模型優化_剪枝_Pruning_

21.剪枝 Pruning

• 此為鐵人賽系列文示範文件，參考TensorFlow Lite官方範例修改而成。
• TF Lite 評估函數參考來源。

• 剪枝 Pruning將無關緊要的權重歸零刪除歸零，在壓縮時能明顯縮小尺寸。
• 經過剪枝且量化的模型可以縮小的原來1/10大小。
• Tensorflow 模型優化模組的prune_low_magnitude()，可以將Keras模型在訓練期間將影響較小的權重修剪歸零。
• 在本範例中，您將使用與示範訓練後量化相同的基準模型進行優化。

# 建立評估模型的dict
MODEL_SIZE = {}
ACCURACY = {}

!pip install -q -U tensorflow_model_optimization

import tensorflow as tf
import tensorflow_model_optimization as tfmot
import numpy as np
import os

建立基本模型

• 模型採用tf.keras.datasets.mnist，用CNN進行建模。

# 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

def model_builder():

  keras = tf.keras

  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, activation='softmax')
  ])

  return model

baseline_model = model_builder()
baseline_model.compile(
    optimizer='adam',
    loss='sparse_categorical_crossentropy',
    metrics=['accuracy']
    )

baseline_model.summary()
baseline_model.save_weights('baseline_weights.h5')

baseline_model.fit(
    train_images, 
    train_labels, 
    epochs=1, 
    shuffle=False
    )

# 儲存未量化模型
baseline_model.save('non_pruned.h5', include_optimizer=False)

# 評估模型並紀錄準確率
_, ACCURACY['baseline Keras model'] = baseline_model.evaluate(test_images, test_labels)

# 紀錄模型大小
MODEL_SIZE['baseline h5'] = os.path.getsize('non_pruned.h5')

ACCURACY

MODEL_SIZE

使用剪枝調整模型

• 進行剪枝，另外因為剪枝模型方法有增加一層包裝層，摘要顯示的參數會增加。

# Get the pruning method
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 pruning schedule.
pruning_params = {
    'pruning_schedule': tfmot.sparsity.keras.PolynomialDecay(
        initial_sparsity=0.50,
        final_sparsity=0.80,
        begin_step=0,
        end_step=end_step)
    }

# Pass in the trained baseline model
model_for_pruning = prune_low_magnitude(
    baseline_model, 
    **pruning_params
    )

# `prune_low_magnitude` requires a recompile.
model_for_pruning.compile(
    optimizer='adam',
    loss='sparse_categorical_crossentropy',
    metrics=['accuracy']
    )

model_for_pruning.summary()

• 查看模型中某一層的權重。
  • 剪枝前，有些微弱的權重。
  • 剪枝後，其中許多將被清零。

# 剪枝前的模型權重
model_for_pruning.weights[1]

• 重新訓練模型。並在 Callback 增加tfmot.sparsity.keras.UpdatePruningStep()參數。

# Callback to update pruning wrappers at each step
callbacks=[tfmot.sparsity.keras.UpdatePruningStep()]

# Train and prune the model
model_for_pruning.fit(
    train_images, 
    train_labels,
    epochs=epochs, 
    validation_split=validation_split,
    callbacks=callbacks
    )

• 重新訓練後已修剪，觀察同一層的權重變化，許多不重要的權重已歸零。

# 剪枝後的模型權重
model_for_pruning.weights[1]

剪枝後移除包裝層

• 剪枝之後，您可以用tfmot.sparsity.keras.strip_pruning()刪除包裝層以具有與基線模型相同的層和參數。
• 此方法也有助於保存模型並導出為*.tflite檔案格式。

# Remove pruning wrappers
model_for_export = tfmot.sparsity.keras.strip_pruning(model_for_pruning)
model_for_export    .summary()

• 因為包裝器已被移除，相同的模型權重，已移置索引[0]。

model_for_export.weights[0]

• 將剪枝後的檔案保存為*.h5，此時模型與修剪前大小相同。但一旦壓縮模型則改善 相當明顯。

# Save Keras model
model_for_export.save('pruned_model.h5', include_optimizer=False)

# Get uncompressed model size of baseline and pruned models
MODEL_SIZE['pruned non quantized h5'] = os.path.getsize('pruned_model.h5')

MODEL_SIZE

模型壓縮3倍術

• 剪枝後的模型再壓縮。
• 壓縮後檔案大小約為原本1/3，這是因為剪枝後歸零的權重可以更有效的壓縮。

import tempfile
import zipfile

_, zipped_file = tempfile.mkstemp('.zip')
with zipfile.ZipFile(zipped_file, 'w', compression=zipfile.ZIP_DEFLATED) as f:
    f.write('pruned_model.h5')


MODEL_SIZE['pruned non quantized h5'] = os.path.getsize('pruned_model.h5')

MODEL_SIZE

模型壓縮10倍術

• 現在嘗試將已精剪枝後的模型再量化。
• 量化原本就會縮小約4倍，將剪枝模型壓縮後再量化，與基本模型相比，這使模型減少了約 10 倍。
• 小10倍精度還能維持水準。

# 剪枝壓縮後再量化模型
converter = tf.lite.TFLiteConverter.from_keras_model(baseline_model)
converter.optimizations = [tf.lite.Optimize.DEFAULT]

tflite_model = converter.convert()

with open('pruned_quantized.tflite', 'wb') as f:
    f.write(tflite_model)

MODEL_SIZE['pruned quantized tflite'] = os.path.getsize('pruned_quantized.tflite')
MODEL_SIZE

• 即便小十倍，精度還維持原本水準。

# A helper function to evaluate the TF Lite model using "test" dataset.
# from: https://www.tensorflow.org/lite/performance/post_training_integer_quant_16x8#evaluate_the_models
def evaluate_model(filemane):
  #Load the model into the interpreters
  interpreter = tf.lite.Interpreter(model_path=str(filemane))
  interpreter.allocate_tensors()

  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 test_image in test_images:
    # 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)

  # Compare prediction results with ground truth labels to calculate accuracy.
  accurate_count = 0
  for index in range(len(prediction_digits)):
    if prediction_digits[index] == test_labels[index]:
      accurate_count += 1
  accuracy = accurate_count * 1.0 / len(prediction_digits)

  return accuracy

# Get accuracy of pruned Keras and TF Lite models

_, ACCURACY['pruned model h5'] = model_for_pruning.evaluate(test_images, test_labels)
ACCURACY['pruned and quantized tflite'] = evaluate_model('pruned_quantized.tflite')

成果

ACCURACY

MODEL_SIZE

參考

• TensorFlow Lite官方範例。
• TF Lite 評估函數參考來源。