How to run prediction with a pretrained Keras application model

Application models in Keras package trained image-classification architectures with downloadable weights, so a local photo can pass through a production-shaped inference path without training a model first. MobileNetV2 keeps that first prediction small while retaining the image loading, model-specific preprocessing, and ImageNet decoding required by larger application models.

The TensorFlow backend must be selected before importing Keras. MobileNetV2 expects an RGB image resized to 224 by 224 pixels, a batch dimension, and its matching preprocess_input() transformation before inference.

Decoded labels apply only to the original ImageNet classifier head loaded with weights=“imagenet” and the default include_top=True. Feature extractors and models with custom classifier heads need their own label mapping instead of decode_predictions().

1. Save a recognizable RGB photo as sample.jpg in the project directory.

   Keras also accepts JPEG, PNG, BMP, and non-animated GIF inputs through load_img(), but the filename passed to the script must match the saved image.

2. Create predict_mobilenet.py with the backend selection and application imports.

   predict_mobilenet.py

   import os
   import sys
    
   os.environ["KERAS_BACKEND"] = "tensorflow"
    
   import numpy as np
   import keras
   from keras.applications.mobilenet_v2 import (
       MobileNetV2,
       decode_predictions,
       preprocess_input,
   )

   Keras reads KERAS_BACKEND during import, so the assignment must remain above import keras.

3. Append the model and image preprocessing block to predict_mobilenet.py.

   image_path = sys.argv[1] if len(sys.argv) > 1 else "sample.jpg"
    
   model = MobileNetV2(weights="imagenet")
   image = keras.utils.load_img(
       image_path,
       color_mode="rgb",
       target_size=(224, 224),
   )
   image_array = keras.utils.img_to_array(image)
   image_batch = np.expand_dims(image_array, axis=0)
   model_input = preprocess_input(image_batch.copy())

   The first model construction downloads the pretrained weights to the local Keras model cache. Copying the batch prevents preprocess_input() from overwriting the original floating-point image array.

4. Append the inference and tensor-shape checks to predict_mobilenet.py.

   predictions = model.predict(model_input, verbose=0)
    
   print(f"model input shape: {model_input.shape}")
   print(f"prediction shape: {predictions.shape}")

5. Append the top-three ImageNet decoding loop to predict_mobilenet.py.

   for rank, (_, label, score) in enumerate(
       decode_predictions(predictions, top=3)[0],
       start=1,
   ):
       print(f"{rank}. {label}: {score:.3f}")

6. Compare the completed predict_mobilenet.py with the consolidated file.

   predict_mobilenet.py

   import os
   import sys
    
   os.environ["KERAS_BACKEND"] = "tensorflow"
    
   import numpy as np
   import keras
   from keras.applications.mobilenet_v2 import (
       MobileNetV2,
       decode_predictions,
       preprocess_input,
   )
    
   image_path = sys.argv[1] if len(sys.argv) > 1 else "sample.jpg"
    
   model = MobileNetV2(weights="imagenet")
   image = keras.utils.load_img(
       image_path,
       color_mode="rgb",
       target_size=(224, 224),
   )
   image_array = keras.utils.img_to_array(image)
   image_batch = np.expand_dims(image_array, axis=0)
   model_input = preprocess_input(image_batch.copy())
    
   predictions = model.predict(model_input, verbose=0)
    
   print(f"model input shape: {model_input.shape}")
   print(f"prediction shape: {predictions.shape}")
    
   for rank, (_, label, score) in enumerate(
       decode_predictions(predictions, top=3)[0],
       start=1,
   ):
       print(f"{rank}. {label}: {score:.3f}")

7. Verify pretrained MobileNetV2 inference by running the completed script with the saved image.

   $ python predict_mobilenet.py sample.jpg
   model input shape: (1, 224, 224, 3)
   prediction shape: (1, 1000)
   1. daisy: 0.715
   2. bee: 0.067
   3. orange: 0.008

   The shown prediction used a red sunflower photo. Scores and labels change with the image, while the one-image input batch and 1000-class output shape remain the same for this ImageNet configuration.