TensorFlow Pre-trained Models
Pre-trained models are neural networks that were already trained on massive datasets by research teams with enormous computing resources. TensorFlow provides dozens of these models through the tf.keras.applications module, ready to use immediately. Each model arrives with its learned weights, its architecture, and often its preprocessing function. You can use them as feature extractors, as starting points for fine-tuning, or as complete classifiers for the 1,000 ImageNet categories they were trained on.
The Library Analogy
Imagine a university library with millions of carefully organized books. Researchers spent decades building and cataloguing the collection. You can walk in and use the entire collection on day one rather than spending decades building your own. Pre-trained models are this library — decades of computer vision research encoded into weights you can download in minutes.
Available Pre-trained Models in tf.keras.applications
Model Size Top-1 Accuracy Parameters Best For ────────────────────────────────────────────────────────────────────────── MobileNetV2 14 MB 71.8% 3.4M Mobile, embedded MobileNetV3Small 10 MB 67.4% 2.5M Ultra-compact devices EfficientNetB0 29 MB 77.1% 5.3M General purpose EfficientNetB4 75 MB 82.6% 19M High accuracy ResNet50 98 MB 74.9% 25M Research baseline ResNet152 232 MB 76.6% 60M Maximum ResNet accuracy VGG16 528 MB 71.3% 138M Simple architecture InceptionV3 92 MB 77.9% 23M Multi-scale features Xception 88 MB 79.0% 22M ImageNet leader NASNetLarge 343 MB 82.5% 88M Top accuracy ────────────────────────────────────────────────────────────────────────── Top-1 Accuracy = % of ImageNet test images predicted correctly (top prediction)
Loading a Pre-trained Model
import tensorflow as tf
# Load MobileNetV2 with ImageNet weights
# include_top=True: includes the final 1000-class classification layer
base_model = tf.keras.applications.MobileNetV2(
input_shape=(224, 224, 3),
include_top=True,
weights='imagenet'
)
# Predict on a new image (for the 1000 ImageNet classes)
import numpy as np
img = tf.keras.utils.load_img('cat.jpg', target_size=(224, 224))
img_array = tf.keras.utils.img_to_array(img)
img_array = tf.expand_dims(img_array, 0)
# Preprocess (each model has its own normalization range)
img_array = tf.keras.applications.mobilenet_v2.preprocess_input(img_array)
predictions = base_model.predict(img_array)
decoded = tf.keras.applications.mobilenet_v2.decode_predictions(predictions, top=5)
for class_id, class_name, confidence in decoded[0]:
print(f"{class_name:20s}: {confidence:.2%}")
# tabby_cat : 52.3%
# tiger_cat : 23.1%
# Egyptian_cat : 11.4%
# Persian_cat : 4.2%
# lynx : 2.1%
Preprocessing Functions
Each model was trained with a specific pixel normalization scheme. Use the matching preprocessing function or your model will produce nonsense predictions.
Model Family Preprocessing Function Output Range ──────────────────────────────────────────────────────────────────────────── MobileNetV2/V3 mobilenet_v2.preprocess_input() [-1, 1] EfficientNet B0-B7 efficientnet.preprocess_input() [0, 255] (no change!) ResNet50/101/152 resnet.preprocess_input() BGR, mean-subtracted VGG16/VGG19 vgg16.preprocess_input() BGR, mean-subtracted InceptionV3 inception_v3.preprocess_input() [-1, 1] Xception xception.preprocess_input() [-1, 1] ────────────────────────────────────────────────────────────────────────────
# Always match the preprocessing to the model img_preprocessed = tf.keras.applications.efficientnet.preprocess_input(img_array) img_preprocessed = tf.keras.applications.resnet.preprocess_input(img_array) img_preprocessed = tf.keras.applications.inception_v3.preprocess_input(img_array)
Loading Without the Top (For Transfer Learning)
# include_top=False removes the final 1000-class Dense layer
# This is the standard setup for transfer learning to custom classes
base = tf.keras.applications.EfficientNetB0(
input_shape=(224, 224, 3),
include_top=False, # ← Remove the ImageNet classifier
weights='imagenet' # ← Keep all the learned features
)
print("Number of layers:", len(base.layers))
print("Output shape:", base.output_shape) # (None, 7, 7, 1280)
Feature Extraction Without Training
# Use a frozen pre-trained model to extract features from your images
base = tf.keras.applications.MobileNetV2(
input_shape=(224, 224, 3),
include_top=False,
weights='imagenet'
)
base.trainable = False
# Extract features from your images
def extract_features(image_batch):
preprocessed = tf.keras.applications.mobilenet_v2.preprocess_input(image_batch)
return base(preprocessed, training=False)
# (batch, 7, 7, 1280) per image
features = extract_features(your_image_batch)
Inspecting Model Architecture
base = tf.keras.applications.ResNet50(include_top=False, weights='imagenet')
# Print layer names and output shapes
for i, layer in enumerate(base.layers[:10]):
print(f"Layer {i:3d}: {layer.name:40s} → {layer.output_shape}")
# Visualize architecture
tf.keras.utils.plot_model(base, to_file='resnet50.png',
show_shapes=True, expand_nested=True)
Model Size and Inference Speed Comparison
Benchmark (single image, CPU): MobileNetV2: ~25ms — fast enough for real-time mobile EfficientNetB0: ~35ms — good accuracy/speed balance ResNet50: ~80ms — slower but widely supported VGG16: ~150ms — very slow, memory-heavy Benchmark (batch of 32, GPU): MobileNetV2: ~8ms — ideal for server-side real-time EfficientNetB4: ~25ms — best accuracy per second ResNet152: ~60ms — comprehensive features
Which Pre-trained Model to Choose
Goal Recommended Model ──────────────────────────────────────────────────────────────── Mobile app deployment MobileNetV2 or MobileNetV3 Best accuracy, any size OK EfficientNetB4 or EfficientNetB7 Research reproducibility ResNet50 (universal baseline) Quick prototype MobileNetV2 (fast + lightweight) Multi-scale feature extraction InceptionV3 Edge device (IoT, microcontroller) MobileNetV3Small or TFLite models ────────────────────────────────────────────────────────────────
Pre-trained models are the biggest productivity multiplier in computer vision. Instead of weeks of training, you achieve state-of-the-art feature extraction in a single download. The next topic shows exactly how to take these models and fine-tune their layers for your specific dataset and task.