【Python】Kerasでひらがな画像を文字分類。〜 データ準備・転移学習でモデル作成・文字分類 〜

photo
├── a
│   ├── a.jpg
│   ├── a0.jpg
│   ├── a1.jpg
│   ├── a10.jpg
│   ├── a11.jpg
│   ├── a12.jpg
│   ├── a2.jpg
│   ├── a3.jpg
│   ├── a4.jpg
│   ├── a5.jpg
│   ├── a6.jpg
│   ├── a7.jpg
│   ├── a8.jpg
│   └── a9.jpg
└── i
    ├── i.jpg
    ├── i0.jpg
    ├── i1.jpg
    ├── i10.jpg
    ├── i11.jpg
    ├── i12.jpg
    ├── i2.jpg
    ├── i3.jpg
    ├── i4.jpg
    ├── i5.jpg
    ├── i6.jpg
    ├── i7.jpg
    ├── i8.jpg
    └── i9.jpg

import os
import cv2
import numpy
import keras
import tensorflow
import re

from keras.callbacks import EarlyStopping
from keras.applications.inception_v3 import InceptionV3
from keras.callbacks import ModelCheckpoint

# モジュールのimport関連です。今回の転移学習の元学習データはInceptionV3です。

# データ読み込み関数
def create_image_list():
    main_path = "/photo/"
    files = os.listdir(main_path)
    files_dirs = [f for f in files if os.path.isdir(os.path.join(main_path, f))]

    description_value = None
    objective_value = []

    now_index_number = 0

    for dir_item in files_dirs:
        if dir_item == ".ipynb_checkpoints":
            pass
        else:
            print(dir_item + "読み込み中")
            sub_path = main_path + "/" + dir_item
            sub_files = os.listdir(sub_path)
            sub_files_dirs = [f for f in sub_files if os.path.isdir(os.path.join(sub_path, f))]

            sub_files = [f for f in sub_files if re.search('jpg', f, re.IGNORECASE)]
            image_list = []
            loop_count = 0
            if len(sub_files) != 0:
                for image_path_item in sub_files:
                    image_path = sub_path + "/" + image_path_item
                    img = cv2.imread(image_path ,cv2.COLOR_BGR2RGB)
                    #一時措置
                    img = cv2.resize(img, (150,150))
                    image_list.append(img.tolist())


                    index_number = now_index_number
                    objective_value.append(index_number)

                    loop_count += 1
                    if loop_count >= 1500:
                        break

                if description_value is None:
                    description_value = numpy.array(image_list)

                else:
                    now_array = numpy.array(image_list)
                    description_value = numpy.concatenate([description_value,now_array], axis=0)

            image_list = None
            now_index_number += 1

    
    return[description_value,objective_value]

# モデル学習コード
def main():
    description_value,objective_value = create_image_list()
    description_value = description_value.astype('float32') / 255#
    objective_value = numpy.array(objective_value)
    objective_value = keras.utils.to_categorical(objective_value, 2)
    print(description_value.shape[1:])

    model = InceptionV3(include_top=True,weights=None, input_shape=description_value.shape[1:], classes=2)
    print(model.summary())
    model.compile(loss='categorical_crossentropy', optimizer="sgd", metrics=['accuracy'])

    MODEL_DIR = "/test_model/"
    if not os.path.exists(MODEL_DIR):  # ディレクトリが存在しない場合、作成する。
        os.makedirs(MODEL_DIR)

    checkpoint = ModelCheckpoint(
        filepath=os.path.join(MODEL_DIR, "test.hdf5"), save_best_only=True)  # 精度が向上した場合のみ保存する。
    es_cb = EarlyStopping(monitor='val_loss', patience=2, verbose=1, mode='auto')

    history = model.fit(description_value, objective_value, batch_size=64, epochs=150, verbose=1, validation_data=(description_value, objective_value), shuffle=True, callbacks=[checkpoint])
    open(os.path.join("/test_model/",'keras_model.json'), 'w').write(model.to_json())


if __name__ == "__main__":
    main()

from keras.models import load_model
import numpy as np
from keras.preprocessing.image import img_to_array, load_img
model = load_model(r"/test_model/test.hdf5")
print(model.input,model.output)

import glab
img_dir = glob.glob(r"/test_image/*.jpg")
print(img_dir)
#今回は['/test_image/test_img.jpg']test_imageフォルダのtest_img.jpg(あの画像)の一枚をテスト

img1 = load_img(img_dir[0], target_size=(150,150))
img_nad = img_to_array(img1)/255
img_nad = img_nad[None, ...]
pred = model.predict(img_nad, batch_size=1, verbose=0)
print(pred)

[[0.49381995 0.50618005]]