嘗試利用 Keras 的 VGG16 預訓練時出現無效形狀錯誤 model

Question

我正在嘗試在我自己的圖像分類問題中利用 kera 的VGG16 model 。 我的代碼主要基於 Francois Chollet 的示例（Python 中深度學習的第 8 章 - 代碼）。

我要預測三個班級。 目錄結構：

data/
  training/
    class_1
    class_2
    class_3

注意：這是我第一次使用 Keras，所以我可能只是做錯了什么。

我對model.fit()的調用失敗並顯示： ValueError: Shapes (32, 1) and (32, 3) are incompatible 。 有關完整的錯誤消息，請參閱此問題的底部。 如果我從 .summary .summary()調用中查看 output，我看不到一層維度 (32, 1)。

import pathlib
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
from tensorflow.keras.utils import image_dataset_from_directory

DATA_DIR = pathlib.Path('./data/')
batch_size = 32
img_width = image_height = 256

train_dataset = image_dataset_from_directory(
    DATA_DIR / "training",
    image_size=img_width_height,
    batch_size=batch_size)

validation_dataset = image_dataset_from_directory(
    DATA_DIR / "validation",
    image_size=img_width_height,
    batch_size=batch_size)

# Found 128400 files belonging to 3 classes.
# Found 15600 files belonging to 3 classes.

vgg16_convolution_base = keras.applications.vgg16.VGG16(
    weights="imagenet",
    include_top=False,
    input_shape=(img_width, image_height, 3))

vgg16_convolution_base.summary()
# block3_conv3 (Conv2D)       (None, 64, 64, 256)       590080    
# block3_pool (MaxPooling2D)  (None, 32, 32, 256)       0         
# block4_conv1 (Conv2D)       (None, 32, 32, 512)       1180160   
# block4_conv2 (Conv2D)       (None, 32, 32, 512)       2359808   
# block4_conv3 (Conv2D)       (None, 32, 32, 512)       2359808   
# block4_pool (MaxPooling2D)  (None, 16, 16, 512)       0         
# block5_conv1 (Conv2D)       (None, 16, 16, 512)       2359808   
# block5_conv2 (Conv2D)       (None, 16, 16, 512)       2359808   
# block5_conv3 (Conv2D)       (None, 16, 16, 512)       2359808   
# block5_pool (MaxPooling2D)  (None, 8, 8, 512)         0

def get_features_and_labels(dataset):
    all_features = []
    all_labels = []
    for images, labels in dataset:
        preprocessed_images = keras.applications.vgg16.preprocess_input(images)
        features = vgg16_convolution_base.predict(preprocessed_images)
        all_features.append(features)
        all_labels.append(labels)
    return np.concatenate(all_features), np.concatenate(all_labels)

train_features, train_labels = get_features_and_labels(train_dataset)
val_features, val_labels = get_features_and_labels(validation_dataset)

print(train_features.shape)
print(train_labels.shape)
# (128400, 8, 8, 512)
# (128400,)

print(val_features.shape)
print(val_labels.shape)
# (15600, 8, 8, 512)
# (15600,)

inputs = keras.Input(shape=(8, 8, 512))

x = layers.Flatten()(inputs)
x = layers.Dense(256)(x)
x = layers.Dropout(0.5)(x)

outputs = layers.Dense(3, activation="softmax")(x)

model = keras.Model(inputs, outputs)

model.compile(loss="categorical_crossentropy",
              optimizer="rmsprop",
              metrics=["accuracy"])

model.summary()
# input_4 (InputLayer)        [(None, 8, 8, 512)]       0         
# flatten_1 (Flatten)         (None, 32768)             0         
# dense_2 (Dense)             (None, 256)               8388864   
# dropout_1 (Dropout)         (None, 256)               0         
# dense_3 (Dense)             (None, 3)                 771       
# ================================================================
# Total params: 8,389,635
# Trainable params: 8,389,635

history = model.fit(
    train_features, train_labels,
    epochs=20,
    validation_data=(val_features, val_labels)

我對model.fit()的調用失敗並顯示： ValueError: Shapes (32, 1) and (32, 3) are incompatible

...
File "C:\Users\x\anaconda3\lib\site-packages\keras\losses.py", line 1990, in categorical_crossentropy
        return backend.categorical_crossentropy(
    File "C:\Users\x\anaconda3\lib\site-packages\keras\backend.py", line 5529, in categorical_crossentropy
        target.shape.assert_is_compatible_with(output.shape)

完整追溯

Answer 1

3 個類的categorical_crossentropy損失和 32 的批大小決定了標簽的形狀（對於每個 bach）為 (32, 3)。

這些標簽目前是有序的： 0 2 1 可以對序號標簽使用SparseCategoricalCrossentropy損失：

loss= tf.keras.losses.SparseCategoricalCrossentropy()

或者，仍然可以使用categorical_crossentropy損失，但結合 one-hot 編碼標簽(1, 0, 0) for 0 ， (0, 1, 0) for 1和(0, 0, 1) for 2 . 下面的代碼片段可以完成這樣的編碼：

#one-hot encoding
num_class = len(set(train_labels))
train_labels=tf.one_hot(indices=train_labels, depth=num_class)
val_labels=tf.one_hot(indices=val_labels, depth=num_class)

數據的性質（有序或無序）有助於確定單熱編碼是首選還是有序編碼。