RSVP for your your local TensorFlow Everywhere event today!

Weight clustering in Keras example

View on Run in Google Colab View source on GitHub Download notebook


Welcome to the end-to-end example for weight clustering, part of the TensorFlow Model Optimization Toolkit.

Other pages

For an introduction to what weight clustering is and to determine if you should use it (including what's supported), see the overview page.

To quickly find the APIs you need for your use case (beyond fully clustering a model with 16 clusters), see the comprehensive guide.


In the tutorial, you will:

  1. Train a tf.keras model for the MNIST dataset from scratch.
  2. Fine-tune the model by applying the weight clustering API and see the accuracy.
  3. Create a 6x smaller TF and TFLite models from clustering.
  4. Create a 8x smaller TFLite model from combining weight clustering and post-training quantization.
  5. See the persistence of accuracy from TF to TFLite.


You can run this Jupyter Notebook in your local virtualenv or colab. For details of setting up dependencies, please refer to the installation guide.

 pip install -q tensorflow-model-optimization
import tensorflow as tf
from tensorflow import keras

import numpy as np
import tempfile
import zipfile
import os

Train a tf.keras model for MNIST without clustering

# Load MNIST dataset
mnist = keras.datasets.mnist
(train_images, train_labels), (test_images, test_labels) = mnist.load_data()

# Normalize the input image so that each pixel value is between 0 to 1.
train_images = train_images / 255.0
test_images  = test_images / 255.0

# Define the model architecture.
model = keras.Sequential([
    keras.layers.InputLayer(input_shape=(28, 28)),
    keras.layers.Reshape(target_shape=(28, 28, 1)),
    keras.layers.Conv2D(filters=12, kernel_size=(3, 3), activation=tf.nn.relu),
    keras.layers.MaxPooling2D(pool_size=(2, 2)),

# Train the digit classification model
Downloading data from
11493376/11490434 [==============================] - 0s 0us/step
Epoch 1/10
1688/1688 [==============================] - 7s 4ms/step - loss: 0.3008 - accuracy: 0.9148 - val_loss: 0.1216 - val_accuracy: 0.9687
Epoch 2/10
1688/1688 [==============================] - 7s 4ms/step - loss: 0.1221 - accuracy: 0.9651 - val_loss: 0.0861 - val_accuracy: 0.9758
Epoch 3/10
1688/1688 [==============================] - 7s 4ms/step - loss: 0.0897 - accuracy: 0.9741 - val_loss: 0.0710 - val_accuracy: 0.9802
Epoch 4/10
1688/1688 [==============================] - 7s 4ms/step - loss: 0.0727 - accuracy: 0.9787 - val_loss: 0.0719 - val_accuracy: 0.9803
Epoch 5/10
1688/1688 [==============================] - 7s 4ms/step - loss: 0.0631 - accuracy: 0.9808 - val_loss: 0.0657 - val_accuracy: 0.9822
Epoch 6/10
1688/1688 [==============================] - 7s 4ms/step - loss: 0.0554 - accuracy: 0.9833 - val_loss: 0.0601 - val_accuracy: 0.9820
Epoch 7/10
1688/1688 [==============================] - 7s 4ms/step - loss: 0.0489 - accuracy: 0.9855 - val_loss: 0.0647 - val_accuracy: 0.9805
Epoch 8/10
1688/1688 [==============================] - 7s 4ms/step - loss: 0.0442 - accuracy: 0.9869 - val_loss: 0.0575 - val_accuracy: 0.9845
Epoch 9/10
1688/1688 [==============================] - 7s 4ms/step - loss: 0.0403 - accuracy: 0.9875 - val_loss: 0.0596 - val_accuracy: 0.9820
Epoch 10/10
1688/1688 [==============================] - 7s 4ms/step - loss: 0.0362 - accuracy: 0.9888 - val_loss: 0.0588 - val_accuracy: 0.9833

<tensorflow.python.keras.callbacks.History at 0x7f0e6f780a58>

Evaluate the baseline model and save it for later usage

_, baseline_model_accuracy = model.evaluate(
    test_images, test_labels, verbose=0)

print('Baseline test accuracy:', baseline_model_accuracy)

_, keras_file = tempfile.mkstemp('.h5')
print('Saving model to: ', keras_file)
tf.keras.models.save_model(model, keras_file, include_optimizer=False)
Baseline test accuracy: 0.9785000085830688
Saving model to:  /tmp/tmpjo5b6jen.h5

Fine-tune the pre-trained model with clustering

Apply the cluster_weights() API to a whole pre-trained model to demonstrate its effectiveness in reducing the model size after applying zip while keeping decent accuracy. For how best to balance the accuracy and compression rate for your use case, please refer to the per layer example in the comprehensive guide.

Define the model and apply the clustering API

Before you pass the model to the clustering API, make sure it is trained and shows some acceptable accuracy.

import tensorflow_model_optimization as tfmot

cluster_weights = tfmot.clustering.keras.cluster_weights
CentroidInitialization = tfmot.clustering.keras.CentroidInitialization

clustering_params = {
  'number_of_clusters': 16,
  'cluster_centroids_init': CentroidInitialization.LINEAR

# Cluster a whole model
clustered_model = cluster_weights(model, **clustering_params)

# Use smaller learning rate for fine-tuning clustered model
opt = tf.keras.optimizers.Adam(learning_rate=1e-5)


Model: "sequential"
Layer (type)                 Output Shape              Param #   
cluster_reshape (ClusterWeig (None, 28, 28, 1)         0         
cluster_conv2d (ClusterWeigh (None, 26, 26, 12)        136       
cluster_max_pooling2d (Clust (None, 13, 13, 12)        0         
cluster_flatten (ClusterWeig (None, 2028)              0         
cluster_dense (ClusterWeight (None, 10)                20306     
Total params: 20,442
Trainable params: 54
Non-trainable params: 20,388

Fine-tune the model and evaluate the accuracy against baseline

Fine-tune the model with clustering for 1 epoch.

# Fine-tune model
108/108 [==============================] - 2s 16ms/step - loss: 0.0453 - accuracy: 0.9851 - val_loss: 0.0699 - val_accuracy: 0.9802

<tensorflow.python.keras.callbacks.History at 0x7f0e543ffeb8>

For this example, there is minimal loss in test accuracy after clustering, compared to the baseline.

_, clustered_model_accuracy = clustered_model.evaluate(
  test_images, test_labels, verbose=0)

print('Baseline test accuracy:', baseline_model_accuracy)
print('Clustered test accuracy:', clustered_model_accuracy)
Baseline test accuracy: 0.9785000085830688
Clustered test accuracy: 0.9746000170707703

Create 6x smaller models from clustering

Both strip_clustering and applying a standard compression algorithm (e.g. via gzip) are necessary to see the compression benefits of clustering.

First, create a compressible model for TensorFlow. Here, strip_clustering removes all variables (e.g. tf.Variable for storing the cluster centroids and the indices) that clustering only needs during training, which would otherwise add to model size during inference.

final_model = tfmot.clustering.keras.strip_clustering(clustered_model)

_, clustered_keras_file = tempfile.mkstemp('.h5')
print('Saving clustered model to: ', clustered_keras_file)
tf.keras.models.save_model(final_model, clustered_keras_file, 
Saving clustered model to:  /tmp/tmpo83fpb0m.h5

Then, create compressible models for TFLite. You can convert the clustered model to a format that's runnable on your targeted backend. TensorFlow Lite is an example you can use to deploy to mobile devices.

clustered_tflite_file = '/tmp/clustered_mnist.tflite'
converter = tf.lite.TFLiteConverter.from_keras_model(final_model)
tflite_clustered_model = converter.convert()
with open(clustered_tflite_file, 'wb') as f:
print('Saved clustered TFLite model to:', clustered_tflite_file)
WARNING:tensorflow:From /tmpfs/src/tf_docs_env/lib/python3.6/site-packages/tensorflow/python/training/tracking/ Model.state_updates (from is deprecated and will be removed in a future version.
Instructions for updating:
This property should not be used in TensorFlow 2.0, as updates are applied automatically.
WARNING:tensorflow:From /tmpfs/src/tf_docs_env/lib/python3.6/site-packages/tensorflow/python/training/tracking/ Layer.updates (from tensorflow.python.keras.engine.base_layer) is deprecated and will be removed in a future version.
Instructions for updating:
This property should not be used in TensorFlow 2.0, as updates are applied automatically.
INFO:tensorflow:Assets written to: /tmp/tmp4gcxcvlh/assets
Saved clustered TFLite model to: /tmp/clustered_mnist.tflite

Define a helper function to actually compress the models via gzip and measure the zipped size.

def get_gzipped_model_size(file):
  # It returns the size of the gzipped model in bytes.
  import os
  import zipfile

  _, zipped_file = tempfile.mkstemp('.zip')
  with zipfile.ZipFile(zipped_file, 'w', compression=zipfile.ZIP_DEFLATED) as f:

  return os.path.getsize(zipped_file)

Compare and see that the models are 6x smaller from clustering

print("Size of gzipped baseline Keras model: %.2f bytes" % (get_gzipped_model_size(keras_file)))
print("Size of gzipped clustered Keras model: %.2f bytes" % (get_gzipped_model_size(clustered_keras_file)))
print("Size of gzipped clustered TFlite model: %.2f bytes" % (get_gzipped_model_size(clustered_tflite_file)))
Size of gzipped baseline Keras model: 78047.00 bytes
Size of gzipped clustered Keras model: 12524.00 bytes
Size of gzipped clustered TFlite model: 12141.00 bytes

Create an 8x smaller TFLite model from combining weight clustering and post-training quantization

You can apply post-training quantization to the clustered model for additional benefits.

converter = tf.lite.TFLiteConverter.from_keras_model(final_model)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
tflite_quant_model = converter.convert()

_, quantized_and_clustered_tflite_file = tempfile.mkstemp('.tflite')

with open(quantized_and_clustered_tflite_file, 'wb') as f:

print('Saved quantized and clustered TFLite model to:', quantized_and_clustered_tflite_file)
print("Size of gzipped baseline Keras model: %.2f bytes" % (get_gzipped_model_size(keras_file)))
print("Size of gzipped clustered and quantized TFlite model: %.2f bytes" % (get_gzipped_model_size(quantized_and_clustered_tflite_file)))
INFO:tensorflow:Assets written to: /tmp/tmpt2flzp4s/assets

INFO:tensorflow:Assets written to: /tmp/tmpt2flzp4s/assets

Saved quantized and clustered TFLite model to: /tmp/tmpgu3loy72.tflite
Size of gzipped baseline Keras model: 78047.00 bytes
Size of gzipped clustered and quantized TFlite model: 9240.00 bytes

See the persistence of accuracy from TF to TFLite

Define a helper function to evaluate the TFLite model on the test dataset.

def eval_model(interpreter):
  input_index = interpreter.get_input_details()[0]["index"]
  output_index = interpreter.get_output_details()[0]["index"]

  # Run predictions on every image in the "test" dataset.
  prediction_digits = []
  for i, test_image in enumerate(test_images):
    if i % 1000 == 0:
      print('Evaluated on {n} results so far.'.format(n=i))
    # Pre-processing: add batch dimension and convert to float32 to match with
    # the model's input data format.
    test_image = np.expand_dims(test_image, axis=0).astype(np.float32)
    interpreter.set_tensor(input_index, test_image)

    # Run inference.

    # Post-processing: remove batch dimension and find the digit with highest
    # probability.
    output = interpreter.tensor(output_index)
    digit = np.argmax(output()[0])

  # Compare prediction results with ground truth labels to calculate accuracy.
  prediction_digits = np.array(prediction_digits)
  accuracy = (prediction_digits == test_labels).mean()
  return accuracy

You evaluate the model, which has been clustered and quantized, and then see the accuracy from TensorFlow persists to the TFLite backend.

interpreter = tf.lite.Interpreter(model_content=tflite_quant_model)

test_accuracy = eval_model(interpreter)

print('Clustered and quantized TFLite test_accuracy:', test_accuracy)
print('Clustered TF test accuracy:', clustered_model_accuracy)
Evaluated on 0 results so far.
Evaluated on 1000 results so far.
Evaluated on 2000 results so far.
Evaluated on 3000 results so far.
Evaluated on 4000 results so far.
Evaluated on 5000 results so far.
Evaluated on 6000 results so far.
Evaluated on 7000 results so far.
Evaluated on 8000 results so far.
Evaluated on 9000 results so far.

Clustered and quantized TFLite test_accuracy: 0.9746
Clustered TF test accuracy: 0.9746000170707703


In this tutorial, you saw how to create clustered models with the TensorFlow Model Optimization Toolkit API. More specifically, you've been through an end-to-end example for creating an 8x smaller model for MNIST with minimal accuracy difference. We encourage you to try this new capability, which can be particularly important for deployment in resource-constrained environments.