关于Keras猫狗分类教程的疑问：Image Generator如何识别标签？

Hey there! Let's break down how Keras' ImageGenerator figures out those cat and dog labels—it's a clever folder structure trick that makes labeling totally automatic, no manual tagging required.

The Key: Folder Names = Labels #

First, let's recap the folder structure the tutorial uses (this is the critical piece):

training/
├─ cats/
│  ├─ cat001.jpg
│  ├─ cat002.jpg
│  └─ ...
└─ dogs/
   ├─ dog001.jpg
   ├─ dog002.jpg
   └─ ...

The exact same structure applies to the validation/ directory too.

How flow_from_directory Does the Magic #

When you use train_datagen.flow_from_directory() (the method the tutorial uses to load data), here's what happens under the hood:

• It scans the root directory you specify (like 'training').
• It looks for direct subfolders of that root directory—each subfolder name becomes a class label.
• Every image inside the cats/ subfolder gets automatically tagged with the cats label; every image in dogs/ gets the dogs label.

Let's tie this to the code snippet from the tutorial (expanded to show the key part):

# Training data generator setup
train_datagen = ImageDataGenerator(
    rescale=1./255,
    shear_range=0.2,
    zoom_range=0.2,
    horizontal_flip=True)

# Load training data from directory
train_generator = train_datagen.flow_from_directory(
    'training',          # Root directory for training images
    target_size=(150, 150),  # Resize all images to this size
    batch_size=32,
    class_mode='binary'  # Since we have 2 classes (cat/dog)
)

• The class_mode='binary' parameter tells Keras to convert the two labels into a binary format (0 and 1). By default, Keras assigns labels based on alphabetical order of the subfolder names. So cats (comes before dogs alphabetically) becomes 0, and dogs becomes 1.
• You can verify this by printing train_generator.class_indices—it will output {'cats': 0, 'dogs': 1}.

Extra Notes to Keep in Mind #

• No manual labels needed: You don't have to create CSV files or add labels to filenames—just organize your images into subfolders named after their classes.
• Multi-class support: If you were working with more than two classes (e.g., cats, dogs, birds), you'd just add a birds/ subfolder, set class_mode='categorical', and Keras would handle one-hot encoded labels automatically.
• Validation set consistency: The validation directory uses the exact same structure, so the validation generator will recognize labels the same way—this ensures your model evaluates on correctly labeled validation data.

Does that make sense now? If you have any follow-up questions about how this interacts with model training or data augmentation, feel free to ask!

内容的提问来源于stack exchange，提问作者vidit02100