'binary_crossentropy', metrics=['accuracy'])

history = model.fit(X, y, verbose=0, epochs=100)

plot_loss_accuracy(history)

視覺化

plot_decision_boundary(lambda x: model.predict(x), X, y)

y_pred = model.predict_classes(X, verbose=0)
print(classification_report(y, y_pred))

上面程式碼的輸出

              precision    recall  f1-score   support

           0       1.00      1.00      1.00       500
           1       1.00      1.00      1.00       500

    accuracy                           1.00      1000
   macro avg       1.00      1.00      1.00      1000
weighted avg       1.00      1.00      1.00      1000

與pytorch的模型一樣，相同的模型也可以達到100%

比較

可以發現keras相對於pytorch來說相對簡單，但是方便的東西，伴隨的是很多東西要使用他的API，筆者認為這裡有利有弊，有些功能可以套件沒有提供，但是你又不能使用非這個套件的方法時，就會相當頭痛，pytorch是使用自由，但是筆者認為相對困難對於新手ˋˊ，筆者也是摸了一陣子才有辦法做出來

以上是今天的文章，希望大家會喜歡:))

想更深入認識 AI . FREE Team ？

自由團隊 官方網站：https://aifreeblog.herokuapp.com/
自由團隊 Github：https://github.com/AI-FREE-Team/
自由團隊 粉絲專頁：https://www.facebook.com/AI.Free.Team/
自由團隊 IG：https://www.instagram.com/aifreeteam/
自由團隊 Youtube：https://www.youtube.com/channel/UCjw6Kuw3kwM_il39NTBJVTg/

文章同步發布於：自由團隊部落格
(想看更多文章？學習更多AI知識？敬請鎖定自由團隊的頻道！)