Introduction

In Keras 2, the old Merge-style API gave way to clearer functional merging helpers such as keras.layers.concatenate and the Concatenate layer. So if you want to "concat" multiple tensors in Keras 2.x, the practical answer is to use the Functional API and concatenate the tensors along the axis you need.

Use concatenate in the Functional API

The most common pattern looks like this:

1from tensorflow import keras
2from tensorflow.keras import layers
3
4left_input = keras.Input(shape=(16,))
5right_input = keras.Input(shape=(8,))
6
7left_branch = layers.Dense(32, activation="relu")(left_input)
8right_branch = layers.Dense(32, activation="relu")(right_input)
9
10merged = layers.concatenate([left_branch, right_branch], axis=-1)
11output = layers.Dense(1)(merged)
12
13model = keras.Model(inputs=[left_input, right_input], outputs=output)
14model.summary()

This is the standard Keras 2 way to concatenate two branches. The important rule is that the input tensors must be compatible on every axis except the axis you are concatenating along.

You can also use the Concatenate layer object

The layer form is equivalent and sometimes cleaner in larger models:

1from tensorflow import keras
2from tensorflow.keras import layers
3
4input_a = keras.Input(shape=(10,))
5input_b = keras.Input(shape=(10,))
6
7x1 = layers.Dense(8, activation="relu")(input_a)
8x2 = layers.Dense(8, activation="relu")(input_b)
9
10merged = layers.Concatenate(axis=-1)([x1, x2])
11output = layers.Dense(2, activation="softmax")(merged)
12
13model = keras.Model(inputs=[input_a, input_b], outputs=output)
14model.summary()

Function form and layer form are both fine. Pick the one that reads more naturally in your model code.

Match the shapes before concatenating

Concatenation does not magically align incompatible tensor shapes. For example, this works:

shape_a = (None, 8)
shape_b = (None, 8)

because the tensors are equal on all dimensions except the concatenation axis, which is the last axis.

But this would fail if one branch produced shape (None, 8) and the other (None, 5, 8) without additional reshaping or pooling.

If branches differ, fix them first using layers such as:

• 'Flatten'
• 'GlobalAveragePooling2D'
• 'Reshape'
• 'Dense'

For example:

1image_input = keras.Input(shape=(32, 32, 3))
2meta_input = keras.Input(shape=(10,))
3
4image_features = layers.Conv2D(16, 3, activation="relu")(image_input)
5image_features = layers.GlobalAveragePooling2D()(image_features)
6
7meta_features = layers.Dense(16, activation="relu")(meta_input)
8
9merged = layers.concatenate([image_features, meta_features])
10output = layers.Dense(1)(merged)
11
12model = keras.Model(inputs=[image_input, meta_input], outputs=output)

This is a very common multi-input architecture pattern.

Why the old Merge naming causes confusion

Older Keras tutorials often mention Merge directly, but modern Keras code is clearer with explicit layers such as:

• 'Concatenate'
• 'Add'
• 'Multiply'
• 'Average'

That makes the graph easier to read than a generic merge concept with multiple modes hidden inside one older API style.

So if you are reading old examples, translate them mentally:

• old merge concat idea
• new Concatenate or layers.concatenate(...)

The underlying modeling idea is the same even if the API naming changed.

Common Pitfalls

The biggest mistake is trying to concatenate tensors with incompatible shapes. Keras will raise a shape error because concatenation is only valid when the non-concatenated dimensions match.

Another common issue is mixing Sequential-style thinking with a multi-input graph. Concatenation belongs naturally in the Functional API because you are joining branches, not just stacking one layer after another in a straight line.

People also copy very old Merge examples without adapting them to the newer Concatenate patterns that Keras 2 code actually expects.

Finally, be deliberate about the axis. Concatenating along the wrong dimension can produce a valid tensor with the wrong semantics.

Summary

• In Keras 2, use layers.concatenate(...) or layers.Concatenate(...) for concat-style merging.
• Concatenated tensors must match on all axes except the concatenation axis.
• The Functional API is the natural place for branch merges.
• Use pooling, flattening, or reshaping when branch outputs are not directly compatible.
• Treat older Merge tutorials as historical syntax and translate them to modern concatenation layers.