Skip to main content

How to Resolve Python Keras/TensorFlow Error "ImportError: cannot import name ..." for "pad_sequences" and "load_img"

When working with Keras for deep learning, particularly with newer versions of TensorFlow (which now fully integrates Keras), you might encounter ImportError or AttributeError messages like cannot import name 'pad_sequences' from 'keras.preprocessing.sequence' or module 'keras.preprocessing.image' has no attribute 'load_img'. These errors indicate that you are trying to import these utilities from their older, now deprecated, locations within the standalone Keras or older Keras-within-TensorFlow structure.

This guide explains these import path changes and provides the correct, modern import statements for TensorFlow 2.x and later.

Understanding the Error: Keras API Relocation in TensorFlow 2.x​

Keras, originally a standalone high-level API for neural networks, became the official high-level API for TensorFlow starting with TensorFlow 2.0.0. This integration led to a restructuring of how Keras components are accessed. Many utilities that were previously in modules like keras.preprocessing have been moved or are now primarily accessed through the tensorflow.keras namespace.

Using import paths from the older, standalone Keras structure (e.g., from keras.preprocessing...) with modern TensorFlow installations (TF 2.x and above) will lead to ImportError or AttributeError because those specific paths no longer exist or are deprecated.

Error: cannot import name 'pad_sequences' from 'keras.preprocessing.sequence'​

  • Error Message Example: 
    ImportError: cannot import name 'pad_sequences' from 'keras.preprocessing.sequence' (.../keras/preprocessing/sequence.py)
  • Cause: You are trying to import pad_sequences using the old path from a standalone Keras or an older TensorFlow version's Keras API.

In modern TensorFlow (TF 2.x+), pad_sequences is available under tensorflow.keras.utils.

main_pad_sequences.py
# ⛔️ Old, incorrect import (causes error with modern TensorFlow)
# from keras.preprocessing.sequence import pad_sequences

# ✅ Correct import for modern TensorFlow/Keras
from tensorflow.keras.utils import pad_sequences

# Example usage (assuming you have sequences to pad)
sequences = [[1, 2, 3], [4, 5], [6]]
padded_sequences = pad_sequences(sequences, padding='post', maxlen=5)

print("pad_sequences function object:", pad_sequences)
print("\nPadded Sequences:")
print(padded_sequences)

This is the standard and recommended way to access pad_sequences.

Alternative (Older): Import from keras_preprocessing.sequence​

The keras_preprocessing package was a separate utility library that Keras used. While pad_sequences might exist there, it's generally better to use the official tensorflow.keras.utils path for better integration and future compatibility.

# Alternative, but prefer tensorflow.keras.utils
from keras_preprocessing.sequence import pad_sequences
print(pad_sequences)

If this works, it might indicate an older environment setup or specific dependencies. Prioritize tensorflow.keras.utils.

Error: AttributeError: module 'keras.preprocessing.image' has no attribute 'load_img' or ImportError: cannot import name 'load_img' from 'keras.preprocessing.image'​

  • Error Message Examples: 
    AttributeError: module 'keras.preprocessing.image' has no attribute 'load_img'
    ImportError: cannot import name 'load_img' from 'keras.preprocessing.image' (.../keras/preprocessing/image.py)
  • Cause: You are trying to access load_img via the old keras.preprocessing.image module, which is deprecated or doesn't expose load_img directly in this manner in modern TensorFlow.

Similar to pad_sequences, load_img is now found in tensorflow.keras.utils.

main_load_img.py
import numpy as np # For example usage

# ⛔️ Old, incorrect import (causes error with modern TensorFlow)
# from keras.preprocessing.image import load_img
# Or trying:
# from keras.preprocessing import image
# img = image.load_img(...)

# ✅ Correct import for modern TensorFlow/Keras
from tensorflow.keras.utils import load_img, img_to_array

# Example usage (replace 'your_image.jpg' with an actual image path)
image_path = 'your_image.jpg'
try:
    # Create a dummy image for testing if you don't have one
    from PIL import Image
    dummy_img = Image.new('RGB', (60, 30), color = 'red')
    dummy_img.save(image_path)

    img = load_img(
        image_path,
        target_size=(100, 100), # Example target size
        color_mode="rgb"        # Example color mode
    )
    print("load_img function object:", load_img)
    print(f"\nLoaded image type: {type(img)}")

    # Convert PIL image to NumPy array for Keras model input
    img_array = img_to_array(img)
    print(f"Image array shape: {img_array.shape}")

except FileNotFoundError:
    print(f"Error: Image file not found at '{image_path}'. Please provide a valid path.")
except Exception as e:
    print(f"An error occurred: {e}")

Always import load_img directly from tensorflow.keras.utils.

General Rule: Prefer tensorflow.keras Imports​

As a general guideline when working with TensorFlow 2.x and newer: always try to import Keras components from the tensorflow.keras namespace first.

  • Old Style (Avoid): 
    import keras
    from keras.layers import Dense
    from keras.models import Sequential
    from keras.preprocessing import sequence, image
  • New Style (Preferred): 
    from tensorflow import keras # Access general Keras API via tf.keras
    from tensorflow.keras.layers import Dense
    from tensorflow.keras.models import Sequential
    from tensorflow.keras.utils import pad_sequences, load_img

    # For ImageDataGenerator (if needed, though tf.data is often preferred now):
    # from tensorflow.keras.preprocessing.image import ImageDataGenerator

Solution for Persistent Issues: Upgrade TensorFlow and Keras​

Outdated versions of tensorflow or keras (if you somehow have a standalone Keras installed alongside an older TensorFlow) can cause these import issues or hide the correct paths. Ensure you are using up-to-date versions.

pip install --upgrade tensorflow keras
# Or use pip3 / python -m pip etc.

# For Conda environments:
conda install -c conda-forge tensorflow keras # Or update
note

keras as a separate pip package mainly refers to Keras 3 which supports multiple backends. If using TensorFlow as the backend, tensorflow itself includes the Keras API.

Troubleshooting: Upgrade All Packages (Use with Caution)​

In complex environments, conflicts might arise from multiple outdated packages. While sometimes risky, upgrading all packages might resolve transitive dependency issues related to Keras and TensorFlow APIs.

  • Method 1: Using pip-review (install first: pip install pip-review)

    pip-review --auto

  • Method 2: Script (use with extreme caution - test thoroughly!)

    import pkg_resources
    from subprocess import call
    packages = [dist.project_name for dist in pkg_resources.working_set]
    print(f"Attempting to upgrade: {' '.join(packages)}")
    call("pip install --upgrade " + ' '.join(packages), shell=True)
note

Recommendation: Prefer upgrading TensorFlow and Keras specifically (Solution 5). Only resort to upgrading everything if specific upgrades fail and you understand the potential risks of breaking other dependencies. Always use a virtual environment.

Conclusion​

ImportErrors or AttributeErrors for Keras utilities like pad_sequences and load_img typically mean you are using outdated import paths not compatible with modern TensorFlow 2.x (which bundles Keras).

The primary solution is to update your import statements to use the tensorflow.keras namespace:

  • For pad_sequences: from tensorflow.keras.utils import pad_sequences
  • For load_img: from tensorflow.keras.utils import load_img

Always prefer importing Keras components via tensorflow.keras... when working with TensorFlow 2 and above. Keeping your tensorflow package up-to-date is also crucial for accessing the latest API structures.

Table of Contents