使用tf.train.example遇TypeError：预期bytes却传入int值求助

Hey there, let's break down this issue you're facing. Even though you think you're passing a bytes type, the error clearly says the value 71 is an int when bytes were expected. Looking at your traceback, the problem happens in your _bytes_feature function when creating tf.train.BytesList(value=value)—so let's dig into why that's happening.

Common Causes & Fixes #

1. Your image data isn't actually a bytes object #

If img_data is a numpy array (like what you get from converting a PIL image), passing it directly to _bytes_feature will treat it as a sequence of integers (each pixel value is an int). You need to explicitly convert it to bytes:

# Assuming img_data is a numpy array
img_data = img_data.tobytes()  # Converts the array to a raw bytes object

2. Your _bytes_feature function isn't handling single bytes correctly #

The tf.train.BytesList expects an iterable of bytes objects. If you're passing a single bytes value instead of a list containing that value, it might iterate over the bytes as individual integers (since bytes are iterable in Python). Fix your function to wrap single values in a list:

def _bytes_feature(value):
    # Handle EagerTensor if needed
    if isinstance(value, type(tf.constant(0))):
        value = value.numpy()
    # Wrap single bytes in a list for BytesList
    return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))

3. You're reading the image in text mode instead of binary mode #

If you're loading your image with open("image.jpg", "r"), you're reading it as text, which can mangle the data and lead to unexpected int values. Always use binary mode for images:

with open("your_image.jpg", "rb") as f:
    img_data = f.read()  # This gives you a proper bytes object directly

Quick Debug Check #

Before calling _bytes_feature(img_data), add these lines to verify what you're actually passing:

print(f"Type of img_data: {type(img_data)}")
print(f"First 10 elements: {img_data[:10]}")

• If it's a bytes object, you'll see output like b'\xff\xd8\xff\xe0\x00\x10JFIF'
• If it's a numpy array or list, you'll see a sequence of integers like [71, 72, 69, ...]

That should help you pinpoint exactly where the int values are sneaking in instead of bytes.

内容的提问来源于stack exchange，提问作者Tris Wang