View on TensorFlow.org | Run in Google Colab | View on GitHub | Download notebook |

## Overview

This is an end to end example showing the usage of the **cluster preserving quantization aware training (CQAT)** API, part of the TensorFlow Model Optimization Toolkit's collaborative optimization pipeline.

### Other pages

For an introduction to the pipeline and other available techniques, see the collaborative optimization overview page.

### Contents

In the tutorial, you will:

- Train a
`tf.keras`

model for the MNIST dataset from scratch. - Fine-tune the model with clustering and see the accuracy.
- Apply QAT and observe the loss of clusters.
- Apply CQAT and observe that the clustering applied earlier has been preserved.
- Generate a TFLite model and observe the effects of applying CQAT on it.
- Compare the achieved CQAT model accuracy with a model quantized using post-training quantization.

## Setup

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
import numpy as np
import tempfile
import zipfile
import os
```

2022-12-14 12:28:30.889554: W tensorflow/compiler/xla/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer.so.7'; dlerror: libnvinfer.so.7: cannot open shared object file: No such file or directory 2022-12-14 12:28:30.889669: W tensorflow/compiler/xla/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer_plugin.so.7'; dlerror: libnvinfer_plugin.so.7: cannot open shared object file: No such file or directory 2022-12-14 12:28:30.889678: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Cannot dlopen some TensorRT libraries. If you would like to use Nvidia GPU with TensorRT, please make sure the missing libraries mentioned above are installed properly.

## Train a tf.keras model for MNIST without clustering

```
# Load MNIST dataset
mnist = tf.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
model = tf.keras.Sequential([
tf.keras.layers.InputLayer(input_shape=(28, 28)),
tf.keras.layers.Reshape(target_shape=(28, 28, 1)),
tf.keras.layers.Conv2D(filters=12, kernel_size=(3, 3),
activation=tf.nn.relu),
tf.keras.layers.MaxPooling2D(pool_size=(2, 2)),
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(10)
])
# Train the digit classification model
model.compile(optimizer='adam',
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=['accuracy'])
model.fit(
train_images,
train_labels,
validation_split=0.1,
epochs=10
)
```

2022-12-14 12:28:32.524781: E tensorflow/compiler/xla/stream_executor/cuda/cuda_driver.cc:267] failed call to cuInit: CUDA_ERROR_NO_DEVICE: no CUDA-capable device is detected Epoch 1/10 1688/1688 [==============================] - 8s 5ms/step - loss: 0.2939 - accuracy: 0.9175 - val_loss: 0.1194 - val_accuracy: 0.9688 Epoch 2/10 1688/1688 [==============================] - 8s 4ms/step - loss: 0.1226 - accuracy: 0.9647 - val_loss: 0.0931 - val_accuracy: 0.9750 Epoch 3/10 1688/1688 [==============================] - 8s 4ms/step - loss: 0.0915 - accuracy: 0.9735 - val_loss: 0.0794 - val_accuracy: 0.9803 Epoch 4/10 1688/1688 [==============================] - 8s 4ms/step - loss: 0.0761 - accuracy: 0.9771 - val_loss: 0.0712 - val_accuracy: 0.9805 Epoch 5/10 1688/1688 [==============================] - 8s 5ms/step - loss: 0.0639 - accuracy: 0.9809 - val_loss: 0.0645 - val_accuracy: 0.9825 Epoch 6/10 1688/1688 [==============================] - 8s 5ms/step - loss: 0.0564 - accuracy: 0.9830 - val_loss: 0.0639 - val_accuracy: 0.9833 Epoch 7/10 1688/1688 [==============================] - 8s 4ms/step - loss: 0.0503 - accuracy: 0.9849 - val_loss: 0.0635 - val_accuracy: 0.9828 Epoch 8/10 1688/1688 [==============================] - 7s 4ms/step - loss: 0.0450 - accuracy: 0.9860 - val_loss: 0.0575 - val_accuracy: 0.9860 Epoch 9/10 1688/1688 [==============================] - 7s 4ms/step - loss: 0.0410 - accuracy: 0.9875 - val_loss: 0.0617 - val_accuracy: 0.9848 Epoch 10/10 1688/1688 [==============================] - 8s 4ms/step - loss: 0.0376 - accuracy: 0.9887 - val_loss: 0.0573 - val_accuracy: 0.9860 <keras.callbacks.History at 0x7fe024e9d9a0>

### 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: 0.9824000000953674 Saving model to: /tmpfs/tmp/tmpsnwgg1_w.h5

## Cluster and fine-tune the model with 8 clusters

Apply the `cluster_weights()`

API to cluster the whole pre-trained model to demonstrate and observe its effectiveness in reducing the model size when applying zip, while maintaining accuracy. For how best to use the API to achieve the best compression rate while maintaining your target accuracy, refer to the clustering comprehensive guide.

### Define the model and apply the clustering API

The model needs to be pre-trained before using the clustering API.

```
import tensorflow_model_optimization as tfmot
cluster_weights = tfmot.clustering.keras.cluster_weights
CentroidInitialization = tfmot.clustering.keras.CentroidInitialization
clustering_params = {
'number_of_clusters': 8,
'cluster_centroids_init': CentroidInitialization.KMEANS_PLUS_PLUS,
'cluster_per_channel': True,
}
clustered_model = cluster_weights(model, **clustering_params)
# Use smaller learning rate for fine-tuning
opt = tf.keras.optimizers.Adam(learning_rate=1e-5)
clustered_model.compile(
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
optimizer=opt,
metrics=['accuracy'])
clustered_model.summary()
```

Model: "sequential" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= cluster_reshape (ClusterWei (None, 28, 28, 1) 0 ghts) cluster_conv2d (ClusterWeig (None, 26, 26, 12) 324 hts) cluster_max_pooling2d (Clus (None, 13, 13, 12) 0 terWeights) cluster_flatten (ClusterWei (None, 2028) 0 ghts) cluster_dense (ClusterWeigh (None, 10) 40578 ts) ================================================================= Total params: 40,902 Trainable params: 20,514 Non-trainable params: 20,388 _________________________________________________________________

### Fine-tune the model and evaluate the accuracy against baseline

Fine-tune the model with clustering for 3 epochs.

```
# Fine-tune model
clustered_model.fit(
train_images,
train_labels,
epochs=3,
validation_split=0.1)
```

Epoch 1/3 1688/1688 [==============================] - 11s 5ms/step - loss: 0.0351 - accuracy: 0.9894 - val_loss: 0.0606 - val_accuracy: 0.9830 Epoch 2/3 1688/1688 [==============================] - 8s 5ms/step - loss: 0.0316 - accuracy: 0.9912 - val_loss: 0.0581 - val_accuracy: 0.9845 Epoch 3/3 1688/1688 [==============================] - 8s 5ms/step - loss: 0.0305 - accuracy: 0.9918 - val_loss: 0.0574 - val_accuracy: 0.9845 <keras.callbacks.History at 0x7fe00a8f72b0>

Define helper functions to calculate and print the number of clustering in each kernel of the model.

```
def print_model_weight_clusters(model):
for layer in model.layers:
if isinstance(layer, tf.keras.layers.Wrapper):
weights = layer.trainable_weights
else:
weights = layer.weights
for weight in weights:
# ignore auxiliary quantization weights
if "quantize_layer" in weight.name:
continue
if "kernel" in weight.name:
unique_count = len(np.unique(weight))
print(
f"{layer.name}/{weight.name}: {unique_count} clusters "
)
```

Check that the model kernels were correctly clustered. We need to strip the clustering wrapper first.

```
stripped_clustered_model = tfmot.clustering.keras.strip_clustering(clustered_model)
print_model_weight_clusters(stripped_clustered_model)
```

conv2d/kernel:0: 96 clusters dense/kernel:0: 8 clusters

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: 0.9824000000953674 Clustered test accuracy: 0.9804999828338623

## Apply QAT and CQAT and check effect on model clusters in both cases

Next, we apply both QAT and cluster preserving QAT (CQAT) on the clustered model and observe that CQAT preserves weight clusters in your clustered model. Note that we stripped clustering wrappers from your model with `tfmot.clustering.keras.strip_clustering`

before applying CQAT API.

```
# QAT
qat_model = tfmot.quantization.keras.quantize_model(stripped_clustered_model)
qat_model.compile(optimizer='adam',
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=['accuracy'])
print('Train qat model:')
qat_model.fit(train_images, train_labels, batch_size=128, epochs=1, validation_split=0.1)
# CQAT
quant_aware_annotate_model = tfmot.quantization.keras.quantize_annotate_model(
stripped_clustered_model)
cqat_model = tfmot.quantization.keras.quantize_apply(
quant_aware_annotate_model,
tfmot.experimental.combine.Default8BitClusterPreserveQuantizeScheme())
cqat_model.compile(optimizer='adam',
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=['accuracy'])
print('Train cqat model:')
cqat_model.fit(train_images, train_labels, batch_size=128, epochs=1, validation_split=0.1)
```

WARNING:tensorflow:From /tmpfs/src/tf_docs_env/lib/python3.9/site-packages/tensorflow/python/autograph/pyct/static_analysis/liveness.py:83: Analyzer.lamba_check (from tensorflow.python.autograph.pyct.static_analysis.liveness) is deprecated and will be removed after 2023-09-23. Instructions for updating: Lambda fuctions will be no more assumed to be used in the statement where they are used, or at least in the same block. https://github.com/tensorflow/tensorflow/issues/56089 Train qat model: 422/422 [==============================] - 4s 8ms/step - loss: 0.0322 - accuracy: 0.9909 - val_loss: 0.0561 - val_accuracy: 0.9857 WARNING:root:Input layer does not contain zero weights, so apply CQAT instead. WARNING:root:Input layer does not contain zero weights, so apply CQAT instead. Train cqat model: WARNING:tensorflow:Gradients do not exist for variables ['conv2d/kernel:0', 'dense/kernel:0'] when minimizing the loss. If you're using `model.compile()`, did you forget to provide a `loss` argument? WARNING:tensorflow:Gradients do not exist for variables ['conv2d/kernel:0', 'dense/kernel:0'] when minimizing the loss. If you're using `model.compile()`, did you forget to provide a `loss` argument? WARNING:tensorflow:Gradients do not exist for variables ['conv2d/kernel:0', 'dense/kernel:0'] when minimizing the loss. If you're using `model.compile()`, did you forget to provide a `loss` argument? WARNING:tensorflow:Gradients do not exist for variables ['conv2d/kernel:0', 'dense/kernel:0'] when minimizing the loss. If you're using `model.compile()`, did you forget to provide a `loss` argument? WARNING:tensorflow:Gradients do not exist for variables ['conv2d/kernel:0', 'dense/kernel:0'] when minimizing the loss. If you're using `model.compile()`, did you forget to provide a `loss` argument? WARNING:tensorflow:Gradients do not exist for variables ['conv2d/kernel:0', 'dense/kernel:0'] when minimizing the loss. If you're using `model.compile()`, did you forget to provide a `loss` argument? WARNING:tensorflow:Gradients do not exist for variables ['conv2d/kernel:0', 'dense/kernel:0'] when minimizing the loss. If you're using `model.compile()`, did you forget to provide a `loss` argument? WARNING:tensorflow:Gradients do not exist for variables ['conv2d/kernel:0', 'dense/kernel:0'] when minimizing the loss. If you're using `model.compile()`, did you forget to provide a `loss` argument? 422/422 [==============================] - 5s 8ms/step - loss: 0.0297 - accuracy: 0.9913 - val_loss: 0.0574 - val_accuracy: 0.9852 <keras.callbacks.History at 0x7fe00a35e4c0>

```
print("QAT Model clusters:")
print_model_weight_clusters(qat_model)
print("CQAT Model clusters:")
print_model_weight_clusters(cqat_model)
```

QAT Model clusters: quant_conv2d/conv2d/kernel:0: 108 clusters quant_dense/dense/kernel:0: 19909 clusters CQAT Model clusters: quant_conv2d/conv2d/kernel:0: 96 clusters quant_dense/dense/kernel:0: 8 clusters

## See compression benefits of CQAT model

Define helper function to get zipped model file.

```
def get_gzipped_model_size(file):
# It returns the size of the gzipped model in kilobytes.
_, zipped_file = tempfile.mkstemp('.zip')
with zipfile.ZipFile(zipped_file, 'w', compression=zipfile.ZIP_DEFLATED) as f:
f.write(file)
return os.path.getsize(zipped_file)/1000
```

Note that this is a small model. Applying clustering and CQAT to a bigger production model would yield a more significant compression.

```
# QAT model
converter = tf.lite.TFLiteConverter.from_keras_model(qat_model)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
qat_tflite_model = converter.convert()
qat_model_file = 'qat_model.tflite'
# Save the model.
with open(qat_model_file, 'wb') as f:
f.write(qat_tflite_model)
# CQAT model
converter = tf.lite.TFLiteConverter.from_keras_model(cqat_model)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
cqat_tflite_model = converter.convert()
cqat_model_file = 'cqat_model.tflite'
# Save the model.
with open(cqat_model_file, 'wb') as f:
f.write(cqat_tflite_model)
print("QAT model size: ", get_gzipped_model_size(qat_model_file), ' KB')
print("CQAT model size: ", get_gzipped_model_size(cqat_model_file), ' KB')
```

WARNING:absl:Found untraced functions such as _update_step_xla, reshape_layer_call_fn, reshape_layer_call_and_return_conditional_losses, conv2d_layer_call_fn, conv2d_layer_call_and_return_conditional_losses while saving (showing 5 of 10). These functions will not be directly callable after loading. INFO:tensorflow:Assets written to: /tmpfs/tmp/tmpywi4vcdu/assets INFO:tensorflow:Assets written to: /tmpfs/tmp/tmpywi4vcdu/assets /tmpfs/src/tf_docs_env/lib/python3.9/site-packages/tensorflow/lite/python/convert.py:765: UserWarning: Statistics for quantized inputs were expected, but not specified; continuing anyway. warnings.warn("Statistics for quantized inputs were expected, but not " 2022-12-14 12:30:33.383678: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:362] Ignored output_format. 2022-12-14 12:30:33.383723: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:365] Ignored drop_control_dependency. WARNING:absl:Found untraced functions such as _update_step_xla, reshape_layer_call_fn, reshape_layer_call_and_return_conditional_losses, conv2d_layer_call_fn, conv2d_layer_call_and_return_conditional_losses while saving (showing 5 of 10). These functions will not be directly callable after loading. INFO:tensorflow:Assets written to: /tmpfs/tmp/tmp3tfs97n3/assets INFO:tensorflow:Assets written to: /tmpfs/tmp/tmp3tfs97n3/assets /tmpfs/src/tf_docs_env/lib/python3.9/site-packages/tensorflow/lite/python/convert.py:765: UserWarning: Statistics for quantized inputs were expected, but not specified; continuing anyway. warnings.warn("Statistics for quantized inputs were expected, but not " 2022-12-14 12:30:37.102259: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:362] Ignored output_format. 2022-12-14 12:30:37.102302: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:365] Ignored drop_control_dependency. QAT model size: 17.65 KB CQAT model size: 10.862 KB

## 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(f"Evaluated on {i} results so far.")
# 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.
interpreter.invoke()
# Post-processing: remove batch dimension and find the digit with highest
# probability.
output = interpreter.tensor(output_index)
digit = np.argmax(output()[0])
prediction_digits.append(digit)
print('\n')
# 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 in the TFLite backend.

```
interpreter = tf.lite.Interpreter(cqat_model_file)
interpreter.allocate_tensors()
cqat_test_accuracy = eval_model(interpreter)
print('Clustered and quantized TFLite test_accuracy:', cqat_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. INFO: Created TensorFlow Lite XNNPACK delegate for CPU. 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: 0.9821 Clustered TF test accuracy: 0.9804999828338623

## Apply post-training quantization and compare to CQAT model

Next, we use post-training quantization (no fine-tuning) on the clustered model and check its accuracy against the CQAT model. This demonstrates why you would need to use CQAT to improve the quantized model's accuracy. The difference may not be very visible, because the MNIST model is quite small and overparametrized.

First, define a generator for the callibration dataset from the first 1000 training images.

```
def mnist_representative_data_gen():
for image in train_images[:1000]:
image = np.expand_dims(image, axis=0).astype(np.float32)
yield [image]
```

Quantize the model and compare accuracy to the previously acquired CQAT model. Note that the model quantized with fine-tuning achieves higher accuracy.

```
converter = tf.lite.TFLiteConverter.from_keras_model(stripped_clustered_model)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.representative_dataset = mnist_representative_data_gen
post_training_tflite_model = converter.convert()
post_training_model_file = 'post_training_model.tflite'
# Save the model.
with open(post_training_model_file, 'wb') as f:
f.write(post_training_tflite_model)
# Compare accuracy
interpreter = tf.lite.Interpreter(post_training_model_file)
interpreter.allocate_tensors()
post_training_test_accuracy = eval_model(interpreter)
print('CQAT TFLite test_accuracy:', cqat_test_accuracy)
print('Post-training (no fine-tuning) TF test accuracy:', post_training_test_accuracy)
```

WARNING:absl:Found untraced functions such as _jit_compiled_convolution_op while saving (showing 1 of 1). These functions will not be directly callable after loading. INFO:tensorflow:Assets written to: /tmpfs/tmp/tmp0hcv3m3r/assets INFO:tensorflow:Assets written to: /tmpfs/tmp/tmp0hcv3m3r/assets /tmpfs/src/tf_docs_env/lib/python3.9/site-packages/tensorflow/lite/python/convert.py:765: UserWarning: Statistics for quantized inputs were expected, but not specified; continuing anyway. warnings.warn("Statistics for quantized inputs were expected, but not " 2022-12-14 12:30:38.838127: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:362] Ignored output_format. 2022-12-14 12:30:38.838169: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:365] Ignored drop_control_dependency. fully_quantize: 0, inference_type: 6, input_inference_type: FLOAT32, output_inference_type: FLOAT32 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. CQAT TFLite test_accuracy: 0.9821 Post-training (no fine-tuning) TF test accuracy: 0.981

## Conclusion

In this tutorial, you learned how to create a model, cluster it using the `cluster_weights()`

API, and apply the cluster preserving quantization aware training (CQAT) to preserve clusters while using QAT. The final CQAT model was compared to the QAT one to show that the clusters are preserved in the former and lost in the latter. Next, the models were converted to TFLite to show the compression benefits of chaining clustering and CQAT model optimization techniques and the TFLite model was evaluated to ensure that the accuracy persists in the TFLite backend. Finally, the CQAT model was compared to a quantized clustered model achieved using the post-training quantization API to demonstrate the advantage of CQAT in recovering the accuracy loss from normal quantization.