基于Keras的多标签分类任务中JSON文件处理方法咨询

Got it, let's break down exactly how to convert your JSON multi-label image data into the binary vector format that works with Keras—just like the tutorial you referenced. Here's a step-by-step solution tailored to your setup:

First, we'll read in your JSON file and restructure it into a usable format. Note: Make sure your JSON has unique image filenames as keys (the example you shared has duplicate "PIC_NAME" keys, which standard JSON parsers will overwrite—fix that first if it's an actual issue in your data).

import json
import numpy as np
import pandas as pd
from sklearn.preprocessing import MultiLabelBinarizer

# Load the JSON label file
with open('your_label_data.json', 'r') as f:
    label_data = json.load(f)

# Convert to a list of (image_filename, label_list) tuples for easy processing
image_label_pairs = [(img_filename, labels) for img_filename, labels in label_data.items()]

We'll use MultiLabelBinarizer (from scikit-learn) to handle the conversion from string labels to binary vectors. First, we'll extract all unique labels from your dataset to define our label set.

# Collect all unique labels across all images
all_labels = [label for _, label_list in image_label_pairs for label in label_list]
unique_labels = sorted(list(set(all_labels)))
print(f"Identified {len(unique_labels)} unique labels: {unique_labels}")

# Initialize the binarizer with our fixed label set (ensures consistent indices every time)
mlb = MultiLabelBinarizer(classes=unique_labels)
mlb.fit([unique_labels])

If you actually need a 23-dimensional vector (like the tutorial mentioned, even though you have 20 labels), just add 3 placeholder labels to the unique_labels list before fitting the binarizer.

Now we'll transform each image's label list into the binary format Keras expects. For example, if an image has labels ["Label2", "Label6"], the vector will have 1s at the indices corresponding to Label2 and Label6, and 0s everywhere else.

# Convert all label lists to binary vectors
binary_label_matrix = mlb.transform([label_list for _, label_list in image_label_pairs])

# Example verification to make sure it works
sample_idx = 0
print(f"\nSample Image: {image_label_pairs[sample_idx][0]}")
print(f"Original Labels: {image_label_pairs[sample_idx][1]}")
print(f"Binary Label Vector: {binary_label_matrix[sample_idx]}")

To match the tutorial's workflow, we can convert this into a CSV file (like the tutorials use) or keep it as a DataFrame for direct use with Keras' ImageDataGenerator.

Option A: Save to CSV (for flow_from_dataframe) #

# Create a DataFrame with image paths and expanded binary label columns
df = pd.DataFrame({
    'image_path': [img_filename for img_filename, _ in image_label_pairs]
})

# Add each label as a separate column (0/1 values)
for label_idx, label in enumerate(unique_labels):
    df[label] = binary_label_matrix[:, label_idx]

# Save to CSV for easy use in Keras
df.to_csv('image_labels_binary.csv', index=False)

Option B: Use directly in Keras #

If you don't want a CSV, you can use the binary_label_matrix directly with a custom data loader, or pair it with image paths for flow_from_dataframe:

from tensorflow.keras.preprocessing.image import ImageDataGenerator

# Initialize image data generator (add augmentations as needed)
datagen = ImageDataGenerator(rescale=1./255)

# Create a generator that loads images and their binary labels
train_generator = datagen.flow_from_dataframe(
    dataframe=df,
    directory='path/to/your/image/folder',  # Path where your images are stored
    x_col='image_path',
    y_col=unique_labels,  # Use the label columns we created
    target_size=(224, 224),  # Adjust to your model's input size
    batch_size=32,
    class_mode='raw'  # Critical for multi-label binary classification
)

• Ensure your image filenames in the JSON match the actual filenames in your image directory (including file extensions like .jpg).
• The class_mode='raw' tells Keras we're passing binary vectors for multi-label classification, instead of single-class labels.
• If you need to map back from binary vectors to string labels later, use mlb.inverse_transform(binary_label_matrix) to convert vectors back to label lists.

内容的提问来源于stack exchange，提问作者glomba