キカガク（KIKAGAKU）で「ファインチューニング」を学ぶ

Photo by Owen Beard on Unsplash

今回は「ファインチューニング - KIKAGAKU」を学ぶ。

Table of Contents [Disable]

• 1.学習内容
• 2.ソースコード
• 3.出力結果
• 4.出力ログ
• 5.関連記事

学習内容

1. 画像分類／データセットの準備
2. CNN モデルの定義／目的関数と最適化手法の選択／モデルの学習／
3. ファインチューニング／全結合層を追加／モデルの学習／予測精度の評価

学習させるのに5日間を要した。（PCを使用している間のみ、バックグラウンドで学習。）

ソースコード

	import numpy as np
	import pandas as pd
	import matplotlib.pyplot as plt
	import cv2
	import tensorflow as tf
	# GPU が使用可能であることを確認
	from tensorflow.python.client import device_lib
	print(device_lib.list_local_devices())
	# データセットの読み込み
	train, test = tf.keras.datasets.cifar10.load_data()
	# 画像の情報
	print("画像情報：",train[0].shape, test[0].shape)
	# ラベルの情報
	print("ラベル情報：",train[1].shape, test[1].shape)
	# 学習用データセットとテスト用データセットに対して正規化
	x_train = train[0] / 255
	x_test = test[0] / 255
	# 目標値の切り分け
	t_train = train[1]
	t_test = test[1]
	# 32bit にキャスト
	x_train, x_test = x_train.astype('float32'), x_test.astype('float32')
	t_train, t_test = t_train.astype('int32'), t_test.astype('int32')
	# CNN モデルの定義
	import os, random
	def reset_seed(seed=0):
	os.environ['PYTHONHASHSEED'] = '0'
	random.seed(seed)
	np.random.seed(seed)
	tf.random.set_seed(seed)
	from tensorflow.keras import models,layers
	# シードの固定
	reset_seed(0)
	# モデルの構築
	model = models.Sequential([
	# 特徴量抽出
	layers.Conv2D(filters=3, kernel_size=(3, 3), activation='relu', input_shape=(32, 32, 3)),
	layers.MaxPool2D(pool_size=(2, 2)),
	# ベクトル化
	layers.Flatten(),
	# 識別
	layers.Dense(100, activation='relu'),
	layers.Dense(10, activation='softmax')
	])
	# パラメータ確認
	print("パラメータ確認：",model.summary())
	# optimizerの設定
	optimizer = tf.keras.optimizers.Adam(lr=0.01)
	# モデルのコンパイル
	model.compile(optimizer=optimizer,
	loss='sparse_categorical_crossentropy',
	metrics=['accuracy'])
	# モデルの学習
	batch_size = 4096
	epochs = 30
	# 学習の実行
	history = model.fit(x_train, t_train,
	batch_size=batch_size,
	epochs=epochs,
	validation_data=(x_test, t_test))
	# ファインチューニング
	from tensorflow.keras.applications import resnet, VGG16
	# 学習済みモデルをインスタンス化
	base_model = VGG16(input_shape=(224, 224, 3),
	include_top=False, weights='imagenet')
	# パラメータ確認
	print("パラメータ確認：",base_model.summary())
	# データセットの準備
	# ランダムにデータを取得する
	train_choice = np.random.randint(low=0, high=50000, size=10000)
	test_choice = np.random.randint(low=0, high=10000, size=5000)
	# データの準備
	x_train = train[0][train_choice]
	x_test = test[0][test_choice]
	t_train = train[1][train_choice].astype('int32')
	t_test = test[1][test_choice].astype('int32')
	_train, _test = [], []
	# 画像サイズを 224 × 224 にリサイズしてリストに格納
	for img in x_train:
	_train.append(cv2.resize(src=img, dsize=(224, 224)))
	for img in x_test:
	_test.append(cv2.resize(src=img, dsize=(224, 224)))
	# リストから ndarray に変換し、正規化
	x_train = np.array(_train, dtype='float32') / 255.0
	x_test = np.array(_test, dtype='float32') / 255.0
	print("サイズの確認：",x_train.shape, x_test.shape)
	# 全結合層の追加
	reset_seed(0)
	# モデルの定義
	finetuned_model = models.Sequential([
	base_model,
	layers.GlobalAveragePooling2D(),
	layers.Dense(512, activation='relu'),
	layers.Dense(10, activation='softmax')
	])
	# 目的関数と最適化手法の選択
	optimizer = tf.keras.optimizers.SGD(lr=0.01)
	# モデルのコンパイル
	finetuned_model.compile(optimizer=optimizer,
	loss='sparse_categorical_crossentropy',
	metrics=['accuracy'])
	# ファインチューニング　モデルの確認
	print("パラメータ確認：",finetuned_model.summary())
	# モデルの学習
	# 順伝播（比較のため）
	loss, accuracy = finetuned_model.evaluate(x_test, t_test)
	# 正解率
	print(f'loss : {loss}, acuracy : {accuracy}')
	# モデルの学習
	history = finetuned_model.fit(x_train, t_train,
	epochs=10,
	batch_size=32,
	validation_data=(x_test, t_test))
	# 予測精度の評価
	results = pd.DataFrame(history.history)
	results.tail(3)
	# 損失を可視化
	results[['loss', 'val_loss']].plot(title='loss')
	plt.xlabel('epochs')
	plt.show()
	# 正解率を可視化
	results[['accuracy', 'val_accuracy']].plot(title='accuracy')
	plt.xlabel('epochs')
	plt.show()

view raw image_recognition_3.py hosted with ❤ by GitHub

出力結果

出力ログ

	[name: "/device:CPU:0"
	device_type: "CPU"
	memory_limit: 268435456
	locality {
	}
	incarnation: 5422697336045806387
	]
	2020-06-01 13:45:54.033667: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2
	画像情報： (50000, 32, 32, 3) (10000, 32, 32, 3)
	ラベル情報： (50000, 1) (10000, 1)
	Model: "sequential"
	_________________________________________________________________
	Layer (type) Output Shape Param #
	=================================================================
	conv2d (Conv2D) (None, 30, 30, 3) 84
	_________________________________________________________________
	max_pooling2d (MaxPooling2D) (None, 15, 15, 3) 0
	_________________________________________________________________
	flatten (Flatten) (None, 675) 0
	_________________________________________________________________
	dense (Dense) (None, 100) 67600
	_________________________________________________________________
	dense_1 (Dense) (None, 10) 1010
	=================================================================
	Total params: 68,694
	Trainable params: 68,694
	Non-trainable params: 0
	_________________________________________________________________
	パラメータ確認： None
	Train on 50000 samples, validate on 10000 samples
	Epoch 1/30
	4096/50000 [=>............................] - ETA: 14s - loss: 2.3118 - accuracy: 0.1062
	8192/50000 [===>..........................] - ETA: 10s - loss: 2.3346 - accuracy: 0.1068
	12288/50000 [======>.......................] - ETA: 8s - loss: 2.2974 - accuracy: 0.1268
	16384/50000 [========>.....................] - ETA: 6s - loss: 2.2829 - accuracy: 0.1295
	20480/50000 [===========>..................] - ETA: 5s - loss: 2.2618 - accuracy: 0.1448
	24576/50000 [=============>................] - ETA: 4s - loss: 2.2438 - accuracy: 0.1610
	28672/50000 [================>.............] - ETA: 3s - loss: 2.2292 - accuracy: 0.1661
	32768/50000 [==================>...........] - ETA: 3s - loss: 2.2130 - accuracy: 0.1773
	36864/50000 [=====================>........] - ETA: 2s - loss: 2.1967 - accuracy: 0.1911
	40960/50000 [=======================>......] - ETA: 1s - loss: 2.1832 - accuracy: 0.2026
	45056/50000 [==========================>...] - ETA: 0s - loss: 2.1667 - accuracy: 0.2116
	49152/50000 [============================>.] - ETA: 0s - loss: 2.1497 - accuracy: 0.2197
	50000/50000 [==============================] - 9s 181us/sample - loss: 2.1462 - accuracy: 0.2209 - val_loss: 1.9381 - val_accuracy: 0.3203
	＜省略＞
	Epoch 30/30
	4096/50000 [=>............................] - ETA: 8s - loss: 0.9545 - accuracy: 0.6646
	8192/50000 [===>..........................] - ETA: 7s - loss: 0.9619 - accuracy: 0.6646
	12288/50000 [======>.......................] - ETA: 6s - loss: 0.9575 - accuracy: 0.6654
	16384/50000 [========>.....................] - ETA: 5s - loss: 0.9641 - accuracy: 0.6627
	20480/50000 [===========>..................] - ETA: 4s - loss: 0.9698 - accuracy: 0.6602
	24576/50000 [=============>................] - ETA: 4s - loss: 0.9693 - accuracy: 0.6608
	28672/50000 [================>.............] - ETA: 3s - loss: 0.9739 - accuracy: 0.6591
	32768/50000 [==================>...........] - ETA: 2s - loss: 0.9754 - accuracy: 0.6598
	36864/50000 [=====================>........] - ETA: 2s - loss: 0.9765 - accuracy: 0.6598
	40960/50000 [=======================>......] - ETA: 1s - loss: 0.9758 - accuracy: 0.6596
	45056/50000 [==========================>...] - ETA: 0s - loss: 0.9786 - accuracy: 0.6584
	49152/50000 [============================>.] - ETA: 0s - loss: 0.9805 - accuracy: 0.6568
	50000/50000 [==============================] - 9s 171us/sample - loss: 0.9810 - accuracy: 0.6566 - val_loss: 1.4533 - val_accuracy: 0.5144
	Model: "vgg16"
	_________________________________________________________________
	Layer (type) Output Shape Param #
	=================================================================
	input_1 (InputLayer) [(None, 224, 224, 3)] 0
	_________________________________________________________________
	block1_conv1 (Conv2D) (None, 224, 224, 64) 1792
	_________________________________________________________________
	block1_conv2 (Conv2D) (None, 224, 224, 64) 36928
	_________________________________________________________________
	block1_pool (MaxPooling2D) (None, 112, 112, 64) 0
	_________________________________________________________________
	block2_conv1 (Conv2D) (None, 112, 112, 128) 73856
	_________________________________________________________________
	block2_conv2 (Conv2D) (None, 112, 112, 128) 147584
	_________________________________________________________________
	block2_pool (MaxPooling2D) (None, 56, 56, 128) 0
	_________________________________________________________________
	block3_conv1 (Conv2D) (None, 56, 56, 256) 295168
	_________________________________________________________________
	block3_conv2 (Conv2D) (None, 56, 56, 256) 590080
	_________________________________________________________________
	block3_conv3 (Conv2D) (None, 56, 56, 256) 590080
	_________________________________________________________________
	block3_pool (MaxPooling2D) (None, 28, 28, 256) 0
	_________________________________________________________________
	block4_conv1 (Conv2D) (None, 28, 28, 512) 1180160
	_________________________________________________________________
	block4_conv2 (Conv2D) (None, 28, 28, 512) 2359808
	_________________________________________________________________
	block4_conv3 (Conv2D) (None, 28, 28, 512) 2359808
	_________________________________________________________________
	block4_pool (MaxPooling2D) (None, 14, 14, 512) 0
	_________________________________________________________________
	block5_conv1 (Conv2D) (None, 14, 14, 512) 2359808
	_________________________________________________________________
	block5_conv2 (Conv2D) (None, 14, 14, 512) 2359808
	_________________________________________________________________
	block5_conv3 (Conv2D) (None, 14, 14, 512) 2359808
	_________________________________________________________________
	block5_pool (MaxPooling2D) (None, 7, 7, 512) 0
	=================================================================
	Total params: 14,714,688
	Trainable params: 14,714,688
	Non-trainable params: 0
	_________________________________________________________________
	パラメータ確認： None
	サイズの確認： (10000, 224, 224, 3) (5000, 224, 224, 3)
	Model: "sequential_1"
	_________________________________________________________________
	Layer (type) Output Shape Param #
	=================================================================
	vgg16 (Model) (None, 7, 7, 512) 14714688
	_________________________________________________________________
	global_average_pooling2d (Gl (None, 512) 0
	_________________________________________________________________
	dense_2 (Dense) (None, 512) 262656
	_________________________________________________________________
	dense_3 (Dense) (None, 10) 5130
	=================================================================
	Total params: 14,982,474
	Trainable params: 14,982,474
	Non-trainable params: 0
	_________________________________________________________________
	パラメータ確認： None
	＜省略＞
	loss : 2.444868545150757, acuracy : 0.09120000153779984
	Train on 10000 samples, validate on 5000 samples
	Epoch 1/10
	＜省略＞
	9920/10000 [============================>.] - ETA: 1:10 - loss: 0.2278 - accuracy: 0.9276
	9952/10000 [============================>.] - ETA: 42s - loss: 0.2279 - accuracy: 0.9276
	9984/10000 [============================>.] - ETA: 14s - loss: 0.2281 - accuracy: 0.9274
	10000/10000 [==============================] - 10428s 1s/sample - loss: 0.2278 - accuracy: 0.9275 - val_loss: 0.4037 - val_accuracy: 0.8640

view raw image_recognition_3.log hosted with ❤ by GitHub