Keras Applications

1.0.8 · maintenance · verified Thu Apr 09

Keras Applications provides reference implementations of popular deep learning models alongside pre-trained weights. Currently at version 1.0.8, this standalone library historically served as the primary source for models like VGG, ResNet, and MobileNet. However, these models have since been fully integrated into the core Keras library and are now primarily accessed via `tf.keras.applications` within TensorFlow-based Keras environments.

Warnings

breaking The standalone `keras-applications` GitHub repository and Python package are officially deprecated. All Keras Application models have been integrated into the core Keras library and the TensorFlow pip package. Direct imports from `keras_applications` may not work or might refer to an outdated version.
Fix: Migrate your imports to `from tensorflow.keras.applications import ...` if using TensorFlow with Keras 2.x/3.x, or `from keras.applications import ...` if using standalone Keras 3 with another backend. Remove `pip install keras-applications` and rely on `tensorflow` or `keras` installation.
gotcha Each Keras Application model expects a specific type of input preprocessing. Failing to call the correct `preprocess_input` function for the chosen model (e.g., converting RGB to BGR, zero-centering pixels, or scaling to [-1, 1]) will lead to unstable training, incorrect feature extraction, or poor prediction accuracy when using pre-trained weights.
Fix: Always import and apply the `preprocess_input` function specific to the model you are using (e.g., `from tensorflow.keras.applications.vgg16 import preprocess_input`).
gotcha With TensorFlow 2.x, Keras is deeply integrated as `tf.keras`. Using `import keras` (standalone Keras 2) and `tf.keras` interchangeably or in the same environment can lead to confusion and `AttributeError` or `ModuleNotFoundError` issues, especially when loading saved models. Keras 3 further changes import paths for multi-backend support.
Fix: Standardize on `from tensorflow.keras import ...` or `import tensorflow.keras as keras` for TensorFlow-based projects. For Keras 3 (standalone `keras` package), use `from keras import ...`. Avoid mixing `keras` and `tf.keras` imports in the same codebase. Consider migrating older Keras 2 code to `tf_keras` if maintaining compatibility is paramount for inactive projects.

Install

pip install keras-applications Install standalone package (legacy)
pip install tensorflow Recommended for modern use (includes Keras Applications via tf.keras)

Imports

ResNet50
```
from tensorflow.keras.applications.resnet50 import ResNet50
```
The standalone `keras_applications` library is deprecated; models are now part of `tf.keras.applications` or `keras.applications` in Keras 3.
preprocess_input
```
from tensorflow.keras.applications.resnet50 import preprocess_input
```
Preprocessing functions are now typically model-specific and found alongside the model import in `tf.keras.applications`.

Quickstart

This quickstart demonstrates how to load a pre-trained `ResNet50` model, preprocess a sample image, and make predictions using the `tf.keras.applications` module. Weights are automatically downloaded upon first instantiation. The example includes creating a dummy image for immediate execution.

import numpy as np
from tensorflow.keras.applications.resnet50 import ResNet50, preprocess_input, decode_predictions
from tensorflow.keras.preprocessing import image
import os

# Create a dummy image for demonstration
dummy_image_path = 'dummy_elephant.jpg'
if not os.path.exists(dummy_image_path):
    try:
        from PIL import Image
        img_data = np.random.randint(0, 255, size=(224, 224, 3), dtype=np.uint8)
        dummy_img = Image.fromarray(img_data)
        dummy_img.save(dummy_image_path)
        print(f"Created dummy image: {dummy_image_path}")
    except ImportError:
        print("Pillow not installed. Skipping dummy image creation.")
        print("Please provide a real image path to run the example.")
        dummy_image_path = None

if dummy_image_path:
    # Load the pre-trained ResNet50 model
    # weights='imagenet' downloads weights if not already present
    model = ResNet50(weights='imagenet')
    print("ResNet50 model loaded successfully.")

    # Load an image and resize it to the target size expected by ResNet50
    img = image.load_img(dummy_image_path, target_size=(224, 224))
    x = image.img_to_array(img)
    x = np.expand_dims(x, axis=0) # Add batch dimension

    # Preprocess the image for the model (e.g., channel reordering, mean subtraction)
    x = preprocess_input(x)

    # Make predictions
    preds = model.predict(x)

    # Decode the top 3 predictions
    decoded_preds = decode_predictions(preds, top=3)[0]
    print('Predicted:', decoded_preds)

    # Clean up dummy image
    os.remove(dummy_image_path)
else:
    print("Quickstart example requires Pillow to create a dummy image.")

view raw JSON →