如何适应tensorflow ImageDataGenerator

Question

我已经建立了我的模型，但不知道如何拟合它。有没有人可以给我一些建议，这样我就可以在处理图像时在我的模型中使用ImageDataGenerator，或者最好使用其他方法，比如使用Dataset

import tensorflow as tf
from tensorflow.keras.layers import Dense, Conv2D, Flatten, Dropout, MaxPooling2D
import numpy as np
import matplotlib.pyplot as plt
import cv2
import os

# const
IMG_HEIGHT = 150
IMG_WIDTH  = 150
BATCH = 32
EPOCHS = 5
train_dir = "data/images/train"
val_dir = "data/images/val"


# train image data generator
train_generator = tf.keras.preprocessing.image.ImageDataGenerator(
    rescale=1./255,
    horizontal_flip=True,
    dtype=tf.float32
)
train_generator.flow_from_directory(
    directory=train_dir,
    target_size=(IMG_WIDTH, IMG_HEIGHT)
)

# validation image data generator
val_generator = tf.keras.preprocessing.image.ImageDataGenerator(
    rescale=1./255,
    horizontal_flip=False
)
val_generator.flow_from_directory(
    directory = val_dir,
    target_size=(IMG_WIDTH, IMG_HEIGHT)
)

# count train cats & dogs
train_cats_len = len(os.listdir(os.path.join(train_dir, "cats")))
train_dogs_len = len(os.listdir(os.path.join(train_dir, "dogs")))
train_len = train_cats_len + train_dogs_len

# count validation cats & dogs
val_cats_len = len(os.listdir(os.path.join(val_dir, "cats")))
val_dogs_len = len(os.listdir(os.path.join(val_dir, "dogs")))
val_len = val_cats_len + val_dogs_len

# build a model
model = tf.keras.Sequential([
    Conv2D(16, 3, padding='same', activation='relu', input_shape=(IMG_HEIGHT, IMG_WIDTH , 3)),
    MaxPooling2D(),
    Dropout(0.2),
    Flatten(),
    Dense(128, activation='relu'),
    Dense(64, activation='relu'),
    Dense(2, activation='sigmoid')
])

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

# fit?

# history = model.fit_generator(
#     train_generator,
#     steps_per_epoch=train_len // BATCH,
#     epochs=EPOCHS,
#     validation_data=val_generator,
#     validation_steps=val_len // BATCH,
#     verbose=True
# )

# raises error:
# ValueError: Failed to find data adapter that can handle input: <class 'tensorflow.python.keras.preprocessing.image.ImageDataGenerator'>, <class 'NoneType'>

我的目录架构：

data-
    |-images-
            |-train-
                   |-cats
                   |-dogs
            |-val-
                  |-cats
                  |-dogs

PS：

我发现article使用了相同的方法，一切似乎都正常，但在我的例子中并非如此

Answer 1

你的问题是你有代码

train_generator.flow_from_directory(
    directory=train_dir,
    target_size=(IMG_WIDTH, IMG_HEIGHT)

您需要将其更改为

train_generator=train_generator.flow_from_directory( directory=train_dir,
    target_size=(IMG_WIDTH, IMG_HEIGHT)

对val_generator执行相同的操作。此外，ImageDataGenerator的默认class_mode是“分类的”。因此，在model.compile中，您应该将损失指定为'categorical_crossentropy‘。在包含2个节点的模型层中，激活函数应该是'softmax‘。顺便说一句，我认为你的模型可能表现不是很好，因为处理数据的特征可能有点简单。我建议添加更多的卷积层和更多的滤波器。下面显示了一个更复杂模型的示例

model = tf.keras.Sequential([
    Conv2D(16, 3, padding='same', activation='relu', input_shape=(IMG_HEIGHT, IMG_WIDTH , 3)),
    MaxPooling2D(),
    Conv2D(32, 3, padding='same', activation='relu', input_shape=(IMG_HEIGHT, IMG_WIDTH , 3)),
    MaxPooling2D(),
    Conv2D(64, 3, padding='same', activation='relu', input_shape=(IMG_HEIGHT, IMG_WIDTH , 3)),
    MaxPooling2D(),
    Conv2D(128, 3, padding='same', activation='relu', input_shape=(IMG_HEIGHT, IMG_WIDTH , 3)),
    MaxPooling2D(),
    Flatten(),
    Dense(128, activation='relu'),
    Dropout(.3),
    Dense(64, activation='relu'),
    Dropout(.3),
    Dense(2, activation='softmax')
])

Answer 2

history = model.fit(train_generator,
                    validation_data=validation_generator,
                    steps_per_epoch=100,
                    epochs=15,
                    validation_steps=50,
                    verbose=2)

您可以按照colab上的示例进行操作