Get in touch
or send us a question?
CONTACT

Transfer learning and fine-tuning

Introduction

Transfer learning consists of taking features learned on one problem, and leveraging them on a new, similar problem. For instance, features from a model that has learned to identify racoons may be useful to kick-start a model meant to identify tanukis.

Transfer learning is usually done for tasks where your dataset has too little data to train a full-scale model from scratch.

The most common incarnation of transfer learning in the context of deep learning is the following workflow:

  1. Take layers from a previously trained model.
  2. Freeze them, so as to avoid destroying any of the information they contain during future training rounds.
  3. Add some new, trainable layers on top of the frozen layers. They will learn to turn the old features into predictions on a new dataset.
  4. Train the new layers on your dataset.

A last, optional step, is fine-tuning, which consists of unfreezing the entire model you obtained above (or part of it), and re-training it on the new data with a very low learning rate. This can potentially achieve meaningful improvements, by incrementally adapting the pretrained features to the new data.

First, we will go over the Keras trainable API in detail, which underlies most transfer learning & fine-tuning workflows.

Then, we’ll demonstrate the typical workflow by taking a model pretrained on the ImageNet dataset, and retraining it on the Kaggle “cats vs dogs” classification dataset.

This is adapted from Deep Learning with Python and the 2016 blog post “building powerful image classification models using very little data”.

Freezing layers: understanding the trainable attribute

Layers & models have three weight attributes:

  • weights is the list of all weights variables of the layer.
  • trainable_weights is the list of those that are meant to be updated (via gradient descent) to minimize the loss during training.
  • non_trainable_weights is the list of those that aren’t meant to be trained. Typically they are updated by the model during the forward pass.

Example: the Dense layer has 2 trainable weights (kernel & bias)

layer = keras.layers.Dense(3)
layer.build((None, 4))  # Create the weights

print("weights:", len(layer.weights))
print("trainable_weights:", len(layer.trainable_weights))
print("non_trainable_weights:", len(layer.non_trainable_weights))
weights: 2
trainable_weights: 2
non_trainable_weights: 0

In general, all weights are trainable weights. The only built-in layer that has non-trainable weights is the BatchNormalization layer. It uses non-trainable weights to keep track of the mean and variance of its inputs during training. To learn how to use non-trainable weights in your own custom layers, see the guide to writing new layers from scratch.

Example: the BatchNormalization layer has 2 trainable weights and 2 non-trainable weights

layer = keras.layers.BatchNormalization()
layer.build((None, 4))  # Create the weights

print("weights:", len(layer.weights))
print("trainable_weights:", len(layer.trainable_weights))
print("non_trainable_weights:", len(layer.non_trainable_weights))
weights: 4
trainable_weights: 2
non_trainable_weights: 2

Layers & models also feature a boolean attribute trainable. Its value can be changed. Setting layer.trainable to False moves all the layer’s weights from trainable to non-trainable. This is called “freezing” the layer: the state of a frozen layer won’t be updated during training (either when training with fit() or when training with any custom loop that relies on trainable_weights to apply gradient updates).

Example: setting trainable to False

layer = keras.layers.Dense(3)
layer.build((None, 4))  # Create the weights
layer.trainable = False  # Freeze the layer

print("weights:", len(layer.weights))
print("trainable_weights:", len(layer.trainable_weights))
print("non_trainable_weights:", len(layer.non_trainable_weights))
weights: 2
trainable_weights: 0
non_trainable_weights: 2

When a trainable weight becomes non-trainable, its value is no longer updated during training.

# Make a model with 2 layers
layer1 = keras.layers.Dense(3, activation="relu")
layer2 = keras.layers.Dense(3, activation="sigmoid")
model = keras.Sequential([keras.Input(shape=(3,)), layer1, layer2])

# Freeze the first layer
layer1.trainable = False

# Keep a copy of the weights of layer1 for later reference
initial_layer1_weights_values = layer1.get_weights()

# Train the model
model.compile(optimizer="adam", loss="mse")
model.fit(np.random.random((2, 3)), np.random.random((2, 3)))

# Check that the weights of layer1 have not changed during training
final_layer1_weights_values = layer1.get_weights()
np.testing.assert_allclose(
    initial_layer1_weights_values[0], final_layer1_weights_values[0]
)
np.testing.assert_allclose(
    initial_layer1_weights_values[1], final_layer1_weights_values[1]
)
1/1 [==============================] - 1s 640ms/step - loss: 0.0945

Do not confuse the layer.trainable attribute with the argument training in layer.__call__() (which controls whether the layer should run its forward pass in inference mode or training mode). For more information, see the Keras FAQ.