Introduction

When Keras says "Data cardinality is ambiguous", it usually means your inputs do not agree on how many samples exist. The first dimension of every training input and target must line up, so this error is almost always about mismatched shapes rather than about the model architecture itself.

Check the First Dimension of Every Input

Keras expects the same sample count across:

• 'x and y in a single-input model'
• every item inside a list or dict of multiple inputs
• sample weights, if you pass them

Here is a simple example that fails:

1import numpy as np
2from tensorflow import keras
3
4x = np.random.rand(100, 10)
5y = np.random.randint(0, 2, size=(90,))
6
7model = keras.Sequential([
8    keras.layers.Dense(16, activation="relu", input_shape=(10,)),
9    keras.layers.Dense(1, activation="sigmoid"),
10])
11
12model.compile(optimizer="adam", loss="binary_crossentropy")
13model.fit(x, y, epochs=5)

x has 100 samples, but y has 90. Keras cannot guess how those should be paired, so it raises the cardinality error.

Print Shapes Before Calling fit

The fastest debugging step is to inspect shapes right before training:

print("x shape:", x.shape)
print("y shape:", y.shape)

For multiple inputs, print all of them:

1images = np.random.rand(100, 64, 64, 3)
2metadata = np.random.rand(95, 5)
3labels = np.random.randint(0, 3, size=(100,))
4
5print(images.shape)
6print(metadata.shape)
7print(labels.shape)

In this example, metadata is shorter than the other arrays, so the model cannot train until the preprocessing pipeline produces aligned samples.

Fix the Upstream Data Pipeline

The real solution is usually not to slice arrays at random, but to find why the sizes drifted apart. Common causes include:

• filtering one dataframe but not the matching labels
• dropping missing values on one input only
• shuffling features and labels separately
• generating windows or sequences with inconsistent logic

A valid multi-input fit call looks like this:

1images = np.random.rand(100, 64, 64, 3)
2metadata = np.random.rand(100, 5)
3labels = np.random.randint(0, 3, size=(100,))
4
5input_image = keras.Input(shape=(64, 64, 3))
6input_meta = keras.Input(shape=(5,))
7
8x1 = keras.layers.Flatten()(input_image)
9x2 = keras.layers.Dense(8, activation="relu")(input_meta)
10x = keras.layers.concatenate([x1, x2])
11output = keras.layers.Dense(3, activation="softmax")(x)
12
13model = keras.Model([input_image, input_meta], output)
14model.compile(optimizer="adam", loss="sparse_categorical_crossentropy")
15model.fit([images, metadata], labels, epochs=3)

Every input now has 100 samples, so Keras can match rows correctly.

Do Not Confuse Batch Size with Cardinality

This error is about the total number of samples, not about the batch size. Changing batch_size=32 to batch_size=16 will not fix mismatched datasets, because Keras still needs the full arrays to align before it can create batches.

That distinction is important because many first attempts focus on training settings instead of the actual data problem.

Common Pitfalls

The most common mistake is checking only the feature array and assuming the labels must match automatically. Print every array that goes into fit.

Another issue is accidentally converting one input into shape (1, n) while another is (n, ...). That can happen after reshaping, wrapping arrays in an extra list, or using np.expand_dims in the wrong place.

People also sometimes "fix" the problem by truncating everything to the shortest array without understanding why lengths diverged. That can hide a serious preprocessing bug.

Finally, if you use pandas, be careful when filtering rows. Index alignment and row dropping can produce arrays that look reasonable individually but no longer represent the same samples.

Summary

• "Data cardinality is ambiguous" usually means your training inputs or labels have different sample counts.
• Check the first dimension of every array passed to fit, evaluate, or predict.
• Fix the upstream preprocessing pipeline so all inputs stay aligned.
• Do not confuse sample-count mismatches with batch-size settings.
• Treat random truncation as a last resort, not as the default solution.