Introduction

This Keras error means two tensors are being concatenated along one axis, but they do not match on the other axes. In convolutional networks, the mismatch is usually in height or width rather than channel count.

The failure often appears in encoder-decoder models such as U-Net, feature pyramid networks, or any architecture that downsamples and later merges features. One branch ends up with shape 523 x 523, another with 522 x 522, and Concatenate rejects the pair.

What Concatenate Requires

If you concatenate along the channel axis, all non-channel dimensions must match exactly. For tensors shaped like (batch, height, width, channels), this means:

• batch can stay dynamic
• height must match
• width must match
• channels may differ only if channels are the concat axis

So these shapes are valid together:

• '(None, 128, 128, 32)'
• '(None, 128, 128, 16)'

These are not:

• '(None, 128, 128, 32)'
• '(None, 127, 128, 16)'

Why Mismatches Happen

The usual causes are:

• odd input sizes combined with pooling or strided convolution
• different padding modes such as same on one branch and valid on another
• upsampling that does not exactly invert the earlier downsampling
• manual cropping on one path but not the other

A classic example is a U-Net with an input size that is not divisible by repeated downsampling factors. After several pool and upsample steps, the skip connection tensor and decoder tensor drift by one pixel.

A Small Reproducible Example

This model fails because the tensors differ in height and width:

1import tensorflow as tf
2from tensorflow.keras import layers, Model
3
4inputs = layers.Input(shape=(523, 523, 3))
5
6x1 = layers.Conv2D(32, 3, padding="same")(inputs)
7x2 = layers.MaxPooling2D(pool_size=2)(x1)
8x2 = layers.UpSampling2D(size=2)(x2)
9
10outputs = layers.Concatenate(axis=-1)([x1, x2])
11model = Model(inputs, outputs)

Depending on the exact chain of layers, x2 may come back at 522 x 522 while x1 remains 523 x 523. That single-pixel difference is enough to trigger the exception.

Ways To Fix It

There is no universal fix. The right one depends on model intent.

Option 1: Use Compatible Input Sizes

If the network repeatedly downsamples by powers of two, choose image sizes divisible by the total factor. This is often the simplest solution for segmentation models.

Option 2: Crop One Branch

If a skip connection is intentionally larger, crop it before concatenation:

1import tensorflow as tf
2from tensorflow.keras import layers, Model
3
4inputs = layers.Input(shape=(523, 523, 3))
5
6x1 = layers.Conv2D(32, 3, padding="same")(inputs)
7x2 = layers.MaxPooling2D(pool_size=2)(x1)
8x2 = layers.UpSampling2D(size=2)(x2)
9x1_cropped = layers.Cropping2D(((0, 1), (0, 1)))(x1)
10
11outputs = layers.Concatenate(axis=-1)([x1_cropped, x2])
12model = Model(inputs, outputs)
13print(model.output_shape)

Option 3: Pad the Smaller Tensor

If preserving the larger spatial resolution matters more, pad the smaller branch instead of cropping the larger one.

Option 4: Align Padding Strategy

Make sure both branches use consistent same or valid behavior, and audit every pooling, strided convolution, transpose convolution, and resize operation.

How To Debug It Quickly

Print the shapes of intermediate tensors before the failing Concatenate:

print(x1.shape)
print(x2.shape)

In functional models, this is often enough to identify the first layer where the spatial sizes diverged. You do not need to guess. Follow the shape chain.

Common Pitfalls

• Assuming UpSampling2D exactly reverses every earlier pooling step.
• Feeding odd-sized images into architectures that repeatedly downsample by two.
• Mixing valid and same padding across branches without checking dimensions.
• Cropping or padding blindly instead of deciding which spatial alignment is semantically correct.
• Looking only at channel counts when the mismatch is actually in height or width.

Summary

• 'Concatenate requires equal shapes on every axis except the concat axis.'
• In CNNs, the mismatch is usually caused by downsampling and upsampling drift.
• Odd spatial dimensions are a common source of one-pixel errors.
• Fixes include resizing inputs, cropping, padding, or making branch padding strategies consistent.
• Print intermediate shapes and locate the first divergence instead of debugging by guesswork.