Go to Qiita Advent Calendar 2024 Top
LoginSignup

Delete article

Deleted articles cannot be recovered.

Draft of this article would be also deleted.

Are you sure you want to delete this article?

This article is a Private article. Only a writer and users who know the URL can access it.
Please change open range to public in publish setting if you want to share this article with other users.

More than 3 years have passed since last update.

@fantahide

【ラビットチャレンジ】深層学習(Day2:CNN)

Last updated at Posted at 2020-05-16

E資格取得のためラビットチャレンジに挑戦しています。

#復習

#####確認テスト
連鎖律の原理を使い、dz/dxを求めよ

z = t^2
\\
t = x + y
\\
\frac{dz}{dt}=2t
\\
\frac{dt}{dx}=1
\\
\frac{dz}{dx}= \frac{dz}{dt}\frac{dt}{dx}
\\
\frac{dz}{dx}=2t=2(x + y)

#####確認テスト
シグモイド関数を微分した時、入力値が0の時に最大値をとる。その値として正しいものを選択肢から選べ
(1) 0.15
(2) 0.25
(3) 0.35
(4) 0.45

$ (x)(1- sigmoid(x))・sigmoid(x)$
0.5の時最大
(1-0.5)・(0.5)=0.25

よって(2)

#Section1 勾配消失問題

###ポイント

  • 誤差逆伝播法が下位層(出力->入力)に進んでいくにつれて勾配が緩やかになっていく。そのため、勾配降下法による、更新では下位層のパラメータはほとんど変わらず、訓練は最適値に収束しなくなる。(パラメータ更新が掛けれない)
  • 勾配消失の解決方法
    • 活性化関数の選択
    • 重みの初期値設定
      • Xavier
        • ReLU関数、シグモイド関数、双曲線正接関数と組み合わせて使う
        • 重みの要素を前の層のノード数の平方根で除算した値(初期値の設定方法)
      • He
        • ReLU関数と組み合わせて使う
        • 重みの要素を、前の層のノード数の平方根で除算した値に対し$\sqrt{2}$を掛け合わせた値(初期値の設定方法)　
    • バッチ正規化
      • ミニバッチ単位で、入力値のデータの偏りを抑制する方法
      • 活性化関数に値を渡す前後に、バッチ正規化の処理を孕んだ層を加える
      • バッチ正規化層への入力値は
        • $u^{l}=w^{(l)}z^{(l-1)}+b^{(l)}$ または $z$

image.png

#####確認テスト

  • 重みの初期値に0を設定すると、どのような問題が発生するか、簡潔に説明せよ

=>全ての値が同じ値で伝わるためパラメータのチューニングが行われなくなる

#####確認テスト

  • 一般的に考えられるバッチ正規化の効果を2点挙げよ

=>精度が上がる。収束時間が短縮される。
=>計算の高速化、勾配消失が起きづらくなる。

###ハンズオン

ソースコード名
2_2_1_vanishing_gradient

vanishing gradient

  • sigmoid - gauss
Jupyter


import sys, os
sys.path.append(os.pardir)  # 親ディレクトリのファイルをインポートするための設定
import numpy as np
from common import layers
from collections import OrderedDict
from common import functions
from data.mnist import load_mnist
import matplotlib.pyplot as plt

# mnistをロード
(x_train, d_train), (x_test, d_test) = load_mnist(normalize=True, one_hot_label=True)
train_size = len(x_train)

print("データ読み込み完了")

# 重み初期値補正係数
wieght_init = 0.01
#入力層サイズ
input_layer_size = 784
#中間層サイズ
hidden_layer_1_size = 40
hidden_layer_2_size = 20

#出力層サイズ
output_layer_size = 10
# 繰り返し数
iters_num = 2000
# ミニバッチサイズ
batch_size = 100
# 学習率
learning_rate = 0.1
# 描写頻度
plot_interval=10

# 初期設定
def init_network():
    network = {} 
    network['W1'] = wieght_init * np.random.randn(input_layer_size, hidden_layer_1_size)
    network['W2'] = wieght_init * np.random.randn(hidden_layer_1_size, hidden_layer_2_size)
    network['W3'] = wieght_init * np.random.randn(hidden_layer_2_size, output_layer_size)

    network['b1'] = np.zeros(hidden_layer_1_size)
    network['b2'] = np.zeros(hidden_layer_2_size)
    network['b3'] = np.zeros(output_layer_size)

    return network

# 順伝播
def forward(network, x):
    W1, W2, W3 = network['W1'], network['W2'], network['W3']
    b1, b2, b3 = network['b1'], network['b2'], network['b3']
    hidden_f = functions.sigmoid
    
    u1 =  np.dot(x, W1) + b1
    z1 = hidden_f(u1)
    u2 =  np.dot(z1, W2) + b2
    z2 = hidden_f(u2)
    u3 =  np.dot(z2, W3) + b3
    y = functions.softmax(u3)
 
    return z1, z2, y

# 誤差逆伝播
def backward(x, d, z1, z2, y):
    grad = {}
    
    W1, W2, W3 = network['W1'], network['W2'], network['W3']
    b1, b2, b3 = network['b1'], network['b2'], network['b3']
    hidden_d_f = functions.d_sigmoid
    last_d_f = functions.d_softmax_with_loss
    
    
    # 出力層でのデルタ
    delta3 = last_d_f(d, y)
    # b3の勾配
    grad['b3'] = np.sum(delta3, axis=0)
    # W3の勾配
    grad['W3'] = np.dot(z2.T, delta3)
    # 2層でのデルタ
    delta2 = np.dot(delta3, W3.T) * hidden_d_f(z2)
    # b2の勾配
    grad['b2'] = np.sum(delta2, axis=0)
    # W2の勾配
    grad['W2'] = np.dot(z1.T, delta2)
    # 1層でのデルタ
    delta1 = np.dot(delta2, W2.T) * hidden_d_f(z1)
    # b1の勾配
    grad['b1'] = np.sum(delta1, axis=0)
    # W1の勾配
    grad['W1'] = np.dot(x.T, delta1)

    return grad

# パラメータの初期化
network = init_network()

accuracies_train = []
accuracies_test = []

# 正答率
def accuracy(x, d):
    z1, z2, y = forward(network, x)
    y = np.argmax(y, axis=1)
    if d.ndim != 1 : d = np.argmax(d, axis=1)
    accuracy = np.sum(y == d) / float(x.shape[0])
    return accuracy

for i in range(iters_num):
    # ランダムにバッチを取得    
    batch_mask = np.random.choice(train_size, batch_size)
    # ミニバッチに対応する教師訓練画像データを取得    
    x_batch = x_train[batch_mask]
    # ミニバッチに対応する訓練正解ラベルデータを取得する
    d_batch = d_train[batch_mask]


    
    z1, z2, y = forward(network, x_batch)
    grad = backward(x_batch, d_batch, z1, z2, y)

    if (i+1)%plot_interval==0:
        accr_test = accuracy(x_test, d_test)
        accuracies_test.append(accr_test)
        
        accr_train = accuracy(x_batch, d_batch)
        accuracies_train.append(accr_train)

        print('Generation: ' + str(i+1) + '. 正答率(トレーニング) = ' + str(accr_train))
        print('                : ' + str(i+1) + '. 正答率(テスト) = ' + str(accr_test))

    # パラメータに勾配適用
    for key in ('W1', 'W2', 'W3', 'b1', 'b2', 'b3'):
        network[key]  -= learning_rate * grad[key]


lists = range(0, iters_num, plot_interval)
plt.plot(lists, accuracies_train, label="training set")
plt.plot(lists, accuracies_test,  label="test set")
plt.legend(loc="lower right")
plt.title("accuracy")
plt.xlabel("count")
plt.ylabel("accuracy")
plt.ylim(0, 1.0)
# グラフの表示
plt.show()

勾配消失が起きている
image.png

  • ReLU - gauss
Jupyter

import sys, os
sys.path.append(os.pardir)  # 親ディレクトリのファイルをインポートするための設定
import numpy as np
from data.mnist import load_mnist
from PIL import Image
import pickle
from common import functions
import matplotlib.pyplot as plt

# mnistをロード
(x_train, d_train), (x_test, d_test) = load_mnist(normalize=True, one_hot_label=True)
train_size = len(x_train)

print("データ読み込み完了")

# 重み初期値補正係数
wieght_init = 0.01
#入力層サイズ
input_layer_size = 784
#中間層サイズ
hidden_layer_1_size = 40
hidden_layer_2_size = 20

#出力層サイズ
output_layer_size = 10
# 繰り返し数
iters_num = 2000
# ミニバッチサイズ
batch_size = 100
# 学習率
learning_rate = 0.1
# 描写頻度
plot_interval=10

# 初期設定
def init_network():
    network = {} 

    network['W1'] = wieght_init * np.random.randn(input_layer_size, hidden_layer_1_size)
    network['W2'] = wieght_init * np.random.randn(hidden_layer_1_size, hidden_layer_2_size)
    network['W3'] = wieght_init * np.random.randn(hidden_layer_2_size, output_layer_size)
            
    network['b1'] = np.zeros(hidden_layer_1_size)
    network['b2'] = np.zeros(hidden_layer_2_size)
    network['b3'] = np.zeros(output_layer_size)

    return network

# 順伝播
def forward(network, x):
    W1, W2, W3 = network['W1'], network['W2'], network['W3']
    b1, b2, b3 = network['b1'], network['b2'], network['b3']
    
    ###########  変更箇所  ##############

    hidden_f = functions.relu

    #################################
    
    u1 =  np.dot(x, W1) + b1
    z1 = hidden_f(u1)
    u2 =  np.dot(z1, W2) + b2
    z2 = hidden_f(u2)
    u3 =  np.dot(z2, W3) + b3
    y = functions.softmax(u3)
 
    return z1, z2, y

# 誤差逆伝播
def backward(x, d, z1, z2, y):
    grad = {}
    
    W1, W2, W3 = network['W1'], network['W2'], network['W3']
    b1, b2, b3 = network['b1'], network['b2'], network['b3']

    ###########  変更箇所  ##############
    
    hidden_d_f = functions.d_relu
    
    #################################
    
    
    # 出力層でのデルタ
    delta3 = functions.d_softmax_with_loss(d, y)
    # b3の勾配
    grad['b3'] = np.sum(delta3, axis=0)
    # W3の勾配
    grad['W3'] = np.dot(z2.T, delta3)
    # 2層でのデルタ
    delta2 = np.dot(delta3, W3.T) * hidden_d_f(z2)
    # b2の勾配
    grad['b2'] = np.sum(delta2, axis=0)
    # W2の勾配
    grad['W2'] = np.dot(z1.T, delta2)
    # 1層でのデルタ
    delta1 = np.dot(delta2, W2.T) * hidden_d_f(z1)
    # b1の勾配
    grad['b1'] = np.sum(delta1, axis=0)
    # W1の勾配
    grad['W1'] = np.dot(x.T, delta1)

    return grad

# パラメータの初期化
network = init_network()

accuracies_train = []
accuracies_test = []

# 正答率
def accuracy(x, d):
    z1, z2, y = forward(network, x)
    y = np.argmax(y, axis=1)
    if d.ndim != 1 : d = np.argmax(d, axis=1)
    accuracy = np.sum(y == d) / float(x.shape[0])
    return accuracy

for i in range(iters_num):
    # ランダムにバッチを取得    
    batch_mask = np.random.choice(train_size, batch_size)
    # ミニバッチに対応する教師訓練画像データを取得    
    x_batch = x_train[batch_mask]
    # ミニバッチに対応する訓練正解ラベルデータを取得する
    d_batch = d_train[batch_mask]


    
    z1, z2, y = forward(network, x_batch)
    grad = backward(x_batch, d_batch, z1, z2, y)

    if (i+1)%plot_interval==0:
        accr_test = accuracy(x_test, d_test)
        accuracies_test.append(accr_test)
        
        accr_train = accuracy(x_batch, d_batch)
        accuracies_train.append(accr_train)

        print('Generation: ' + str(i+1) + '. 正答率(トレーニング) = ' + str(accr_train))
        print('                : ' + str(i+1) + '. 正答率(テスト) = ' + str(accr_test))
        
    # パラメータに勾配適用
    for key in ('W1', 'W2', 'W3', 'b1', 'b2', 'b3'):
        network[key]  -= learning_rate * grad[key]


lists = range(0, iters_num, plot_interval)
plt.plot(lists, accuracies_train, label="training set")
plt.plot(lists, accuracies_test,  label="test set")
plt.legend(loc="lower right")
plt.title("accuracy")
plt.xlabel("count")
plt.ylabel("accuracy")
plt.ylim(0, 1.0)
# グラフの表示
plt.show()

ReLU関数x>0の時 y=xになるので勾配消失は起きていない

image.png

  • sigmoid - Xavier
Jupyter
import sys, os
sys.path.append(os.pardir)  # 親ディレクトリのファイルをインポートするための設定
import numpy as np
from data.mnist import load_mnist
from PIL import Image
import pickle
from common import functions
import matplotlib.pyplot as plt

# mnistをロード
(x_train, d_train), (x_test, d_test) = load_mnist(normalize=True, one_hot_label=True)
train_size = len(x_train)

print("データ読み込み完了")

#入力層サイズ
input_layer_size = 784
#中間層サイズ
hidden_layer_1_size = 40
hidden_layer_2_size = 20
#出力層サイズ
output_layer_size = 10
# 繰り返し数
iters_num = 2000
# ミニバッチサイズ
batch_size = 100
# 学習率
learning_rate = 0.1
# 描写頻度
plot_interval=10

# 初期設定
def init_network():
    network = {} 
    
    ###########  変更箇所  ##############
    
    # Xavierの初期値
    network['W1'] = np.random.randn(input_layer_size, hidden_layer_1_size) / (np.sqrt(input_layer_size))
    network['W2'] = np.random.randn(hidden_layer_1_size, hidden_layer_2_size) / (np.sqrt(hidden_layer_1_size))
    network['W3'] = np.random.randn(hidden_layer_2_size, output_layer_size) / (np.sqrt(hidden_layer_2_size))
    
    #################################
    
    network['b1'] = np.zeros(hidden_layer_1_size)
    network['b2'] = np.zeros(hidden_layer_2_size)
    network['b3'] = np.zeros(output_layer_size)

    return network

# 順伝播
def forward(network, x):
    W1, W2, W3 = network['W1'], network['W2'], network['W3']
    b1, b2, b3 = network['b1'], network['b2'], network['b3']
    hidden_f = functions.sigmoid
    
    u1 =  np.dot(x, W1) + b1
    z1 = hidden_f(u1)
    u2 =  np.dot(z1, W2) + b2
    z2 = hidden_f(u2)
    u3 =  np.dot(z2, W3) + b3
    y = functions.softmax(u3)
 
    return z1, z2, y

# 誤差逆伝播
def backward(x, d, z1, z2, y):
    grad = {}
    
    W1, W2, W3 = network['W1'], network['W2'], network['W3']
    b1, b2, b3 = network['b1'], network['b2'], network['b3']
    hidden_d_f = functions.d_sigmoid
    
    # 出力層でのデルタ
    delta3 = functions.d_softmax_with_loss(d, y)
    # b3の勾配
    grad['b3'] = np.sum(delta3, axis=0)
    # W3の勾配
    grad['W3'] = np.dot(z2.T, delta3)
    # 2層でのデルタ
    delta2 = np.dot(delta3, W3.T) * hidden_d_f(z2)
    # b2の勾配
    grad['b2'] = np.sum(delta2, axis=0)
    # W2の勾配
    grad['W2'] = np.dot(z1.T, delta2)
    # 1層でのデルタ
    delta1 = np.dot(delta2, W2.T) * hidden_d_f(z1)
    # b1の勾配
    grad['b1'] = np.sum(delta1, axis=0)
    # W1の勾配
    grad['W1'] = np.dot(x.T, delta1)

    return grad

# パラメータの初期化
network = init_network()

accuracies_train = []
accuracies_test = []

# 正答率
def accuracy(x, d):
    z1, z2, y = forward(network, x)
    y = np.argmax(y, axis=1)
    if d.ndim != 1 : d = np.argmax(d, axis=1)
    accuracy = np.sum(y == d) / float(x.shape[0])
    return accuracy

for i in range(iters_num):
    # ランダムにバッチを取得    
    batch_mask = np.random.choice(train_size, batch_size)
    # ミニバッチに対応する教師訓練画像データを取得    
    x_batch = x_train[batch_mask]
    # ミニバッチに対応する訓練正解ラベルデータを取得する
    d_batch = d_train[batch_mask]


    
    z1, z2, y = forward(network, x_batch)
    grad = backward(x_batch, d_batch, z1, z2, y)

    if (i+1)%plot_interval==0:
        accr_test = accuracy(x_test, d_test)
        accuracies_test.append(accr_test)
        
        accr_train = accuracy(x_batch, d_batch)
        accuracies_train.append(accr_train)

        print('Generation: ' + str(i+1) + '. 正答率(トレーニング) = ' + str(accr_train))
        print('                : ' + str(i+1) + '. 正答率(テスト) = ' + str(accr_test))
        
    # パラメータに勾配適用
    for key in ('W1', 'W2', 'W3', 'b1', 'b2', 'b3'):
        network[key]  -= learning_rate * grad[key]


lists = range(0, iters_num, plot_interval)
plt.plot(lists, accuracies_train, label="training set")
plt.plot(lists, accuracies_test,  label="test set")
plt.legend(loc="lower right")
plt.title("accuracy")
plt.xlabel("count")
plt.ylabel("accuracy")
plt.ylim(0, 1.0)
# グラフの表示
plt.show()

初期値の設定方法をXavierに変えるだけで勾配消失は起きなくなった。

image.png

  • ReLU - He
Jupyter
import sys, os
sys.path.append(os.pardir)  # 親ディレクトリのファイルをインポートするための設定
import numpy as np
from data.mnist import load_mnist
from PIL import Image
import pickle
from common import functions
import matplotlib.pyplot as plt

# mnistをロード
(x_train, d_train), (x_test, d_test) = load_mnist(normalize=True, one_hot_label=True)
train_size = len(x_train)

print("データ読み込み完了")

# 重み初期値補正係数
wieght_init = 0.01
#入力層サイズ
input_layer_size = 784
#中間層サイズ
hidden_layer_1_size = 40
hidden_layer_2_size = 20

#出力層サイズ
output_layer_size = 10
# 繰り返し数
iters_num = 2000
# ミニバッチサイズ
batch_size = 100
# 学習率
learning_rate = 0.1
# 描写頻度
plot_interval=10

# 初期設定
def init_network():
    network = {} 

    ###########  変更箇所  ##############

    # Heの初期値
    network['W1'] = np.random.randn(input_layer_size, hidden_layer_1_size) / np.sqrt(input_layer_size) * np.sqrt(2)
    network['W2'] = np.random.randn(hidden_layer_1_size, hidden_layer_2_size) / np.sqrt(hidden_layer_1_size) * np.sqrt(2)
    network['W3'] = np.random.randn(hidden_layer_2_size, output_layer_size) / np.sqrt(hidden_layer_2_size) * np.sqrt(2)
        
    #################################
    
    network['b1'] = np.zeros(hidden_layer_1_size)
    network['b2'] = np.zeros(hidden_layer_2_size)
    network['b3'] = np.zeros(output_layer_size)

    return network

# 順伝播
def forward(network, x):
    W1, W2, W3 = network['W1'], network['W2'], network['W3']
    b1, b2, b3 = network['b1'], network['b2'], network['b3']

    ###########  変更箇所  ##############
    
    hidden_f = functions.relu
    
    #################################
    
    u1 =  np.dot(x, W1) + b1
    z1 = hidden_f(u1)
    u2 =  np.dot(z1, W2) + b2
    z2 = hidden_f(u2)
    u3 =  np.dot(z2, W3) + b3
    y = functions.softmax(u3)
 
    return z1, z2, y

# 誤差逆伝播
def backward(x, d, z1, z2, y):
    grad = {}
    
    W1, W2, W3 = network['W1'], network['W2'], network['W3']
    b1, b2, b3 = network['b1'], network['b2'], network['b3']
    
    ###########  変更箇所  ##############
    
    hidden_d_f = functions.d_relu
    
    #################################
    
    # 出力層でのデルタ
    delta3 = functions.d_softmax_with_loss(d, y)
    # b3の勾配
    grad['b3'] = np.sum(delta3, axis=0)
    # W3の勾配
    grad['W3'] = np.dot(z2.T, delta3)
    # 2層でのデルタ
    delta2 = np.dot(delta3, W3.T) * hidden_d_f(z2)
    # b2の勾配
    grad['b2'] = np.sum(delta2, axis=0)
    # W2の勾配
    grad['W2'] = np.dot(z1.T, delta2)
    # 1層でのデルタ
    delta1 = np.dot(delta2, W2.T) * hidden_d_f(z1)
    # b1の勾配
    grad['b1'] = np.sum(delta1, axis=0)
    # W1の勾配
    grad['W1'] = np.dot(x.T, delta1)

    return grad

# パラメータの初期化
network = init_network()

accuracies_train = []
accuracies_test = []

# 正答率
def accuracy(x, d):
    z1, z2, y = forward(network, x)
    y = np.argmax(y, axis=1)
    if d.ndim != 1 : d = np.argmax(d, axis=1)
    accuracy = np.sum(y == d) / float(x.shape[0])
    return accuracy

for i in range(iters_num):
    # ランダムにバッチを取得    
    batch_mask = np.random.choice(train_size, batch_size)
    # ミニバッチに対応する教師訓練画像データを取得    
    x_batch = x_train[batch_mask]
    # ミニバッチに対応する訓練正解ラベルデータを取得する
    d_batch = d_train[batch_mask]


    
    z1, z2, y = forward(network, x_batch)
    grad = backward(x_batch, d_batch, z1, z2, y)

    if (i+1)%plot_interval==0:
        accr_test = accuracy(x_test, d_test)
        accuracies_test.append(accr_test)
        
        accr_train = accuracy(x_batch, d_batch)
        accuracies_train.append(accr_train)

        print('Generation: ' + str(i+1) + '. 正答率(トレーニング) = ' + str(accr_train))
        print('                : ' + str(i+1) + '. 正答率(テスト) = ' + str(accr_test))
        
    # パラメータに勾配適用
    for key in ('W1', 'W2', 'W3', 'b1', 'b2', 'b3'):
        network[key]  -= learning_rate * grad[key]


lists = range(0, iters_num, plot_interval)
plt.plot(lists, accuracies_train, label="training set")
plt.plot(lists, accuracies_test,  label="test set")
plt.legend(loc="lower right")
plt.title("accuracy")
plt.xlabel("count")
plt.ylabel("accuracy")
plt.ylim(0, 1.0)
# グラフの表示
plt.show()

image.png

  • 中間層の変更(Relu)

L1=40,L2=20

image.png

L1=100,L2=80

image.png

L1=300,L2= 100

image.png

  • Relu - Xavier
    image.png

  • Sigmoid - Xavier

image.png

  • Relu - He

image.png

  • sigmoid - He

image.png

####考察
勾配消失を起こさないためには初期値の設定方法が重要である。
中間層を増やすと、少ない反復思考で良い結果始め、最終的な精度も上がる。

####参考
バッチ正規化
https://www.renom.jp/ja/notebooks/tutorial/basic_algorithm/batch_normalization/notebook.html
https://www.slideshare.net/artimpact/batch-normalization-effectiveness20190206

#Section2 学習率最適化手法

###ポイント

  • 学習率が大きい場合、最適値にいつまでもたどり着かず発散してしまう。
  • 学習率が小さい場合、発散することはないが、小さすぎると収束までに時間がかかる、大域局所最適値に収束し辛くなる。
  • 学習最適化手法を利用して学習率を最適化する
  • 学習率最適化手法
    • モメンタム
    • AdaGrad
    • RMSProp
    • Adam

##モメンタム
image.png


V_t = \mu V_{t-1}-\epsilon \nabla E
\\
w^{(t+1)} = w^{(t)}+ V_t

慣性(ハイパーパラメータ)$\mu$は0.05-0.09が多い

#####確認テスト
モメンタム,AdaGrad,RMSPropの特徴を簡潔に説明せよ

  • モメンタム
    • 大域最適解になる。最適値にいくまでの時間が早い
  • AdaGrad
    • 勾配の緩やかな斜面に対して有効
  • RMSProp
    • ハイパーパラメータの調整が必要な場合が少ない

##AdaGrad

image.png

h_0 =\theta
\\
h_t = h_{t-1}+ (\nabla E)^2
\\
w^{(t+1)} = w^{(t)}-\epsilon \frac{1}{\sqrt{h_t} + \theta } \nabla E

$\theta$が足されているのは分母が0にならないようにするため

##RMSProp
image.png
AdaGradの課題(鞍点)を克服

h_t = \alpha + (1- \alpha)(\nabla E)^2
\\
w^{(t+1)} = w^{(t)}-\epsilon \frac{1}{\sqrt{h_t} + \theta } \nabla E

##Adam
image.png

一般的に多く多用される

###ハンズオン

ソースコード名
2_4_optimizer
  • SGD
Jupyter

import sys, os
sys.path.append(os.pardir)  # 親ディレクトリのファイルをインポートするための設定
import numpy as np
from collections import OrderedDict
from common import layers
from data.mnist import load_mnist
import matplotlib.pyplot as plt
from multi_layer_net import MultiLayerNet


# データの読み込み
(x_train, d_train), (x_test, d_test) = load_mnist(normalize=True, one_hot_label=True)

print("データ読み込み完了")

# batch_normalizationの設定 ================================
# use_batchnorm = True
use_batchnorm = False
# ====================================================


network = MultiLayerNet(input_size=784, hidden_size_list=[40, 20], output_size=10, activation='sigmoid', weight_init_std=0.01,
                       use_batchnorm=use_batchnorm)

iters_num = 1000
train_size = x_train.shape[0]
batch_size = 100
learning_rate = 0.01

train_loss_list = []
accuracies_train = []
accuracies_test = []

plot_interval=10

for i in range(iters_num):
    batch_mask = np.random.choice(train_size, batch_size)
    x_batch = x_train[batch_mask]
    d_batch = d_train[batch_mask]

    # 勾配
    grad = network.gradient(x_batch, d_batch)
    
    for key in ('W1', 'W2', 'W3', 'b1', 'b2', 'b3'):
        network.params[key] -= learning_rate * grad[key]
        
        loss = network.loss(x_batch, d_batch)
        train_loss_list.append(loss)
    
    
    if (i + 1) % plot_interval == 0:
        accr_test = network.accuracy(x_test, d_test)
        accuracies_test.append(accr_test)        
        accr_train = network.accuracy(x_batch, d_batch)
        accuracies_train.append(accr_train)
        
        print('Generation: ' + str(i+1) + '. 正答率(トレーニング) = ' + str(accr_train))
        print('                : ' + str(i+1) + '. 正答率(テスト) = ' + str(accr_test))

        
lists = range(0, iters_num, plot_interval)
plt.plot(lists, accuracies_train, label="training set")
plt.plot(lists, accuracies_test,  label="test set")
plt.legend(loc="lower right")
plt.title("accuracy")
plt.xlabel("count")
plt.ylabel("accuracy")
plt.ylim(0, 1.0)
# グラフの表示
plt.show()

  • learing rate 0.001
    image.png

  • learing rate 1.00
    image.png

  • learing rate 0.0001
    image.png

  • モメンタム

Jupyter

# データの読み込み
(x_train, d_train), (x_test, d_test) = load_mnist(normalize=True, one_hot_label=True)

print("データ読み込み完了")

# batch_normalizationの設定 ================================
# use_batchnorm = True
use_batchnorm = False
# ====================================================

network = MultiLayerNet(input_size=784, hidden_size_list=[40, 20], output_size=10, activation='sigmoid', weight_init_std=0.01,
                       use_batchnorm=use_batchnorm)

iters_num = 1000
train_size = x_train.shape[0]
batch_size = 100
learning_rate = 0.01
# 慣性
momentum = 0.9

train_loss_list = []
accuracies_train = []
accuracies_test = []

plot_interval=10

for i in range(iters_num):
    batch_mask = np.random.choice(train_size, batch_size)
    x_batch = x_train[batch_mask]
    d_batch = d_train[batch_mask]

    # 勾配
    grad = network.gradient(x_batch, d_batch)
    if i == 0:
        v = {}
    for key in ('W1', 'W2', 'W3', 'b1', 'b2', 'b3'):
        if i == 0:
            v[key] = np.zeros_like(network.params[key])
        v[key] = momentum * v[key] - learning_rate * grad[key]
        network.params[key] += v[key]

        loss = network.loss(x_batch, d_batch)
        train_loss_list.append(loss)
        
    if (i + 1) % plot_interval == 0:
        accr_test = network.accuracy(x_test, d_test)
        accuracies_test.append(accr_test)        
        accr_train = network.accuracy(x_batch, d_batch)
        accuracies_train.append(accr_train)

        print('Generation: ' + str(i+1) + '. 正答率(トレーニング) = ' + str(accr_train))
        print('                : ' + str(i+1) + '. 正答率(テスト) = ' + str(accr_test))
        
        
lists = range(0, iters_num, plot_interval)
plt.plot(lists, accuracies_train, label="training set")
plt.plot(lists, accuracies_test,  label="test set")
plt.legend(loc="lower right")
plt.title("accuracy")
plt.xlabel("count")
plt.ylabel("accuracy")
plt.ylim(0, 1.0)
# グラフの表示
plt.show()

  • $\mu=0.9$
    image.png

  • $\mu=1.0$
    image.png

  • $\mu=0.5$
    image.png

  • $\mu=0.3$

image.png

  • MomentumをもとにAdaGradを作ってみよう
    • θ = 1e-4 とする
Jupyter
# AdaGradを作ってみよう
# データの読み込み
(x_train, d_train), (x_test, d_test) = load_mnist(normalize=True, one_hot_label=True)

print("データ読み込み完了")

# batch_normalizationの設定 ================================
# use_batchnorm = True
use_batchnorm = False
# ====================================================

network = MultiLayerNet(input_size=784, hidden_size_list=[40, 20], output_size=10, activation='sigmoid', weight_init_std=0.01,
                       use_batchnorm=use_batchnorm)

iters_num = 1000
# iters_num = 500 # 処理を短縮

train_size = x_train.shape[0]
batch_size = 100
learning_rate = 0.01

# AdaGradでは不必要
# =============================

#momentum = 0.9 

# =============================

train_loss_list = []
accuracies_train = []
accuracies_test = []

plot_interval=10

for i in range(iters_num):
    batch_mask = np.random.choice(train_size, batch_size)
    x_batch = x_train[batch_mask]
    d_batch = d_train[batch_mask]

    # 勾配
    grad = network.gradient(x_batch, d_batch)
    if i == 0:
        h = {}
    for key in ('W1', 'W2', 'W3', 'b1', 'b2', 'b3'):
            
        # 変更しよう  
        # ===========================================
        if i == 0:
            h[key] = np.zeros_like(network.params[key])        
        #h[key] = momentum * h[key] - learning_rate * grad[key]
        h[key] += grad[key]*grad[key]
        network.params[key] -= learning_rate * grad[key] / (np.sqrt(h[key])+1e-7)

        # ===========================================
        
        loss = network.loss(x_batch, d_batch)
        train_loss_list.append(loss)
        
    if (i + 1) % plot_interval == 0:
        accr_test = network.accuracy(x_test, d_test)
        accuracies_test.append(accr_test)        
        accr_train = network.accuracy(x_batch, d_batch)
        accuracies_train.append(accr_train)

        print('Generation: ' + str(i+1) + '. 正答率(トレーニング) = ' + str(accr_train))
        print('                : ' + str(i+1) + '. 正答率(テスト) = ' + str(accr_test))
        
        
lists = range(0, iters_num, plot_interval)
plt.plot(lists, accuracies_train, label="training set")
plt.plot(lists, accuracies_test,  label="test set")
plt.legend(loc="lower right")
plt.title("accuracy")
plt.xlabel("count")
plt.ylabel("accuracy")
plt.ylim(0, 1.0)
# グラフの表示
plt.show()

image.png

  • RSMprop
Jupyter

# データの読み込み
(x_train, d_train), (x_test, d_test) = load_mnist(normalize=True, one_hot_label=True)

print("データ読み込み完了")

# batch_normalizationの設定 ================================
# use_batchnorm = True
use_batchnorm = False
# ====================================================

network = MultiLayerNet(input_size=784, hidden_size_list=[40, 20], output_size=10, activation='sigmoid', weight_init_std=0.01,
                       use_batchnorm=use_batchnorm)

iters_num = 1000
train_size = x_train.shape[0]
batch_size = 100
learning_rate = 0.01
decay_rate = 0.99

train_loss_list = []
accuracies_train = []
accuracies_test = []

plot_interval=10

for i in range(iters_num):
    batch_mask = np.random.choice(train_size, batch_size)
    x_batch = x_train[batch_mask]
    d_batch = d_train[batch_mask]

    # 勾配
    grad = network.gradient(x_batch, d_batch)
    if i == 0:
        h = {}
    for key in ('W1', 'W2', 'W3', 'b1', 'b2', 'b3'):
        if i == 0:
            h[key] = np.zeros_like(network.params[key])
        h[key] *= decay_rate
        h[key] += (1 - decay_rate) * np.square(grad[key])
        network.params[key] -= learning_rate * grad[key] / (np.sqrt(h[key]) + 1e-7)

        loss = network.loss(x_batch, d_batch)
        train_loss_list.append(loss)                
        
    if (i + 1) % plot_interval == 0:
        accr_test = network.accuracy(x_test, d_test)
        accuracies_test.append(accr_test)        
        accr_train = network.accuracy(x_batch, d_batch)
        accuracies_train.append(accr_train)
        
        print('Generation: ' + str(i+1) + '. 正答率(トレーニング) = ' + str(accr_train))
        print('                : ' + str(i+1) + '. 正答率(テスト) = ' + str(accr_test))
        
        
lists = range(0, iters_num, plot_interval)
plt.plot(lists, accuracies_train, label="training set")
plt.plot(lists, accuracies_test,  label="test set")
plt.legend(loc="lower right")
plt.title("accuracy")
plt.xlabel("count")
plt.ylabel("accuracy")
plt.ylim(0, 1.0)
# グラフの表示
plt.show()
  • learning rate 0.01

image.png

  • learning rate 0.0001

image.png

  • learning rate 0.1

image.png

  • decay_rate 0.01

image.png

  • decay_rate 1 $(1-\alpha)がゼロになる。$

image.png

  • Adam
Jupyter


# データの読み込み
(x_train, d_train), (x_test, d_test) = load_mnist(normalize=True, one_hot_label=True)

print("データ読み込み完了")

# batch_normalizationの設定 ================================
# use_batchnorm = True
use_batchnorm = False
# ====================================================

network = MultiLayerNet(input_size=784, hidden_size_list=[40, 20], output_size=10, activation='sigmoid', weight_init_std=0.01,
                       use_batchnorm=use_batchnorm)

iters_num = 1000
train_size = x_train.shape[0]
batch_size = 100
learning_rate = 0.01
beta1 = 0.9
beta2 = 0.999

train_loss_list = []
accuracies_train = []
accuracies_test = []

plot_interval=10

for i in range(iters_num):
    batch_mask = np.random.choice(train_size, batch_size)
    x_batch = x_train[batch_mask]
    d_batch = d_train[batch_mask]

    # 勾配
    grad = network.gradient(x_batch, d_batch)
    if i == 0:
        m = {}
        v = {}
    learning_rate_t  = learning_rate * np.sqrt(1.0 - beta2 ** (i + 1)) / (1.0 - beta1 ** (i + 1))    
    for key in ('W1', 'W2', 'W3', 'b1', 'b2', 'b3'):
        if i == 0:
            m[key] = np.zeros_like(network.params[key])
            v[key] = np.zeros_like(network.params[key])
            
        m[key] += (1 - beta1) * (grad[key] - m[key])
        v[key] += (1 - beta2) * (grad[key] ** 2 - v[key])            
        network.params[key] -= learning_rate_t * m[key] / (np.sqrt(v[key]) + 1e-7)                
        
        
    if (i + 1) % plot_interval == 0:
        accr_test = network.accuracy(x_test, d_test)
        accuracies_test.append(accr_test)        
        accr_train = network.accuracy(x_batch, d_batch)
        accuracies_train.append(accr_train)
        loss = network.loss(x_batch, d_batch)
        train_loss_list.append(loss)        
        
        print('Generation: ' + str(i+1) + '. 正答率(トレーニング) = ' + str(accr_train))
        print('                : ' + str(i+1) + '. 正答率(テスト) = ' + str(accr_test))
                

lists = range(0, iters_num, plot_interval)
plt.plot(lists, accuracies_train, label="training set")
plt.plot(lists, accuracies_test,  label="test set")
plt.legend(loc="lower right")
plt.title("accuracy")
plt.xlabel("count")
plt.ylabel("accuracy")
plt.ylim(0, 1.0)
# グラフの表示
plt.show()
  • Adam

image.png

  • Adam(バッチ正規化)
    image.png

####考察
ハンズオンの結果からRMSPropとAdamは高い精度を出している。
RMSPropのパラメータ設定において1を設定すると $(1-\alpha)がゼロにってしまい学習しなくなるので注意が必要。

####参考
https://qiita.com/ZoneTsuyoshi/items/8ef6fa1e154d176e25b8
https://www.slideshare.net/nishio/ss-66840545

#Section3 過学習

image.png

###ポイント

  • テスト誤差と訓練誤差とで学習曲線が乖離すること

  • 原因

    • パラメータ数が多い
    • パラメータの値が適切でない
    • ノードが多い
    • ネットワークの自由度が高いのが原因

  • 正則化手法を利用して過学習を抑制する

    • ネットワークの自由度を制約する。

  • Weight decay(荷重減衰)

    • image.png

  • 正則化手法

    • L1正則化、L2正則化
    • ドロップアウト

  • L1正則化、L2正則化

    • 誤差関数にPノルムを加える
    • P=1はL1正規化、P=2はL2正規化
    • L2は精度向上しやすい。計算リソースが大きい場合はL2を使う。
E_n(w) + \frac {1}{p} \lambda　||x||_p    
\\
||x||_p = (|x_1|^p + \cdots +|x_n|^p)^{\frac{1}{p}}
  • ドロップアウト
    • ランダムにノードを削除して学習
    • データ量を変化させずに異なるモデルを学習させていると解釈できる

#####確認テスト
image.png

=>(a)

#####確認テスト

  • L1正則化を表しているのはどちらか

image.png

=>Lasso推定量

#####例題チャレンジ
image.png

=>(4)

#####例題チャレンジ
image.png

=>(3)

#####例題チャレンジ

image.png

=>(4)

###ハンズオン

ソースコード名
2_5_overfiting
Jupyter
import sys, os
sys.path.append(os.pardir)  # 親ディレクトリのファイルをインポートするための設定
import numpy as np
from collections import OrderedDict
from common import layers
from data.mnist import load_mnist
import matplotlib.pyplot as plt
from multi_layer_net import MultiLayerNet
from common import optimizer


(x_train, d_train), (x_test, d_test) = load_mnist(normalize=True)

print("データ読み込み完了")

# 過学習を再現するために、学習データを削減
x_train = x_train[:300]
d_train = d_train[:300]

network = MultiLayerNet(input_size=784, hidden_size_list=[100, 100, 100, 100, 100, 100], output_size=10)
optimizer = optimizer.SGD(learning_rate=0.01)

iters_num = 1000
train_size = x_train.shape[0]
batch_size = 100

train_loss_list = []
accuracies_train = []
accuracies_test = []

plot_interval=10


for i in range(iters_num):
    batch_mask = np.random.choice(train_size, batch_size)
    x_batch = x_train[batch_mask]
    d_batch = d_train[batch_mask]

    grad = network.gradient(x_batch, d_batch)
    optimizer.update(network.params, grad)

    loss = network.loss(x_batch, d_batch)
    train_loss_list.append(loss)
        
    if (i+1) % plot_interval == 0:
        accr_train = network.accuracy(x_train, d_train)
        accr_test = network.accuracy(x_test, d_test)
        accuracies_train.append(accr_train)
        accuracies_test.append(accr_test)

        print('Generation: ' + str(i+1) + '. 正答率(トレーニング) = ' + str(accr_train))
        print('                : ' + str(i+1) + '. 正答率(テスト) = ' + str(accr_test))        

lists = range(0, iters_num, plot_interval)
plt.plot(lists, accuracies_train, label="training set")
plt.plot(lists, accuracies_test,  label="test set")
plt.legend(loc="lower right")
plt.title("accuracy")
plt.xlabel("count")
plt.ylabel("accuracy")
plt.ylim(0, 1.0)
# グラフの表示
plt.show()

過学習
image.png

weight decay

  • L2
Jupyter
from common import optimizer

(x_train, d_train), (x_test, d_test) = load_mnist(normalize=True)

print("データ読み込み完了")

# 過学習を再現するために、学習データを削減
x_train = x_train[:300]
d_train = d_train[:300]


network = MultiLayerNet(input_size=784, hidden_size_list=[100, 100, 100, 100, 100, 100], output_size=10)


iters_num = 1000
train_size = x_train.shape[0]
batch_size = 100
learning_rate=0.01

train_loss_list = []
accuracies_train = []
accuracies_test = []

plot_interval=10
hidden_layer_num = network.hidden_layer_num

# 正則化強度設定 ======================================
weight_decay_lambda = 0.1
# =================================================

for i in range(iters_num):
    batch_mask = np.random.choice(train_size, batch_size)
    x_batch = x_train[batch_mask]
    d_batch = d_train[batch_mask]

    grad = network.gradient(x_batch, d_batch)
    weight_decay = 0
    
    for idx in range(1, hidden_layer_num+1):
        grad['W' + str(idx)] = network.layers['Affine' + str(idx)].dW + weight_decay_lambda * network.params['W' + str(idx)]
        grad['b' + str(idx)] = network.layers['Affine' + str(idx)].db
        network.params['W' + str(idx)] -= learning_rate * grad['W' + str(idx)]
        network.params['b' + str(idx)] -= learning_rate * grad['b' + str(idx)]        
        weight_decay += 0.5 * weight_decay_lambda * np.sqrt(np.sum(network.params['W' + str(idx)] ** 2))

    loss = network.loss(x_batch, d_batch) + weight_decay
    train_loss_list.append(loss)        
        
    if (i+1) % plot_interval == 0:
        accr_train = network.accuracy(x_train, d_train)
        accr_test = network.accuracy(x_test, d_test)
        accuracies_train.append(accr_train)
        accuracies_test.append(accr_test)
        
        print('Generation: ' + str(i+1) + '. 正答率(トレーニング) = ' + str(accr_train))
        print('                : ' + str(i+1) + '. 正答率(テスト) = ' + str(accr_test))               


lists = range(0, iters_num, plot_interval)
plt.plot(lists, accuracies_train, label="training set")
plt.plot(lists, accuracies_test,  label="test set")
plt.legend(loc="lower right")
plt.title("accuracy")
plt.xlabel("count")
plt.ylabel("accuracy")
plt.ylim(0, 1.0)
# グラフの表示
plt.show()

  • 強度0.1
    image.png

  • 強度0.2
    image.png

  • 強度0.05
    image.png

  • L1

Jupyter
(x_train, d_train), (x_test, d_test) = load_mnist(normalize=True)

print("データ読み込み完了")

# 過学習を再現するために、学習データを削減
x_train = x_train[:300]
d_train = d_train[:300]

network = MultiLayerNet(input_size=784, hidden_size_list=[100, 100, 100, 100, 100, 100], output_size=10)


iters_num = 1000
train_size = x_train.shape[0]
batch_size = 100
learning_rate=0.1

train_loss_list = []
accuracies_train = []
accuracies_test = []

plot_interval=10
hidden_layer_num = network.hidden_layer_num

# 正則化強度設定 ======================================
weight_decay_lambda = 0.005
# =================================================

for i in range(iters_num):
    batch_mask = np.random.choice(train_size, batch_size)
    x_batch = x_train[batch_mask]
    d_batch = d_train[batch_mask]

    grad = network.gradient(x_batch, d_batch)
    weight_decay = 0
    
    for idx in range(1, hidden_layer_num+1):
        grad['W' + str(idx)] = network.layers['Affine' + str(idx)].dW + weight_decay_lambda * np.sign(network.params['W' + str(idx)])
        grad['b' + str(idx)] = network.layers['Affine' + str(idx)].db
        network.params['W' + str(idx)] -= learning_rate * grad['W' + str(idx)]
        network.params['b' + str(idx)] -= learning_rate * grad['b' + str(idx)]        
        weight_decay += weight_decay_lambda * np.sum(np.abs(network.params['W' + str(idx)]))

    loss = network.loss(x_batch, d_batch) + weight_decay
    train_loss_list.append(loss)        
        
    if (i+1) % plot_interval == 0:
        accr_train = network.accuracy(x_train, d_train)
        accr_test = network.accuracy(x_test, d_test)
        accuracies_train.append(accr_train)
        accuracies_test.append(accr_test)
        
        print('Generation: ' + str(i+1) + '. 正答率(トレーニング) = ' + str(accr_train))
        print('                : ' + str(i+1) + '. 正答率(テスト) = ' + str(accr_test))               
                
lists = range(0, iters_num, plot_interval)
plt.plot(lists, accuracies_train, label="training set")
plt.plot(lists, accuracies_test,  label="test set")
plt.legend(loc="lower right")
plt.title("accuracy")
plt.xlabel("count")
plt.ylabel("accuracy")
plt.ylim(0, 1.0)
# グラフの表示
plt.show()

image.png

  • ドロップアウト
Jupyter
#dropout_ratioの閾値を儲ける

class Dropout:
    def __init__(self, dropout_ratio=0.5):
        self.dropout_ratio = dropout_ratio
        self.mask = None

    def forward(self, x, train_flg=True):
        if train_flg:
            self.mask = np.random.rand(*x.shape) > self.dropout_ratio
            return x * self.mask
        else:
            return x * (1.0 - self.dropout_ratio)

    def backward(self, dout):
        return dout * self.mask
Jupyter
from common import optimizer
(x_train, d_train), (x_test, d_test) = load_mnist(normalize=True)

print("データ読み込み完了")

# 過学習を再現するために、学習データを削減
x_train = x_train[:300]
d_train = d_train[:300]

# ドロップアウト設定 ======================================
use_dropout = True
dropout_ratio = 0.15
# ====================================================

network = MultiLayerNet(input_size=784, hidden_size_list=[100, 100, 100, 100, 100, 100], output_size=10,
                        weight_decay_lambda=weight_decay_lambda, use_dropout = use_dropout, dropout_ratio = dropout_ratio)
optimizer = optimizer.SGD(learning_rate=0.01)
# optimizer = optimizer.Momentum(learning_rate=0.01, momentum=0.9)
# optimizer = optimizer.AdaGrad(learning_rate=0.01)
# optimizer = optimizer.Adam()

iters_num = 1000
train_size = x_train.shape[0]
batch_size = 100

train_loss_list = []
accuracies_train = []
accuracies_test = []

plot_interval=10


for i in range(iters_num):
    batch_mask = np.random.choice(train_size, batch_size)
    x_batch = x_train[batch_mask]
    d_batch = d_train[batch_mask]

    grad = network.gradient(x_batch, d_batch)
    optimizer.update(network.params, grad)

    loss = network.loss(x_batch, d_batch)
    train_loss_list.append(loss)    
    
    if (i+1) % plot_interval == 0:
        accr_train = network.accuracy(x_train, d_train)
        accr_test = network.accuracy(x_test, d_test)
        accuracies_train.append(accr_train)
        accuracies_test.append(accr_test)

        print('Generation: ' + str(i+1) + '. 正答率(トレーニング) = ' + str(accr_train))
        print('                : ' + str(i+1) + '. 正答率(テスト) = ' + str(accr_test))        
        
lists = range(0, iters_num, plot_interval)
plt.plot(lists, accuracies_train, label="training set")
plt.plot(lists, accuracies_test,  label="test set")
plt.legend(loc="lower right")
plt.title("accuracy")
plt.xlabel("count")
plt.ylabel("accuracy")
plt.ylim(0, 1.0)
# グラフの表示
plt.show()

  • dropout 0.15 + SGD
    image.png

  • dropout 0.15 + モメンタム
    image.png

  • dropout 0.15 + AdaGrad
    image.png

  • dropout 0.15 + Adam
    image.png

  • dropout なし + Adam
    image.png

  • dropout 0.5 + Adam

image.png

  • dropout 0.01 + Adam
    image.png

  • dropout 0.2 + Adam
    image.png

  • Dropout + L1

Jupyter
from common import optimizer
(x_train, d_train), (x_test, d_test) = load_mnist(normalize=True)

print("データ読み込み完了")

# 過学習を再現するために、学習データを削減
x_train = x_train[:300]
d_train = d_train[:300]

# ドロップアウト設定 ======================================
use_dropout = True
dropout_ratio = 0.08
# ====================================================

network = MultiLayerNet(input_size=784, hidden_size_list=[100, 100, 100, 100, 100, 100], output_size=10,
                        use_dropout = use_dropout, dropout_ratio = dropout_ratio)

iters_num = 1000
train_size = x_train.shape[0]
batch_size = 100
learning_rate=0.01

train_loss_list = []
accuracies_train = []
accuracies_test = []
hidden_layer_num = network.hidden_layer_num

plot_interval=10

# 正則化強度設定 ======================================
weight_decay_lambda=0.004
# =================================================

for i in range(iters_num):
    batch_mask = np.random.choice(train_size, batch_size)
    x_batch = x_train[batch_mask]
    d_batch = d_train[batch_mask]

    grad = network.gradient(x_batch, d_batch)
    weight_decay = 0
    
    for idx in range(1, hidden_layer_num+1):
        grad['W' + str(idx)] = network.layers['Affine' + str(idx)].dW + weight_decay_lambda * np.sign(network.params['W' + str(idx)])
        grad['b' + str(idx)] = network.layers['Affine' + str(idx)].db
        network.params['W' + str(idx)] -= learning_rate * grad['W' + str(idx)]
        network.params['b' + str(idx)] -= learning_rate * grad['b' + str(idx)]        
        weight_decay += weight_decay_lambda * np.sum(np.abs(network.params['W' + str(idx)]))

    loss = network.loss(x_batch, d_batch) + weight_decay
    train_loss_list.append(loss)        
        
    if (i+1) % plot_interval == 0:
        accr_train = network.accuracy(x_train, d_train)
        accr_test = network.accuracy(x_test, d_test)
        accuracies_train.append(accr_train)
        accuracies_test.append(accr_test)
        
        print('Generation: ' + str(i+1) + '. 正答率(トレーニング) = ' + str(accr_train))
        print('                : ' + str(i+1) + '. 正答率(テスト) = ' + str(accr_test))               
        
lists = range(0, iters_num, plot_interval)
plt.plot(lists, accuracies_train, label="training set")
plt.plot(lists, accuracies_test,  label="test set")
plt.legend(loc="lower right")
plt.title("accuracy")
plt.xlabel("count")
plt.ylabel("accuracy")
plt.ylim(0, 1.0)
# グラフの表示
plt.show()

image.png

####考察
L1,L2正則化において正則化強度を強くするほど過学習は抑制できるがテストデータの学習精度も下がる。
ドロップアウト率を大きくしすぎると抑制されすぎる、小さすぎると過学習の抑制ができなくなる。

####参考
https://python.atelierkobato.com/sign/

#Section4 畳み込みニューラルネットワークの概念
(消えてしまったので作成し直し、、)

image.png

image.png

###ポイント

  • 三次元画像にも対応

##4-1畳み込み層
###4-1-1バイアス
出力画像にバイアスを追加する。
image.png

###4-1-2パディング
0を入れるなどして入力画像の調整を行う

image.png

###4-1-3ストライド

フィルタの移動幅を決める

image.png

###4-1-4チャンネル
フィルタの数

image.png

##4-2 プーリング層

  • Max pooling
  • Avarage pooling

image.png

#####確認テスト

  • サイズ6×6の入力画像をサイズ2×2のフィルタで畳み込んだ時の出力画像のサイズを答えよ。スライドとパディングは1とする。

=>7×7


OH = \frac{H+2P-FH}{S}+1
\\
OW = \frac{H+2P-FW}{S}+1

###ハンズオン

ソースコード名
2_6_simple_convolution_network
  • im2col
Jupyter


import sys, os
sys.path.append(os.pardir)
import pickle
import numpy as np
from collections import OrderedDict
from common import layers
from common import optimizer
from data.mnist import load_mnist
import matplotlib.pyplot as plt

# 画像データを２次元配列に変換
'''
input_data: 入力値
filter_h: フィルターの高さ
filter_w: フィルターの横幅
stride: ストライド
pad: パディング
'''
def im2col(input_data, filter_h, filter_w, stride=1, pad=0):
    # N: number, C: channel, H: height, W: width
    N, C, H, W = input_data.shape
    out_h = (H + 2 * pad - filter_h)//stride + 1
    out_w = (W + 2 * pad - filter_w)//stride + 1

    img = np.pad(input_data, [(0,0), (0,0), (pad, pad), (pad, pad)], 'constant')
    col = np.zeros((N, C, filter_h, filter_w, out_h, out_w))

    for y in range(filter_h):
        y_max = y + stride * out_h
        for x in range(filter_w):
            x_max = x + stride * out_w
            col[:, :, y, x, :, :] = img[:, :, y:y_max:stride, x:x_max:stride]
    
    col = col.transpose(0, 4, 5, 1, 2, 3) # (N, C, filter_h, filter_w, out_h, out_w) -> (N, filter_w, out_h, out_w, C, filter_h)    
    
    col = col.reshape(N * out_h * out_w, -1)
    return col

# im2colの処理確認
input_data = np.random.rand(2, 1, 4, 4)*100//1 # number, channel, height, widthを表す
print('========== input_data ===========\n', input_data)
print('==============================')
filter_h = 3
filter_w = 3
stride = 1
pad = 0
col = im2col(input_data, filter_h=filter_h, filter_w=filter_w, stride=stride, pad=pad)
print('============= col ==============\n', col)
print('==============================')

２次元に変更

出力



========== input_data ===========
 [[[[36. 81. 79. 54.]
   [ 6. 72.  5. 87.]
   [57.  4. 37.  8.]
   [78. 42. 79. 61.]]]


 [[[33. 38. 44. 78.]
   [19. 72. 53. 34.]
   [92. 18. 52. 31.]
   [13. 38. 10. 93.]]]]
==============================
============= col ==============
 [[36. 81. 79.  6. 72.  5. 57.  4. 37.]
 [81. 79. 54. 72.  5. 87.  4. 37.  8.]
 [ 6. 72.  5. 57.  4. 37. 78. 42. 79.]
 [72.  5. 87.  4. 37.  8. 42. 79. 61.]
 [33. 38. 44. 19. 72. 53. 92. 18. 52.]
 [38. 44. 78. 72. 53. 34. 18. 52. 31.]
 [19. 72. 53. 92. 18. 52. 13. 38. 10.]
 [72. 53. 34. 18. 52. 31. 38. 10. 93.]]
==============================
  • col2im
Jupyter

# ２次元配列を画像データに変換
def col2im(col, input_shape, filter_h, filter_w, stride=1, pad=0):
    # N: number, C: channel, H: height, W: width
    N, C, H, W = input_shape
    # 切り捨て除算    
    out_h = (H + 2 * pad - filter_h)//stride + 1
    out_w = (W + 2 * pad - filter_w)//stride + 1
    col = col.reshape(N, out_h, out_w, C, filter_h, filter_w).transpose(0, 3, 4, 5, 1, 2) # (N, filter_h, filter_w, out_h, out_w, C)

    img = np.zeros((N, C, H + 2 * pad + stride - 1, W + 2 * pad + stride - 1))
    for y in range(filter_h):
        y_max = y + stride * out_h
        for x in range(filter_w):
            x_max = x + stride * out_w
            img[:, :, y:y_max:stride, x:x_max:stride] += col[:, :, y, x, :, :]

    return img[:, :, pad:H + pad, pad:W + pad]

image=col2im(col, input_shape=input_data.shape, filter_h=filter_h, filter_w=filter_w, stride=stride, pad=pad)
print(image)
出力

[[[[ 36. 162. 158.  54.]
   [ 12. 288.  20. 174.]
   [114.  16. 148.  16.]
   [ 78.  84. 158.  61.]]]


 [[[ 33.  76.  88.  78.]
   [ 38. 288. 212.  68.]
   [184.  72. 208.  62.]
   [ 13.  76.  20.  93.]]]]
  • プーリング
Jupyter

class Pooling:
    def __init__(self, pool_h, pool_w, stride=1, pad=0):
        self.pool_h = pool_h
        self.pool_w = pool_w
        self.stride = stride
        self.pad = pad
        
        self.x = None
        self.arg_max = None

    def forward(self, x):
        N, C, H, W = x.shape
        out_h = int(1 + (H - self.pool_h) / self.stride)
        out_w = int(1 + (W - self.pool_w) / self.stride)
        
        # xを行列に変換
        col = im2col(x, self.pool_h, self.pool_w, self.stride, self.pad)
        # プーリングのサイズに合わせてリサイズ
        col = col.reshape(-1, self.pool_h*self.pool_w)
        
        # 行ごとに最大値を求める
        arg_max = np.argmax(col, axis=1)
        out = np.max(col, axis=1)
        # 整形
        out = out.reshape(N, out_h, out_w, C).transpose(0, 3, 1, 2)

        self.x = x
        self.arg_max = arg_max

        return out

    def backward(self, dout):
        dout = dout.transpose(0, 2, 3, 1)
        
        pool_size = self.pool_h * self.pool_w
        dmax = np.zeros((dout.size, pool_size))
        dmax[np.arange(self.arg_max.size), self.arg_max.flatten()] = dout.flatten()
        dmax = dmax.reshape(dout.shape + (pool_size,)) 
        
        dcol = dmax.reshape(dmax.shape[0] * dmax.shape[1] * dmax.shape[2], -1)
        dx = col2im(dcol, self.x.shape, self.pool_h, self.pool_w, self.stride, self.pad)
        
        return dx
Jupyter
from common import optimizer

# データの読み込み
(x_train, d_train), (x_test, d_test) = load_mnist(flatten=False)

print("データ読み込み完了")

# 処理に時間のかかる場合はデータを削減 
x_train, d_train = x_train[:5000], d_train[:5000]
x_test, d_test = x_test[:1000], d_test[:1000]


network = SimpleConvNet(input_dim=(1,28,28), conv_param = {'filter_num': 30, 'filter_size': 5, 'pad': 0, 'stride': 1},
                        hidden_size=100, output_size=10, weight_init_std=0.01)

optimizer = optimizer.Adam()

iters_num = 1000
train_size = x_train.shape[0]
batch_size = 100

train_loss_list = []
accuracies_train = []
accuracies_test = []

plot_interval=10



for i in range(iters_num):
    batch_mask = np.random.choice(train_size, batch_size)
    x_batch = x_train[batch_mask]
    d_batch = d_train[batch_mask]
    
    grad = network.gradient(x_batch, d_batch)
    optimizer.update(network.params, grad)

    loss = network.loss(x_batch, d_batch)
    train_loss_list.append(loss)

    if (i+1) % plot_interval == 0:
        accr_train = network.accuracy(x_train, d_train)
        accr_test = network.accuracy(x_test, d_test)
        accuracies_train.append(accr_train)
        accuracies_test.append(accr_test)
        
        print('Generation: ' + str(i+1) + '. 正答率(トレーニング) = ' + str(accr_train))
        print('                : ' + str(i+1) + '. 正答率(テスト) = ' + str(accr_test))               

lists = range(0, iters_num, plot_interval)
plt.plot(lists, accuracies_train, label="training set")
plt.plot(lists, accuracies_test,  label="test set")
plt.legend(loc="lower right")
plt.title("accuracy")
plt.xlabel("count")
plt.ylabel("accuracy")
plt.ylim(0, 1.0)
# グラフの表示
plt.show()

image.png

####考察
入力データと出力データは同じものにならない。
CNNは非常に時間がかかるので高スペックのマシーンかGPUを使用することをおすすめする。

####参考
https://qiita.com/FukuharaYohei/items/73cce8f5707a353e3c3a

#Section5 最新のCNN

##AlexNet
###ポイント

  • 2012年の画像認識コンテストで優勝したモデル
  • 物体認識のために、初めて深層学習の概念および畳み込みニューラルネットワークの概念を取り入れたアーキテクチャである。
  • 全結合の出力にドロップアウトを用いて過学習を抑制している。

image.png

####考察
画像認識分野において深層学習＋CNNが有効であることを世界中に証明した。

####参考
https://axa.biopapyrus.jp/deep-learning/cnn/image-classification/alexnet.html

0

Register as a new user and use Qiita more conveniently

  1. You get articles that match your needs
  2. You can efficiently read back useful information
  3. You can use dark theme
What you can do with signing up
Sign upLogin

Delete article

Deleted articles cannot be recovered.

Draft of this article would be also deleted.

Are you sure you want to delete this article?