使用TensorFlow InceptionV3预处理函数时的错误问题

代码实现参数及数据集加载逻辑 #

# Here my args, they are pretty much the same for all three functions:
training_preprocessing_args = dict(
    labels='inferred',
    label_mode='int',
    class_names=classes,
    color_mode='rgb',
    image_size=hyper_parameter["image_size"],
    shuffle=True,
    seed=seed,
    validation_split=None,
    subset=None,
    interpolation='bilinear',
    follow_links=False,
    crop_to_aspect_ratio=False
)

logging.info("Training Data:")
train_dataset:tf.data.Dataset =  tf.keras.utils.image_dataset_from_directory(directory=PATH_DATA_TRAINING, **training_preprocessing_args)

logging.info("Testing Data:")
test_dataset:tf.data.Dataset =  tf.keras.utils.image_dataset_from_directory(directory=PATH_DATA_TESTING, **testing_preprocessing_args)

logging.info("Validation Data:")
validation_dataset:tf.data.Dataset =  tf.keras.utils.image_dataset_from_directory(directory=PATH_DATA_VALIDATION, **validation_preprocessing_args)

logging.info("Preprocessing:")
train_dataset = tf.keras.applications.inception_v3.preprocess_input(tf.cast(train_dataset, tf.float32))
validation_dataset = tf.keras.applications.inception_v3.preprocess_input(tf.cast(validation_dataset, tf.float32))
test_dataset = tf.keras.applications.inception_v3.preprocess_input(tf.cast(test_dataset, tf.float32))

报错信息 #

15-12-2022 23:21:15 INFO     Training Data:
Found 6988 files belonging to 10 classes.
2022-12-15 23:21:16.075523: I tensorflow/core/platform/cpu_feature_guard.cc:193] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations:  AVX AVX2
To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.
INFO:tensorflow:Converted call: <function paths_and_labels_to_dataset.<locals>.<lambda> at 0x000002063D761FC0>
    args: (<tf.Tensor 'args_0:0' shape=() dtype=string>,)
    kwargs: {}

15-12-2022 23:21:16 INFO     Converted call: <function paths_and_labels_to_dataset.<locals>.<lambda> at 0x000002063D761FC0>
    args: (<tf.Tensor 'args_0:0' shape=() dtype=string>,)
    kwargs: {}

INFO:tensorflow:Allowlisted: <function paths_and_labels_to_dataset.<locals>.<lambda> at 0x000002063D761FC0>: DoNotConvert rule for keras
15-12-2022 23:21:16 INFO     Allowlisted: <function paths_and_labels_to_dataset.<locals>.<lambda> at 0x000002063D761FC0>: DoNotConvert rule for keras
15-12-2022 23:21:16 INFO     Testing Data:
Found 1699 files belonging to 10 classes.
INFO:tensorflow:Converted call: <function paths_and_labels_to_dataset.<locals>.<lambda> at 0x000002063D763490>
    args: (<tf.Tensor 'args_0:0' shape=() dtype=string>,)
    kwargs: {}

15-12-2022 23:21:16 INFO     Converted call: <function paths_and_labels_to_dataset.<locals>.<lambda> at 0x000002063D763490>
    args: (<tf.Tensor 'args_0:0' shape=() dtype=string>,)
    kwargs: {}

INFO:tensorflow:Allowlisted: <function paths_and_labels_to_dataset.<locals>.<lambda> at 0x000002063D763490>: DoNotConvert rule for keras
15-12-2022 23:21:16 INFO     Allowlisted: <function paths_and_labels_to_dataset.<locals>.<lambda> at 0x000002063D763490>: DoNotConvert rule for keras
15-12-2022 23:21:16 INFO     Validation Data:
Found 1700 files belonging to 10 classes.
INFO:tensorflow:Converted call: <function paths_and_labels_to_dataset.<locals>.<lambda> at 0x000002063D761BD0>
    args: (<tf.Tensor 'args_0:0' shape=() dtype=string>,)
    kwargs: {}

15-12-2022 23:21:16 INFO     Converted call: <function paths_and_labels_to_dataset.<locals>.<lambda> at 0x000002063D761BD0>
    args: (<tf.Tensor 'args_0:0' shape=() dtype=string>,)
    kwargs: {}

INFO:tensorflow:Allowlisted: <function paths_and_labels_to_dataset.<locals>.<lambda> at 0x000002063D761BD0>: DoNotConvert rule for keras
15-12-2022 23:21:16 INFO     Allowlisted: <function paths_and_labels_to_dataset.<locals>.<lambda> at 0x000002063D761BD0>: DoNotConvert rule for keras
15-12-2022 23:21:16 INFO     Preprocessing:
Traceback (most recent call last):
  File "_CORE\main.py", line 27, in <module>
    main()
  File "_CORE\main.py", line 17, in main
    data:tuple = run_preprocessing()
  File "_CORE\preprocessing\run.py", line 10, in run_preprocessing
    data = create_datasets()
  File "_CORE\preprocessing\CreateDataset.py", line 23, in create_datasets
    train_dataset = tf.keras.applications.inception_v3.preprocess_input(train_dataset)#tf.cast(train_dataset, tf.float32))
  File "_ENV\_ENV_1\lib\site-packages\keras\applications\inception_v3.py", line 448, in preprocess_input
    return imagenet_utils.preprocess_input(
  File "_ENV\_ENV_1\lib\site-packages\keras\applications\imagenet_utils.py", line 123, in preprocess_input
    return _preprocess_symbolic_input(x, data_format=data_format, mode=mode)
  File "_ENV\_ENV_1\lib\site-packages\keras\applications\imagenet_utils.py", line 271, in _preprocess_symbolic_input
    x /= 127.5
TypeError: unsupported operand type(s) for /=: 'BatchDataset' and 'float'

参考官方示例 #

i = tf.keras.layers.Input([None, None, 3], dtype = tf.uint8)
x = tf.cast(i, tf.float32)
x = tf.keras.applications.mobilenet.preprocess_input(x)
core = tf.keras.applications.MobileNet()
x = core(x)
model = tf.keras.Model(inputs=[i], outputs=[x])

image = tf.image.decode_png(tf.io.read_file('file.png'))
result = model(image)

问题分析与解决方法 #

错误原因 #

你直接将BatchDataset对象传给了preprocess_input函数，但该函数的处理对象是单个张量（Tensor），而非整个数据集。tf.keras.utils.image_dataset_from_directory返回的是包含（图像张量，标签张量）的批次数据集，无法直接作为预处理函数的参数。

解决方法 #

方法1：用map函数处理数据集

针对数据集中的每个元素（图像+标签）单独执行预处理：

def preprocess_fn(image, label):
    # 转换图像张量类型
    image = tf.cast(image, tf.float32)
    # 应用InceptionV3的预处理逻辑
    image = tf.keras.applications.inception_v3.preprocess_input(image)
    return image, label

# 对三个数据集分别应用预处理
train_dataset = train_dataset.map(preprocess_fn)
validation_dataset = validation_dataset.map(preprocess_fn)
test_dataset = test_dataset.map(preprocess_fn)

方法2：将预处理整合到模型中（适配官方示例）

把预处理逻辑作为模型的输入层部分，数据传入模型时自动完成预处理：

# 假设你的输入尺寸是hyper_parameter["image_size"]，例如(299,299)
input_layer = tf.keras.layers.Input(shape=(*hyper_parameter["image_size"], 3), dtype=tf.uint8)
x = tf.cast(input_layer, tf.float32)
x = tf.keras.applications.inception_v3.preprocess_input(x)
# 加载预训练的InceptionV3（根据需求设置参数）
base_model = tf.keras.applications.InceptionV3(weights='imagenet', include_top=False, input_tensor=x)
# 添加自定义分类层（示例）
x = base_model.output
x = tf.keras.layers.GlobalAveragePooling2D()(x)
output_layer = tf.keras.layers.Dense(len(classes), activation='softmax')(x)

model = tf.keras.Model(inputs=input_layer, outputs=output_layer)

这种方式下，你可以直接将image_dataset_from_directory返回的原始数据集传入模型训练，无需提前处理。

关于尺寸调整 #

你设置的image_size=hyper_parameter["image_size"]已经让image_dataset_from_directory自动将所有图像调整到指定尺寸，不需要额外处理。预处理函数仅负责像素值归一化、缩放等操作，不承担尺寸调整的工作。

内容的提问来源于stack exchange，提问作者JustSomeCoder