Understanding non_trainable_weights in Keras: A Complete Guide with Examples | by Karthik Karunakaran, Ph.D. | Mar, 2025

Deep learning models often require precise control over which parameters are updated during training. In Keras, the non_trainable_weights property helps manage such parameters efficiently. Whether you’re fine-tuning a pre-trained model or implementing custom layers, knowing how to use non_trainable_weights correctly can improve performance and flexibility.

Keras models and layers have two types of weight attributes:

• Trainable Weights: These are updated during backpropagation.
• Non-Trainable Weights: These remain constant during training, useful for storing fixed parameters like statistics in batch normalization.

The non_trainable_weights property allows access to these parameters, ensuring they are used without being modified by gradient updates.

• Pre-trained Model Fine-Tuning: Freezing layers to retain learned features.
• Custom Layers: Defining stateful layers with fixed parameters.
• Efficiency: Reducing the number of trainable parameters optimizes training speed.

To access non-trainable weights in a model, use:

from tensorflow import keras
# Load a pre-trained model
base_model = keras.applications.MobileNetV2(weights='imagenet', include_top=False)
# Freeze all layers
for layer in base_model.layers:
layer.trainable = False
print("Non-trainable weights:", base_model.non_trainable_weights)

Here, all layers are frozen, making their weights part of non_trainable_weights.

You can define a custom layer with fixed parameters:

import tensorflow as tf
class CustomLayer(keras.layers.Layer):
def __init__(self, **kwargs):
super().__init__(**kwargs)
self.fixed_weight = self.add_weight(shape=(1,), initializer="ones", trainable=False)
def call(self, inputs):
return inputs * self.fixed_weight
layer = CustomLayer()
print("Non-trainable weights:", layer.non_trainable_weights)

Here, fixed_weight remains unchanged during training.

Even though they are not updated automatically, non-trainable weights can be manually modified:

layer.fixed_weight.assign([2.0])
print("Updated non-trainable weight:", layer.fixed_weight.numpy())

1. Use for Frozen Layers: When fine-tuning pre-trained models, set trainable = False for layers.
2. Manually Update When Needed: If updates are required, assign values explicitly.
3. Monitor Parameter Count: Use model.summary() to check trainable vs. non-trainable parameters.

Understanding and leveraging non_trainable_weights in Keras is crucial for optimizing deep learning workflows. Whether you’re customizing layers or fine-tuning models, managing trainable and non-trainable weights can significantly enhance model efficiency.

Interested in mastering AI and deep learning? Check out my Udemy courses: Karthik K on Udemy

Share your thoughts in the comments! Have you used non_trainable_weights before? How did it impact your model training?