Skip to main content

Python TensorFlow/Keras: How to Resolve "ImportError: cannot import name 'to_categorical' from 'keras.utils'"

When working with deep learning models in modern versions of TensorFlow (2.x and newer), you may encounter the ImportError: cannot import name 'to_categorical' from 'keras.utils'. This error occurs because Keras, which was once a separate library, is now fully integrated into TensorFlow. As a result, the import paths for all its modules have changed.

The error is a clear sign that you are using an outdated import statement from a legacy, standalone Keras project in a modern TensorFlow environment. This guide will explain the change and show you the simple, one-line fix to resolve this error.

Understanding the Error: The Keras and TensorFlow Merger​

In the era of TensorFlow 1.x, Keras was a popular high-level API that could be installed and used as a standalone library. Its modules were imported directly from the keras package (e.g., from keras.layers import Dense).

With the release of TensorFlow 2.0, Keras was adopted as the official high-level API and is now bundled directly within TensorFlow. All of its functionality is accessible through the tensorflow.keras namespace. The standalone keras package is no longer the standard for TensorFlow development, and using its old import paths will lead to errors.

Reproducing the ImportError​

The error is triggered when you use the old, standalone keras import path in a modern TensorFlow environment.

Example of code causing the error:

import numpy as np
# This legacy import path is the cause of the error.
from keras.utils import to_categorical

try:
    # Example labels for a 3-class classification problem
    labels = np.array([0, 1, 2, 1, 0])
    
    # Attempting to use the function
    categorical_labels = to_categorical(labels, num_classes=3)
    print(categorical_labels)

except ImportError as e:
    print(f"Error: {e}")
except NameError:
    # This will be triggered if the import fails
    print("Import failed, could not call to_categorical.")

Output (if standalone keras is not properly configured or missing):

Import failed, could not call to_categorical.

Or if you try to run it, the traceback points to the import line:

ImportError: cannot import name 'to_categorical' from 'keras.utils'

The Solution: Update the Import Path​

The fix is to change the import statement to use the official tensorflow.keras namespace. This ensures you are using the Keras API that is integrated with your TensorFlow installation.

Solution:

import numpy as np
# ✅ Correct: Import from the tensorflow.keras namespace
from tensorflow.keras.utils import to_categorical

# Example labels for a 3-class classification problem
labels = np.array([0, 1, 2, 1, 0])

# This now works correctly
categorical_labels = to_categorical(labels, num_classes=3)

print("One-hot encoded labels:")
print(categorical_labels)

Output:

One-hot encoded labels:
[[1. 0. 0.]
 [0. 1. 0.]
 [0. 0. 1.]
 [0. 1. 0.]
 [1. 0. 0.]]
note

A Universal Fix for Keras Imports

This solution applies to all Keras modules, not just to_categorical. If you are migrating an older project, you will need to update all your Keras imports.

  • from keras.models import Sequential ➔ from tensorflow.keras.models import Sequential
  • from keras.layers import Dense, Conv2D ➔ from tensorflow.keras.layers import Dense, Conv2D
  • from keras.optimizers import Adam ➔ from tensorflow.keras.optimizers import Adam

Conclusion​

The ImportError: cannot import name 'to_categorical' from 'keras.utils' is a clear indication of an outdated import path. The solution is simple and universally applicable for any modern TensorFlow project.

Old Import Path (Legacy)Correct Import Path (TensorFlow 2+)
from keras.utils import to_categoricalfrom tensorflow.keras.utils import to_categorical
from keras.models import Modelfrom tensorflow.keras.models import Model
from keras.layers import Inputfrom tensorflow.keras.layers import Input

To resolve the error, replace all keras. import prefixes with tensorflow.keras.. This aligns your code with the modern structure of the TensorFlow library and ensures access to all of its integrated features.

Table of Contents