Introduction

This Keras shape error usually happens because sparse_categorical_crossentropy and TruePositives expect different label formats. In a sparse multiclass setup, y_true is typically a vector of integer class IDs, while TruePositives is designed for binary or same-shape label predictions, so it cannot compare a class-index vector against a full probability matrix directly.

Why the Shapes Do Not Match

With sparse multiclass classification, your model output often has shape:

batch_size, num_classes

For example, 2,64 means two samples and sixty-four class probabilities per sample.

But sparse labels usually have shape:

batch_size

or sometimes batch_size, 1

So if you compile a model like this:

1model.compile(
2    optimizer="adam",
3    loss="sparse_categorical_crossentropy",
4    metrics=[tf.keras.metrics.TruePositives()]
5)

Keras tries to feed TruePositives tensors that do not represent the same kind of object. One side is sparse integer labels, the other is dense per-class predictions.

That is why the metric raises an incompatible-shape error.

What TruePositives Is For

TruePositives is meant for binary-style comparisons where predictions and labels align elementwise after thresholding. It is a natural fit for:

• binary classification
• multilabel classification with same-shape targets and predictions

It is not the right built-in metric for ordinary sparse multiclass classification.

Use the Right Multiclass Metric Instead

For sparse multiclass problems, the most common built-in metric is SparseCategoricalAccuracy.

1import tensorflow as tf
2
3model.compile(
4    optimizer="adam",
5    loss="sparse_categorical_crossentropy",
6    metrics=[tf.keras.metrics.SparseCategoricalAccuracy()]
7)

This metric understands that y_true is sparse class indices and y_pred is a probability distribution across classes.

That is why it works cleanly with sparse_categorical_crossentropy.

If You Need True Positives Per Class

If your real goal is class-wise precision, recall, or true-positive counts, convert the predictions into class IDs first and then compare class labels in a custom metric or after evaluation.

A simple custom sparse accuracy-style metric might look like this:

1import tensorflow as tf
2
3
4def sparse_argmax_match(y_true, y_pred):
5    y_true = tf.cast(tf.reshape(y_true, [-1]), tf.int64)
6    y_pred = tf.argmax(y_pred, axis=-1, output_type=tf.int64)
7    matches = tf.cast(tf.equal(y_true, y_pred), tf.float32)
8    return tf.reduce_mean(matches)

For per-class true positives, it is often cleaner to compute a confusion matrix after prediction rather than trying to force the generic TruePositives metric into a sparse multiclass training loop.

Example of a Correct Sparse Multiclass Setup

1import tensorflow as tf
2from tensorflow import keras
3
4model = keras.Sequential([
5    keras.layers.Input(shape=(20,)),
6    keras.layers.Dense(32, activation="relu"),
7    keras.layers.Dense(64, activation="softmax")
8])
9
10model.compile(
11    optimizer="adam",
12    loss="sparse_categorical_crossentropy",
13    metrics=[keras.metrics.SparseCategoricalAccuracy()]
14)

Here the output layer has 64 units, the labels are integer class IDs, and the metric matches that representation.

When One-Hot Metrics Are Appropriate

If you intentionally convert labels to one-hot vectors, then the setup changes. At that point you would usually use:

• 'categorical_crossentropy instead of sparse_categorical_crossentropy'
• metrics that expect one-hot or same-shape tensors

The key is consistency. Loss, labels, model output, and metrics all have to agree on the representation.

Common Pitfalls

Mixing sparse labels with metrics designed for elementwise binary comparisons is the main cause of this error.

Using a softmax output with one-hot-oriented metrics while keeping integer labels is another representation mismatch.

Trying to measure multiclass true positives inside training with the wrong built-in metric can also obscure the actual evaluation goal. Often a confusion matrix after prediction is the better tool.

Finally, always inspect the shapes of y_true and y_pred before debugging deeper. Most of these errors become obvious once the tensor shapes are printed.

Summary

• 'sparse_categorical_crossentropy expects sparse integer labels, while TruePositives expects same-shape binary-style comparisons'
• the shape error occurs because those expectations do not align in standard multiclass classification
• use SparseCategoricalAccuracy or another sparse-compatible metric during training
• if you need per-class true-positive counts, compute them from predicted class IDs or a confusion matrix
• keep loss, labels, output layer, and metrics consistent with the same label representation