Keras张量二值化问题：自定义损失函数报ValueError错误

Got it, let's break down what's going wrong here and how to fix your binarization logic.

Why You're Seeing the Error #

The root issue is that you're probably using regular Python conditional logic (like if/else or list comprehensions) to binarize your tensor. In Keras/TensorFlow, tensors are symbolic (they represent computations, not actual values) during graph construction. Python's native conditionals can't handle these symbolic tensors—they'll either return None or break the computation graph entirely, which triggers that "None values not supported" error.

Your test with just K.flatten(y_pred) works because that's a Keras backend function designed to operate on symbolic tensors, so it plays nicely with the computation graph.

Correct Ways to Binarize Tensors in Loss Functions #

You need to use Keras/TensorFlow backend functions to handle the binarization—these are built to work with symbolic tensors and maintain the computation graph. Here are a few reliable approaches:

1. Use K.round() (Simplest for 0.5 Threshold)

Since you want values >0.5 to become 1.0 and others to be 0.0, K.round() does exactly this out of the box:

from tensorflow.keras import backend as K

def binary_custom_loss(y_true, y_pred):
    # Flatten both tensors to match shapes
    y_true_flat = K.flatten(y_true)
    y_pred_flat = K.flatten(y_pred)
    
    # Binarize predictions: >0.5 → 1.0, ≤0.5 → 0.0
    y_pred_binarized = K.round(y_pred_flat)
    
    # Replace with your preferred loss calculation (e.g., MSE, MAE)
    return K.mean(K.square(y_true_flat - y_pred_binarized))

2. Explicit Threshold with K.greater() + K.cast()

If you want more control over the threshold (or just want to be explicit), use these functions to create a boolean mask and cast it to floats:

from tensorflow.keras import backend as K

def binary_custom_loss(y_true, y_pred):
    y_true_flat = K.flatten(y_true)
    y_pred_flat = K.flatten(y_pred)
    
    # Create a boolean tensor where elements are >0.5
    above_threshold = K.greater(y_pred_flat, 0.5)
    # Convert booleans to floats: True → 1.0, False → 0.0
    y_pred_binarized = K.cast(above_threshold, K.floatx())
    
    return K.mean(K.square(y_true_flat - y_pred_binarized))

3. Use K.where() for Conditional Assignment

Another option is K.where(), which lets you define values for elements that meet or don't meet a condition:

from tensorflow.keras import backend as K

def binary_custom_loss(y_true, y_pred):
    y_true_flat = K.flatten(y_true)
    y_pred_flat = K.flatten(y_pred)
    
    # Assign 1.0 to values >0.5, else 0.0
    y_pred_binarized = K.where(
        y_pred_flat > 0.5,
        K.ones_like(y_pred_flat),  # Value if condition is True
        K.zeros_like(y_pred_flat)  # Value if condition is False
    )
    
    return K.mean(K.square(y_true_flat - y_pred_binarized))

Key Notes #

• Always use Keras/TensorFlow backend functions when manipulating tensors in loss functions, metrics, or custom layers—never rely on Python's native loops or conditionals for symbolic tensor operations.
• If you're working on a binary classification task, consider if you really need to binarize predictions in the loss function. For example, BinaryCrossentropy works directly with probability outputs (no need to binarize) and is more numerically stable for classification tasks. But if your use case specifically requires binarized predictions in the loss, the methods above will work.

内容的提问来源于stack exchange，提问作者FiReTiTi