Pruning in Keras example

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

Overview

Welcome to an end-to-end example for magnitude-based weight pruning.

Other pages

For an introduction to what pruning 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 pruning a model with 80% sparsity), see the comprehensive guide.

Summary

In this tutorial, you will:

  1. Train a tf.keras model for MNIST from scratch.
  2. Fine tune the model by applying the pruning API and see the accuracy.
  3. Create 3x smaller TF and TFLite models from pruning.
  4. Create a 10x smaller TFLite model from combining pruning and post-training quantization.
  5. See the persistence of accuracy from TF to TFLite.

Setup

 pip install -q tensorflow-model-optimization
WARNING: You are using pip version 20.2.2; however, version 20.2.3 is available.
You should consider upgrading via the '/tmpfs/src/tf_docs_env/bin/python -m pip install --upgrade pip' command.

import tempfile
import os

import tensorflow as tf
import numpy as np

from tensorflow import keras

%load_ext tensorboard

Train a model for MNIST without pruning

# 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='relu'),
  keras.layers.MaxPooling2D(pool_size=(2, 2)),
  keras.layers.Flatten(),
  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,
  epochs=4,
  validation_split=0.1,
)
Epoch 1/4
1688/1688 [==============================] - 7s 4ms/step - loss: 0.3422 - accuracy: 0.9004 - val_loss: 0.1760 - val_accuracy: 0.9498
Epoch 2/4
1688/1688 [==============================] - 7s 4ms/step - loss: 0.1813 - accuracy: 0.9457 - val_loss: 0.1176 - val_accuracy: 0.9698
Epoch 3/4
1688/1688 [==============================] - 7s 4ms/step - loss: 0.1220 - accuracy: 0.9648 - val_loss: 0.0864 - val_accuracy: 0.9770
Epoch 4/4
1688/1688 [==============================] - 7s 4ms/step - loss: 0.0874 - accuracy: 0.9740 - val_loss: 0.0763 - val_accuracy: 0.9787

<tensorflow.python.keras.callbacks.History at 0x7f32cbeb9550>

Evaluate baseline test accuracy and save the model 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')
tf.keras.models.save_model(model, keras_file, include_optimizer=False)
print('Saved baseline model to:', keras_file)
Baseline test accuracy: 0.972599983215332
Saved baseline model to: /tmp/tmp6quew9ig.h5

Fine-tune pre-trained model with pruning

Define the model

You will apply pruning to the whole model and see this in the model summary.

In this example, you start the model with 50% sparsity (50% zeros in weights) and end with 80% sparsity.

In the comprehensive guide, you can see how to prune some layers for model accuracy improvements.

import tensorflow_model_optimization as tfmot

prune_low_magnitude = tfmot.sparsity.keras.prune_low_magnitude

# Compute end step to finish pruning after 2 epochs.
batch_size = 128
epochs = 2
validation_split = 0.1 # 10% of training set will be used for validation set. 

num_images = train_images.shape[0] * (1 - validation_split)
end_step = np.ceil(num_images / batch_size).astype(np.int32) * epochs

# Define model for pruning.
pruning_params = {
      'pruning_schedule': tfmot.sparsity.keras.PolynomialDecay(initial_sparsity=0.50,
                                                               final_sparsity=0.80,
                                                               begin_step=0,
                                                               end_step=end_step)
}

model_for_pruning = prune_low_magnitude(model, **pruning_params)

# `prune_low_magnitude` requires a recompile.
model_for_pruning.compile(optimizer='adam',
              loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])

model_for_pruning.summary()
WARNING:tensorflow:From /tmpfs/src/tf_docs_env/lib/python3.6/site-packages/tensorflow_model_optimization/python/core/sparsity/keras/pruning_wrapper.py:220: Layer.add_variable (from tensorflow.python.keras.engine.base_layer) is deprecated and will be removed in a future version.
Instructions for updating:
Please use `layer.add_weight` method instead.
Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
prune_low_magnitude_reshape  (None, 28, 28, 1)         1         
_________________________________________________________________
prune_low_magnitude_conv2d ( (None, 26, 26, 12)        230       
_________________________________________________________________
prune_low_magnitude_max_pool (None, 13, 13, 12)        1         
_________________________________________________________________
prune_low_magnitude_flatten  (None, 2028)              1         
_________________________________________________________________
prune_low_magnitude_dense (P (None, 10)                40572     
=================================================================
Total params: 40,805
Trainable params: 20,410
Non-trainable params: 20,395
_________________________________________________________________

Train and evaluate the model against baseline

Fine tune with pruning for two epochs.

tfmot.sparsity.keras.UpdatePruningStep is required during training, and tfmot.sparsity.keras.PruningSummaries provides logs for tracking progress and debugging.

logdir = tempfile.mkdtemp()

callbacks = [
  tfmot.sparsity.keras.UpdatePruningStep(),
  tfmot.sparsity.keras.PruningSummaries(log_dir=logdir),
]
  
model_for_pruning.fit(train_images, train_labels,
                  batch_size=batch_size, epochs=epochs, validation_split=validation_split,
                  callbacks=callbacks)
Epoch 1/2
  1/422 [..............................] - ETA: 0s - loss: 0.2689 - accuracy: 0.8984WARNING:tensorflow:From /tmpfs/src/tf_docs_env/lib/python3.6/site-packages/tensorflow/python/ops/summary_ops_v2.py:1277: stop (from tensorflow.python.eager.profiler) is deprecated and will be removed after 2020-07-01.
Instructions for updating:
use `tf.profiler.experimental.stop` instead.
422/422 [==============================] - 3s 7ms/step - loss: 0.1105 - accuracy: 0.9691 - val_loss: 0.1247 - val_accuracy: 0.9682
Epoch 2/2
422/422 [==============================] - 3s 6ms/step - loss: 0.1197 - accuracy: 0.9667 - val_loss: 0.0969 - val_accuracy: 0.9763

<tensorflow.python.keras.callbacks.History at 0x7f32422a9550>

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

_, model_for_pruning_accuracy = model_for_pruning.evaluate(
   test_images, test_labels, verbose=0)

print('Baseline test accuracy:', baseline_model_accuracy) 
print('Pruned test accuracy:', model_for_pruning_accuracy)
Baseline test accuracy: 0.972599983215332
Pruned test accuracy: 0.9689000248908997

The logs show the progression of sparsity on a per-layer basis.

%tensorboard --logdir={logdir}

For non-Colab users, you can see the results of a previous run of this code block on TensorBoard.dev.

Create 3x smaller models from pruning

Both tfmot.sparsity.keras.strip_pruning and applying a standard compression algorithm (e.g. via gzip) are necessary to see the compression benefits of pruning.

  • strip_pruning is necessary since it removes every tf.Variable that pruning only needs during training, which would otherwise add to model size during inference
  • Applying a standard compression algorithm is necessary since the serialized weight matrices are the same size as they were before pruning. However, pruning makes most of the weights zeros, which is added redundancy that algorithms can utilize to further compress the model.

First, create a compressible model for TensorFlow.

model_for_export = tfmot.sparsity.keras.strip_pruning(model_for_pruning)

_, pruned_keras_file = tempfile.mkstemp('.h5')
tf.keras.models.save_model(model_for_export, pruned_keras_file, include_optimizer=False)
print('Saved pruned Keras model to:', pruned_keras_file)
Saved pruned Keras model to: /tmp/tmpu92n0irx.h5

Then, create a compressible model for TFLite.

converter = tf.lite.TFLiteConverter.from_keras_model(model_for_export)
pruned_tflite_model = converter.convert()

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

with open(pruned_tflite_file, 'wb') as f:
  f.write(pruned_tflite_model)

print('Saved pruned TFLite model to:', pruned_tflite_file)
WARNING:tensorflow:From /tmpfs/src/tf_docs_env/lib/python3.6/site-packages/tensorflow/python/training/tracking/tracking.py:111: Model.state_updates (from tensorflow.python.keras.engine.training) 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/tracking.py:111: 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/tmpunez1uhy/assets
Saved pruned TFLite model to: /tmp/tmp9oa2swr6.tflite

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

def get_gzipped_model_size(file):
  # Returns size of 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:
    f.write(file)

  return os.path.getsize(zipped_file)

Compare and see that the models are 3x smaller from pruning.

print("Size of gzipped baseline Keras model: %.2f bytes" % (get_gzipped_model_size(keras_file)))
print("Size of gzipped pruned Keras model: %.2f bytes" % (get_gzipped_model_size(pruned_keras_file)))
print("Size of gzipped pruned TFlite model: %.2f bytes" % (get_gzipped_model_size(pruned_tflite_file)))
Size of gzipped baseline Keras model: 78048.00 bytes
Size of gzipped pruned Keras model: 25680.00 bytes
Size of gzipped pruned TFlite model: 24946.00 bytes

Create a 10x smaller model from combining pruning and quantization

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

converter = tf.lite.TFLiteConverter.from_keras_model(model_for_export)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
quantized_and_pruned_tflite_model = converter.convert()

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

with open(quantized_and_pruned_tflite_file, 'wb') as f:
  f.write(quantized_and_pruned_tflite_model)

print('Saved quantized and pruned TFLite model to:', quantized_and_pruned_tflite_file)

print("Size of gzipped baseline Keras model: %.2f bytes" % (get_gzipped_model_size(keras_file)))
print("Size of gzipped pruned and quantized TFlite model: %.2f bytes" % (get_gzipped_model_size(quantized_and_pruned_tflite_file)))
INFO:tensorflow:Assets written to: /tmp/tmpf68nyuwr/assets

INFO:tensorflow:Assets written to: /tmp/tmpf68nyuwr/assets

Saved quantized and pruned TFLite model to: /tmp/tmp85dhxupl.tflite
Size of gzipped baseline Keras model: 78048.00 bytes
Size of gzipped pruned and quantized TFlite model: 7663.00 bytes

See persistence of accuracy from TF to TFLite

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

import numpy as np

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

  # Run predictions on ever y 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.
    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 pruned and quantized model and see that the accuracy from TensorFlow persists to the TFLite backend.

interpreter = tf.lite.Interpreter(model_content=quantized_and_pruned_tflite_model)
interpreter.allocate_tensors()

test_accuracy = evaluate_model(interpreter)

print('Pruned and quantized TFLite test_accuracy:', test_accuracy)
print('Pruned TF test accuracy:', model_for_pruning_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.


Pruned and quantized TFLite test_accuracy: 0.9692
Pruned TF test accuracy: 0.9689000248908997

Conclusion

In this tutorial, you saw how to create sparse models with the TensorFlow Model Optimization Toolkit API for both TensorFlow and TFLite. You then combined pruning with post-training quantization for additional benefits.

You created a 10x 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.