train model using tf.data.Dataset of tf.keras.preprocessing.image_dataset_from_directory is very slow keras

Question

I am using tf.keras.preprocessing.image_dataset_from_directory to load my large dataset. the problem is the training phase is so slow when I exploit this method in fit_generator() although I use the google Colab GPU. The code is:

image_size = (224, 224)
batch_size = 32

data = tf.keras.preprocessing.image_dataset_from_directory(
     '/content/drive/My Drive/dataScience/september exam/data/trainImg',
    seed=1337,
    image_size=image_size,
    batch_size=batch_size,
)

for the training:

model.fit_generator(train_dataset,
             epochs=50,
             verbose=1)

Answer 1

您可以尝试将图像形状缩小到 128x128，减少batch_size 并使用model.fit()的GPU，您应该使用model.fit()希望这有助于您进行时间优化。