5 Best Ways to Use TensorFlow with Pre-Trained Models in Python

Rate this post

πŸ’‘ Problem Formulation: Leveraging pre-trained models can dramatically speed up the development process for Machine Learning projects. However, many developers struggle with the correct methodology for compiling these models using TensorFlow in Python. Let’s assume you have a pre-trained model and you want to efficiently compile it to recognize image patterns or classify text data. This article will demonstrate how to apply various methods to compile and fine-tune a pre-trained model using TensorFlow in Python.

Method 1: Loading and Compiling a Pre-Trained Model Directly

TensorFlow’s Keras API simplifies the process of loading pre-trained models. The ‘compile()’ function is essential for configuring the learning process before training the model. This involves specifying the optimizer, loss function, and metrics for evaluation. The advantage of this method is its straightforwardness, as TensorFlow provides several out-of-the-box models through Keras applications.

Here’s an example:

from tensorflow.keras.applications import VGG16
from tensorflow.keras.optimizers import Adam

# Load pre-trained VGG16 model with weights trained on ImageNet
model = VGG16(weights='imagenet')

# Compile the model
model.compile(optimizer=Adam(), loss='categorical_crossentropy', metrics=['accuracy'])

Output: Model compiled with Adam optimizer, categorical crossentropy loss, and accuracy metric.

In the given snippet, the VGG16 model is loaded with pre-trained ImageNet weights, and then the compile method is called with an Adam optimizer and categorical crossentropy as the loss function. The accuracy is added as a metric to track the model’s performance during training. This setup is typically used for image classification tasks.

Method 2: Importing a Pre-Trained Model for Feature Extraction

Sometimes, we are interested in using a pre-trained model as a feature extractor. In this scenario, we remove the top layer and prevent the weights from being updated during training. After extracting features, you can add custom layers on top for the specific task at hand and compile the model.

Here’s an example:

from tensorflow.keras.applications import ResNet50
from tensorflow.keras.layers import Dense, GlobalAveragePooling2D
from tensorflow.keras.models import Model
from tensorflow.keras.optimizers import RMSprop

# Load ResNet50 without the top classification layer
base_model = ResNet50(weights='imagenet', include_top=False)

# Freeze the base model weights
base_model.trainable = False

# Add a global spatial average pooling layer and a dense classifier on top
x = base_model.output
x = GlobalAveragePooling2D()(x)
predictions = Dense(10, activation='softmax')(x)
model = Model(inputs=base_model.input, outputs=predictions)

# Compile the model
model.compile(optimizer=RMSprop(), loss='categorical_crossentropy', metrics=['accuracy'])

Output: Model set up for feature extraction with frozen pre-trained weights, using RMSprop optimizer.

This code first imports the ResNet50 model without the top layer for use as a feature extractor. The model’s layers are frozen to preserve pre-trained weights. A new set of layers is added for a new task (classifying 10 categories here), and the composite model is compiled with the RMSprop optimizer.

Method 3: Fine-Tuning a Few Top Layers of a Pre-Trained Model

Fine-tuning is a technique where you unfreeze some of the top layers of a pre-trained model and update their weights during training to adapt to the new data. This method requires setting the trainable parameter of these layers to true. It combines the strengths of transfer learning and can achieve higher accuracy on the new task.

Here’s an example:

from tensorflow.keras.applications import InceptionV3
from tensorflow.keras.optimizers import SGD

# Load InceptionV3, without the top layer
base_model = InceptionV3(weights='imagenet', include_top=False)

# Unfreeze the top layers of the model
for layer in base_model.layers[-50:]:
    layer.trainable = True

# Compile the model
model.compile(optimizer=SGD(lr=0.0001, momentum=0.9), loss='categorical_crossentropy', metrics=['accuracy'])

Output: Model configured for fine-tuning the top layers using a SGD optimizer with a low learning rate.

The InceptionV3 model is loaded, and the top 50 layers are unfrozen to enable fine-tuning. The model is compiled with a Stochastic Gradient Descent (SGD) optimizer with modified learning rate and momentum to update the weights carefully without large disturbances.

Method 4: Customizing and Compiling a Pre-Trained Model with a New Task-Specific Layer

For more custom applications, developers can add a new top layer tailored to the specific task, such as a new classifier for a different number of classes. This layer is initialized without pre-trained weights and typically is trained from scratch.

Here’s an example:

from tensorflow.keras.applications import MobileNetV2
from tensorflow.keras.layers import Input, Flatten, Dense
from tensorflow.keras.models import Model

# Load MobileNetV2 model without the top layer
input_tensor = Input(shape=(128, 128, 3))
base_model = MobileNetV2(input_shape=(128, 128, 3), include_top=False, weights='imagenet')

# Adding custom layers
x = base_model.output
x = Flatten()(x)
custom_predictions = Dense(5, activation='softmax')(x)
custom_model = Model(inputs=base_model.input, outputs=custom_predictions)

# Compile the model
custom_model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])

Output: A newly compiled model tailored to a task with 5 classes, using the Adam optimizer.

In this example, a MobileNetV2 model is used as the base. The top layer is replaced with a custom Flatten layer and a Dense layer with 5 outputs corresponding to 5 new categories. The newly created model then gets compiled with default settings for the Adam optimizer.

Bonus One-Liner Method 5: Quick Pre-Trained Model Compilation with a Single Line

For rapid prototyping, TensorFlow’s Keras API provides a one-liner to load, compile, and prepare a pre-trained model for transfer learning, given you are fine with default settings.

Here’s an example:

model = tf.keras.applications.MobileNetV2(weights='imagenet', classes=1000)
model.compile('adam', 'categorical_crossentropy', ['accuracy'])

Output: Pre-trained MobileNetV2 model loaded with default parameters and ready to train.

This code demonstrates the simplest way to load and compile a pre-trained MobileNetV2 model. It loads the weights, sets the number of classes to 1000 (default for ImageNet), and compiles the model with an Adam optimizer and categorical_crossentropy loss.


  • Method 1: Direct compilation of a pre-trained model. Strengths: Quick and easy. Weaknesses: Less flexibility for customization.
  • Method 2: Feature extraction with frozen base model layers. Strengths: Efficient reuse of learned features. Weaknesses: May not fine-tune features for the specific task.
  • Method 3: Fine-tuning top layers. Strengths: Tailors the pre-trained model more closely to new data. Weaknesses: Risk of overfitting if not managed properly.
  • Method 4: Adding custom top layers. Strengths: High degree of customization. Weaknesses: May require more data and training time.
  • Method 5: One-liner compilation. Strengths: Extremely efficient for standard tasks. Weaknesses: No customization and assumes the default parameters are adequate.