Keras损失函数为何返回张量而非单个数值？

Great question! This is a super common point of confusion when first digging into Keras loss logic—let’s break down exactly what’s going on here.

First, let’s revisit the MSE code you shared:

def mean_squared_error(y_true, y_pred):
    return K.mean(K.square(y_pred - y_true), axis=-1)

When you feed in tensors of shape (a, b, c, d), the axis=-1 argument tells Keras to calculate the mean only along the last dimension (the d dimension). That’s why you end up with an output tensor of shape (a, b, c) instead of a single number. But why is this the default behavior?

1. Flexibility for Diverse Tasks #

Keras builds loss functions this way to support a wide range of use cases:

• Spatial/sequence tasks: For image segmentation or time-series models, you might want to track loss per pixel or per time step (instead of collapsing everything to a single value right away). This lets you apply custom weights to specific positions (like masking padding in sequences) or analyze model performance across different parts of your input.
• Weighted per-sample loss: If you need to assign different importance weights to individual samples in your batch, having per-sample (or per-position) loss tensors makes it easy to multiply by a weight tensor before computing the final scalar loss.

2. Keras Handles the Final Reduction Automatically #

Don’t stress—you don’t have to manually shrink the tensor to a scalar for training! When you compile your model with loss='mean_squared_error', Keras automatically runs a final reduction step to turn the tensor into a single scalar loss value. By default, it takes the mean of all values in the loss tensor (you can customize this with the loss_reduction parameter in newer Keras versions if needed).

For example, if your loss tensor is shape (a, b, c), Keras will compute the average of every element in that tensor to get the single loss number you see printed during training epochs.

3. How to Get a Scalar Loss Manually #

If you ever want to compute a scalar loss yourself (for debugging or custom logic), just take the mean over all remaining dimensions:

scalar_loss = K.mean(mean_squared_error(y_true, y_pred))

This will collapse the (a, b, c) tensor down to a single numerical value.

内容的提问来源于stack exchange，提问作者Vivek Pandey