【CNN】バッチ正規化を導入した畳み込みニューラルネットワークを０から実装する（mnist画像識別・Accuracy98%以上）

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
from dataset.mnist import load_mnist

(train_data, train_label), (test_data, test_label) = load_mnist(flatten=True, normalize=True, one_hot_label=True)
print(train_data.shape)
print(train_label.shape)
print(test_data.shape)
print(test_label.shape)

pip install tensorflow

import tensorflow
print(tensorflow.__version__)
# >> 2.9.1

from tensorflow import keras
print(keras.__version__)
# >> 2.9.0

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from sklearn.model_selection import train_test_split
import pickle
from common.optimizer import Adam
from common.ConvNet import ConvNet
from keras.preprocessing.image import ImageDataGenerator

train_data = train_data.astype('float32')
#train_data = train_data.reshape(-1, 28*28) : flatten=Trueの代わりになる
#データの正規化
#train_data = (train_data - train_data.min()) / train_data.max() : normalize=Trueの代わりになる
print("train_data.shape=", train_data.shape)
# >> train_data.shape= (60000, 784)

epochs = 15
batch_size = 64

optimizer = Adam()

# CNNのオブジェクト生成
snet = ConvNet()

train_loss = []
test_loss = []
train_accuracy = []
test_accuracy = []

##データ拡張器
##datagen = ImageDataGenerator(rotation_range = 10, height_shift_range = 0.1,width_shift_range = 0.1,zoom_range= 0.1)

def __train(epochs):
    # 学習に時間がかかりすぎるので、ここでは4000データだけ使用した
    train, test, train_labels, test_labels = train_test_split(train_data[:4000], train_label[:4000], test_size=0.3,shuffle=True) 
    train = train.reshape(-1, 1, 28, 28)
    test = test.reshape(-1, 1, 28, 28)
    x = train
    t = train_labels
    x = x.reshape(-1,1,28,28) # 配列形式の変形

    # 繰り返し回数
    xsize = x.shape[0]
    iter_num = np.ceil(xsize / batch_size).astype(np.int)

    x = x.reshape(-1,1,28,28) # 配列形式の変形
    for epoch in range(epochs):
        print("epoch=%s"%epoch)

        # シャッフル
        idx = np.arange(xsize)
        np.random.shuffle(idx)

        for it in range(iter_num):
            """
            ランダムなミニバッチを順番に取り出す
            """
            print("it=", it)
            mask = idx[batch_size*it : batch_size*(it+1)]

            # ミニバッチの生成
            x_train = x[mask]
            t_train = t[mask]

            ##g = datagen.flow(x_train, batch_size = batch_size,shuffle=False)
            ##batches = g.next()
            batches = x_train

            # 勾配の計算 (誤差逆伝播法を用いる) 
            grads = snet.gradient(batches, t_train)

            # 更新
            optimizer.update(snet.params, grads)

        ## 学習経過の記録

        # 訓練データにおけるloss
        #     print("calculating train_loss")    
        train_loss.append(snet.loss(x,  t))

        #     print("calculating test_loss")
        # テストデータにおけるloss
        test_loss.append(snet.loss(test, test_labels))

        #     print("calculating train_accuracy")
        # 訓練データにて精度を確認
        train_accuracy.append(snet.accuracy(x, t))
    
        #     print("calculating test_accuracy")
        # テストデータにて精度を算出
        test_accuracy.append(snet.accuracy(test, test_labels))
        
        
    return train_accuracy, train_loss, test_accuracy, test_loss   

train_accuracy, train_loss, test_accuracy, test_loss  = __train(epochs)

# 評価性能を可視化する
df_log = pd.DataFrame({"train_loss":train_loss,
             "test_loss":test_loss,
             "train_accuracy":train_accuracy,
             "test_accuracy":test_accuracy})
df_log.plot()
plt.ylabel("loss or accuracy")
plt.xlabel("epochs")
plt.show()

test_data = test_data.reshape(-1, 1, 28, 28)
snet.accuracy(test_data, test_label)
# >> 0.9828

# 上記コードより抜粋
datagen = ImageDataGenerator(rotation_range = 10, height_shift_range = 0.1,width_shift_range = 0.1,zoom_range= 0.1)
...
for epoch in range(epochs):
    g = datagen.flow(x_train, batch_size = batch_size,shuffle=False)
    batches = g.next()

# 重みW, バイアスbの保存
with open("mnist_weight.pickle", "wb") as f:
    pickle.dump(snet.params, f)
# バッチ正規化の学習結果である移動平均を保存
with open("mnist_mv.pickle", "wb") as f:
    pickle.dump(snet.mv, f)

import pickle
from common.ConvNet_test import ConvNet_test

def func_predict(test_data, test_label):

    model = ConvNet_test()

    accuracy = model.accuracy(test_data, test_label)
    loss  = model.loss(test_data, test_label)
    
    return loss, accuracy 

func_predict(test_data, test_label)
# >> 0.982