TensorFlow代码TypeError索引无效错误排查及preprocess_image函数数据传入修复方案

Let's break down what's causing this error and how to fix it step by step.

Why the Error Happens #

The TypeError: Only integers, slices (:), ellipsis (...), tf.newaxis (None) and scalar tf.int32/tf.int64 tensors are valid indices, got 'image' comes down to a simple mismatch:

• When you use tf.keras.preprocessing.image_dataset_from_directory, each element in the resulting dataset is a raw image tensor, not a dictionary with an "image" key. But your preprocess_image function is trying to index into data["image"] as if data were a dictionary—this is why the index type is invalid.
• Also, your code references two undefined global variables: padding and image_size—we'll need to define those first too.

Fix 1: Adjust the Preprocessing Function to Match Your Dataset #

If You Don't Need Bounding Box Logic (Simplest Fix) #

If the bounding box cropping code was copied from another project and you don't actually need it for your current task, modify the preprocess_image function to work directly with image tensors:

First, define the missing global variables at the top of your script:

# Define missing global variables (tweak values based on your needs)
padding = 0.2  # Example padding ratio (only needed if you add bbox logic later)
image_size = 64  # Final output image size

Then update the preprocessing function:

def round_to_int(float_value):
    return tf.cast(tf.math.round(float_value), dtype=tf.int32)

def preprocess_image(image):
    # Skip bbox logic since we don't have bbox data from image_dataset_from_directory
    # Resize (optional if you already set image_size in dataset_from_directory)
    image = tf.image.resize(
        image,
        size=[image_size, image_size],
        method=tf.image.ResizeMethod.AREA
    )
    # Normalize pixel values and clamp to the valid [0, 1] range
    return tf.clip_by_value(image / 255.0, 0.0, 1.0)

If You Do Need Bounding Box Data #

If your task requires cropping based on bounding boxes, image_dataset_from_directory won't work—it can't load paired image+bbox data. You'll need to load your data as structured dictionaries instead:

1. Store your image paths and bbox data in a CSV file (e.g., bbox_data.csv) with columns like image_path, ymin, xmin, ymax, xmax.
2. Load the dataset with custom logic:

import pandas as pd

# Load bbox metadata from CSV
df = pd.read_csv("bbox_data.csv")
image_paths = df["image_path"].tolist()
bboxes = df[["ymin", "xmin", "ymax", "xmax"]].values.tolist()

def load_image_and_bbox(image_path, bbox):
    # Load and preprocess the image
    image = tf.io.read_file(image_path)
    image = tf.image.decode_jpeg(image, channels=3)
    image = tf.image.resize(image, (64, 64))
    # Return a dictionary matching what your original preprocess_image expects
    return {"image": image, "bbox": bbox}

# Build the structured dataset
dataset = tf.data.Dataset.from_tensor_slices((image_paths, bboxes))
dataset = dataset.map(load_image_and_bbox, num_parallel_calls=tf.data.AUTOTUNE)
dataset = dataset.batch(20).shuffle(100, seed=123)

# Now your original preprocess_image function will work as intended
train = dataset.map(preprocess_image, num_parallel_calls=tf.data.AUTOTUNE)

Fix 2: Verify the Dataset Pipeline #

After making these changes, run your code—you shouldn't see the index error anymore. For the simple case (no bboxes), your dataset will output normalized images ready for training. For the bbox case, the cropping logic will correctly use the bounding box values to adjust the image.

内容的提问来源于stack exchange，提问作者Ahtisham Ul Haq