tensorflowで美術作品の画像データから芸術様式を予測するモデルの作成

import pandas as pd
import tensorflow as tf

df = pd.read_csv('./all_data_info.csv.zip')

datagen = tf.keras.preprocessing.image.ImageDataGenerator(
    rescale=1./255,
    rotation_range=40,
    zoom_range=[0.75, 1.25], # ランダムにzoomをしてみる
    horizontal_flip=True,
    vertical_flip=True,
)

IMG_HEIGHT, IMG_WIDTH = 224, 224
BATCH_SIZE = 48

train_generator = datagen.flow_from_dataframe(
    dataframe=df[df['in_train']==True],
    directory='./train',                        
    x_col='new_filename',                     
    y_col='style',                     
    batch_size=BATCH_SIZE,
    target_size=(IMG_HEIGHT, IMG_WIDTH)
)

test_generator = datagen.flow_from_dataframe(
    dataframe=df[df['in_train']==False],
    directory='./test',                 
    x_col='new_filename',                     
    y_col='style',                     
    batch_size=BATCH_SIZE,
    target_size=(IMG_HEIGHT, IMG_WIDTH)                        
)

# model init
classes = list(df['style'])

input_tensor = tf.keras.layers.Input(shape=(IMG_HEIGHT, IMG_WIDTH, 3))

with tf.device('/device:XLA_GPU:0'):

    vgg19 = tf.keras.applications.VGG19(
        weights='imagenet',
        include_top=False,
        input_shape=(IMG_HEIGHT, IMG_WIDTH, 3),
        input_tensor=input_tensor
    )
    
    # add some layer for output
    x = vgg19.output
    x = tf.keras.layers.Flatten()(x)

    x = tf.keras.layers.Dense(724, activation='relu')(x)
    x = tf.keras.layers.Dense(512, activation='relu')(x)

    predictions = tf.keras.layers.Dense(len(classes),
                                        activation='softmax')(x)
    
    model = tf.keras.Model(inputs=vgg19.input, outputs=predictions)
    
    # 20層までfreeze
    for layer in model.layers[:22]:
        layer.trainable = False

    # 20層以降、学習させる
    for layer in model.layers[22:]:
        layer.trainable = True

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

dpath = './models'
fpath = os.path.join(dpath, 'weights.{epoch:02d}-{loss:.2f}-{accuracy:.2f}-{val_loss:.2f}-{val_accuracy:.2f}.hdf5')

cp_cb = tf.keras.callbacks.ModelCheckpoint(filepath=fpath, monitor='val_loss',
  verbose=1, save_best_only=True, mode='auto',
  save_freq='epoch'
)

# notify to slack
from tf_notification_callback import SlackCallback
slack_cb = SlackCallback(webhookURL='https://hooks.slack.com/services/Txxxxx/Bxxxx/bxxxxxxxxxxx',
                         channel='tf-nortify',
                         modelName='vgg19 Model',
                         loss_metrics=['loss', 'val_loss'],
                         acc_metrics=['accuracy', 'val_accuracy'],
                         getSummary=True)

N_TRAIN = len(df[df['in_train']==True])
N_TEST = len(df[df['in_train']==False])
N_EPOCHS = 20

history = model.fit(
    train_generator,
    steps_per_epoch=N_TRAIN // BATCH_SIZE,
    epochs=N_EPOCHS,
    validation_data=test_generator,
    validation_steps=N_TEST // BATCH_SIZE,
    callbacks=[cp_cb, slack_cb],
)

...
1066/1066 [==============================] - 10213s 10s/step - loss: 1.5340 - accuracy: 0.4436 - val_loss: 1.6417 - val_accuracy: 0.4040