InterviewsVector

How to handle class weight for multiple outputs in keras?

TensorFlow & GPUskeras

Quick answer

Keras does not accept class_weight for models with multiple outputs — it raises 'class_weight is only supported for Models with a single output'. The workaround is sample_weight: turn each output's class weights into a per-sample weight array, then pass them to model.fit as a dict keyed by output name, e.g. sample_weight={'out1': sw1, 'out2': sw2}. Alternatively, bake the class weighting into a separate custom loss per output using tf.gather.

Short answer: Keras rejects class_weight on a multi-output model — it raises "class_weight is only supported for Models with a single output." Convert each output's class weights into a per-sample weight array, then pass them to fit as a dict keyed by output name: sample_weight={"out1": sw1, "out2": sw2}. Or bake the weighting into a separate custom loss per output.

class_weight maps one {class: weight} dict to one set of labels. With several outputs, Keras can't tell which output the dict applies to, so it errors. The two reliable workarounds are below.

Option 1 — sample_weight as a per-output dict (recommended)

Turn each output's class weights into an array with one weight per training sample, then hand fit a dict keyed by the output layer names. This is the direct equivalent of class_weight, and it's the approach the Keras docs point you to.

import numpy as np
 
# class weights per output (from the class imbalance of each label set)
cw_out1 = {0: 1.0, 1: 5.0}            # binary, positive class is rare
cw_out2 = {0: 1.0, 1: 2.0, 2: 3.0}    # 3-class
 
# map each label to its weight -> one weight per sample, per output
sw_out1 = np.array([cw_out1[y] for y in y1_train])
sw_out2 = np.array([cw_out2[y] for y in y2_train])
 
model.compile(
    optimizer="adam",
    loss={"out1": "sparse_categorical_crossentropy",
          "out2": "sparse_categorical_crossentropy"},
)
 
model.fit(
    x_train,
    {"out1": y1_train, "out2": y2_train},
    sample_weight={"out1": sw_out1, "out2": sw_out2},
    epochs=10,
    batch_size=32,
)

The keys ("out1", "out2") must match the name= of each output layer. Note the per-output weight arrays are built independently — the earlier mistake is weighting only one output and passing a single array.

Option 2 — a weighted loss per output

If you'd rather keep the weighting inside the loss, give each output its own loss. Keras calls a per-output loss with that output's y_true/y_pred (not a list of all of them), so tf.gather the right weight for each label:

import tensorflow as tf
 
def make_weighted_loss(class_weights):
    cw = tf.constant(class_weights, dtype=tf.float32)   # shape [num_classes]
 
    def loss(y_true, y_pred):
        per_example = tf.keras.losses.sparse_categorical_crossentropy(y_true, y_pred)
        idx = tf.cast(tf.reshape(y_true, [-1]), tf.int32)
        weights = tf.gather(cw, idx)                     # weight per sample
        return tf.reduce_mean(per_example * weights)
 
    return loss
 
model.compile(
    optimizer="adam",
    loss={"out1": make_weighted_loss([1.0, 5.0]),
          "out2": make_weighted_loss([1.0, 2.0, 3.0])},
)
model.fit(x_train, {"out1": y1_train, "out2": y2_train}, epochs=10, batch_size=32)

Which to use

  • sample_weight dict — simplest, no custom code, mirrors class_weight semantics exactly. Prefer this.
  • Custom per-output loss — when you need more than class weighting (e.g. focal loss, or combining the weight with a per-sample importance).

Balance the outputs against each other with loss_weights={"out1": 1.0, "out2": 0.5} in compile — that's a separate knob from the class weighting above.

Related guides

Sources

Key takeaways

  • •Keras rejects class_weight for multi-output models and raises an error.
  • •The workaround is sample_weight: turn your class weights into a per-sample weight array for each output.
  • •Pass the sample weights as a dict keyed by output name to model.fit(sample_weight=...).
  • •Alternatively, bake the class weighting directly into a custom loss for each output.

Frequently asked questions

Why does class_weight fail with multiple outputs in Keras?

Keras only supports class_weight for single-output models; with multiple outputs it cannot map one class-weight dict to several label sets, so it errors.

How do I weight classes per output?

Convert each output's class weights into a per-sample weight array and pass them to model.fit as a sample_weight dict keyed by the output layer names.

Is a custom loss an option?

Yes. You can multiply each output's loss by its class weights inside a custom loss function, which gives full control over how each output is weighted.

By Mohammad Wasi

Software Engineering Leader & Technical Author · Updated September 9, 2026

See all posts

Related Posts