Introduction

This Keras error appears when you try to build a Functional API Model with outputs that are not part of the symbolic layer graph created from keras.Input. In plain terms, the outputs= argument must come from calling Keras layers or models on the input graph, not from unrelated tensors, NumPy values, or standalone TensorFlow objects.

What Keras Expects

In the Functional API, the pattern is:

1. create an input tensor with keras.Input
2. pass that tensor through layers
3. use the resulting symbolic tensor as the model output

Correct example:

1import tensorflow as tf
2
3inputs = tf.keras.Input(shape=(8,))
4x = tf.keras.layers.Dense(16, activation="relu")(inputs)
5outputs = tf.keras.layers.Dense(1)(x)
6
7model = tf.keras.Model(inputs=inputs, outputs=outputs)
8model.summary()

Here, outputs comes from layer calls that are connected to inputs, so Keras can build the graph.

A Typical Failing Pattern

This error often happens when the output is not connected to the functional graph.

Bad example:

1import tensorflow as tf
2
3inputs = tf.keras.Input(shape=(8,))
4x = tf.keras.layers.Dense(16, activation="relu")(inputs)
5outputs = tf.constant([1.0, 2.0, 3.0])
6
7model = tf.keras.Model(inputs=inputs, outputs=outputs)

outputs in this case is just a regular TensorFlow constant. It is not the output of a Keras layer and has no symbolic relationship to the input graph, so the model constructor rejects it.

Another Common Cause: Leaving the Keras Graph

You can also trigger the error by mixing Keras tensors with non-Keras operations in the wrong way. The fix is usually to keep custom computation inside a Keras layer, a Lambda layer, or a nested model.

Correct pattern with Lambda:

1import tensorflow as tf
2
3inputs = tf.keras.Input(shape=(8,))
4x = tf.keras.layers.Dense(16, activation="relu")(inputs)
5outputs = tf.keras.layers.Lambda(lambda t: tf.math.square(t))(x)
6
7model = tf.keras.Model(inputs=inputs, outputs=outputs)

Now the output is still produced by a layer in the Keras graph, even though it uses a TensorFlow operation internally.

Use Models as Layers When Needed

Another clean pattern is to build a submodel and call it on the input, because Keras models are themselves layer-like objects.

1import tensorflow as tf
2
3def build_backbone():
4    inp = tf.keras.Input(shape=(8,))
5    x = tf.keras.layers.Dense(16, activation="relu")(inp)
6    out = tf.keras.layers.Dense(4)(x)
7    return tf.keras.Model(inp, out)
8
9backbone = build_backbone()
10
11inputs = tf.keras.Input(shape=(8,))
12features = backbone(inputs)
13outputs = tf.keras.layers.Dense(1)(features)
14
15model = tf.keras.Model(inputs, outputs)

This works because features is still the output of a Keras model call connected to the current input graph.

Raw Arrays and Python Values Are Not Model Outputs

Another version of the same mistake is trying to pass plain Python or NumPy values as outputs:

1import numpy as np
2import tensorflow as tf
3
4inputs = tf.keras.Input(shape=(4,))
5outputs = np.array([1.0, 2.0])
6
7model = tf.keras.Model(inputs, outputs)

This fails for the same reason: Keras cannot infer a symbolic computation graph from ordinary data.

If you want a fixed transformation or constant-like behavior in the graph, wrap it in a Keras layer or combine it with the input through layer-compatible operations.

A Good Mental Model

Think of Functional API model construction as graph wiring, not arbitrary tensor collection. The constructor is not asking:

• “Do you have some tensor-like thing?”

It is asking:

• “Do you have output tensors that were produced by layer calls from these inputs?”

If the answer is no, Keras cannot serialize, summarize, trace, or train the model correctly.

Common Fixes

When you see this error, check whether the supposed output:

• comes from keras.Input
• passes through Keras layers or submodels
• stays inside the symbolic graph
• avoids being replaced by raw constants or NumPy arrays

If you have custom math, the usual fix is to wrap it in:

• 'tf.keras.layers.Lambda'
• a custom Layer
• a nested Model

That keeps the output attached to the graph Keras is trying to build.

Common Pitfalls

• Passing a plain TensorFlow constant as the model output.
• Mixing NumPy arrays or Python values directly into outputs=.
• Performing custom computation outside a Keras layer and breaking the symbolic graph.
• Forgetting that the Functional API requires outputs to be connected to the declared inputs.
• Trying to build a model from intermediate values that were not produced by calling a layer or model.

Summary

• Functional Keras models must be built from outputs that originate from keras.Input through layer or model calls.
• Raw constants, NumPy arrays, and disconnected tensors cannot be used as model outputs.
• If you need custom math, wrap it in a Lambda layer or a custom layer.
• Submodels work cleanly because a Keras model can be called like a layer.
• The fix is usually to reconnect the output to the symbolic Keras graph, not just to change tensor types.