 Ver en TensorFlow.org |  Ejecutar en Google Colab |  Ver en GitHub |  Descargar cuaderno |
Descripción general
Este es un fin al ejemplo extremo que muestra el uso de la escasez y el grupo preservar API cuantificación conscientes de entrenamiento (PCQAT), que forma parte de la tubería de colaboración optimización del modelo de optimización de TensorFlow Toolkit.
Otras paginas
Para una introducción a la tubería y otras técnicas disponibles, consulte la página de información general de colaboración optimización .
Contenido
En el tutorial, podrá:
- Entrenar a un
tf.kerasmodelo para el conjunto de datos MNIST desde cero. - Ajuste el modelo con la poda y observe la precisión y observe que el modelo se poda con éxito.
- Aplique la dispersión preservando la agrupación en el modelo podado y observe que la escasez aplicada anteriormente se ha conservado.
- Aplique QAT y observe la pérdida de escasez y racimos.
- Aplique PCQAT y observe que tanto la escasez como la agrupación aplicada anteriormente se han conservado.
- Genere un modelo TFLite y observe los efectos de aplicarle PCQAT.
- Compare los tamaños de los diferentes modelos para observar los beneficios de la compresión de aplicar la dispersión seguida de las técnicas de optimización colaborativa de la dispersión preservando la agrupación en clústeres y PCQAT.
- Compare la precisión del modelo totalmente optimizado con la precisión del modelo de referencia no optimizado.
Configuración
Puede ejecutar este Notebook Jupyter en su local de virtualenv o colab . Para los detalles de la creación de dependencias, consulte la guía de instalación .
pip install -q tensorflow-model-optimization
import tensorflow as tf
import numpy as np
import tempfile
import zipfile
import os
Entrene un modelo tf.keras para que MNIST sea podado y agrupado
# 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)
])
opt = tf.keras.optimizers.Adam(learning_rate=1e-3)
# Train the digit classification model
model.compile(optimizer=opt,
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=['accuracy'])
model.fit(
train_images,
train_labels,
validation_split=0.1,
epochs=10
)
Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/mnist.npz 11493376/11490434 [==============================] - 0s 0us/step 11501568/11490434 [==============================] - 0s 0us/step 2021-09-02 11:14:14.164834: E tensorflow/stream_executor/cuda/cuda_driver.cc:271] failed call to cuInit: CUDA_ERROR_NO_DEVICE: no CUDA-capable device is detected Epoch 1/10 1688/1688 [==============================] - 8s 5ms/step - loss: 0.2842 - accuracy: 0.9215 - val_loss: 0.1078 - val_accuracy: 0.9713 Epoch 2/10 1688/1688 [==============================] - 8s 5ms/step - loss: 0.1110 - accuracy: 0.9684 - val_loss: 0.0773 - val_accuracy: 0.9783 Epoch 3/10 1688/1688 [==============================] - 8s 4ms/step - loss: 0.0821 - accuracy: 0.9760 - val_loss: 0.0676 - val_accuracy: 0.9803 Epoch 4/10 1688/1688 [==============================] - 8s 4ms/step - loss: 0.0684 - accuracy: 0.9799 - val_loss: 0.0600 - val_accuracy: 0.9825 Epoch 5/10 1688/1688 [==============================] - 7s 4ms/step - loss: 0.0590 - accuracy: 0.9828 - val_loss: 0.0601 - val_accuracy: 0.9838 Epoch 6/10 1688/1688 [==============================] - 7s 4ms/step - loss: 0.0522 - accuracy: 0.9845 - val_loss: 0.0599 - val_accuracy: 0.9835 Epoch 7/10 1688/1688 [==============================] - 7s 4ms/step - loss: 0.0472 - accuracy: 0.9863 - val_loss: 0.0544 - val_accuracy: 0.9862 Epoch 8/10 1688/1688 [==============================] - 7s 4ms/step - loss: 0.0422 - accuracy: 0.9868 - val_loss: 0.0579 - val_accuracy: 0.9848 Epoch 9/10 1688/1688 [==============================] - 7s 4ms/step - loss: 0.0384 - accuracy: 0.9884 - val_loss: 0.0569 - val_accuracy: 0.9847 Epoch 10/10 1688/1688 [==============================] - 7s 4ms/step - loss: 0.0347 - accuracy: 0.9892 - val_loss: 0.0559 - val_accuracy: 0.9840 <keras.callbacks.History at 0x7f6a8212c550>
Evalúe el modelo de línea de base y guárdelo para su uso posterior
_, 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: 0.9811000227928162 Saving model to: /tmp/tmprlekfdwb.h5
Pode y afine el modelo al 50% de escasez
Aplicar la prune_low_magnitude() API para lograr el modelo de podado que ha de ser agrupado en la siguiente etapa. Consulte la guía completa poda para obtener más información sobre la API de poda.
Definir el modelo y aplicar la API de dispersión
Tenga en cuenta que se utiliza el modelo previamente entrenado.
import tensorflow_model_optimization as tfmot
prune_low_magnitude = tfmot.sparsity.keras.prune_low_magnitude
pruning_params = {
'pruning_schedule': tfmot.sparsity.keras.ConstantSparsity(0.5, begin_step=0, frequency=100)
}
callbacks = [
tfmot.sparsity.keras.UpdatePruningStep()
]
pruned_model = prune_low_magnitude(model, **pruning_params)
# Use smaller learning rate for fine-tuning
opt = tf.keras.optimizers.Adam(learning_rate=1e-5)
pruned_model.compile(
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
optimizer=opt,
metrics=['accuracy'])
/tmpfs/src/tf_docs_env/lib/python3.7/site-packages/keras/engine/base_layer.py:2223: UserWarning: `layer.add_variable` is deprecated and will be removed in a future version. Please use `layer.add_weight` method instead.
warnings.warn('`layer.add_variable` is deprecated and '
Ajuste el modelo, verifique la escasez y evalúe la precisión con respecto a la línea de base
Afina el modelo con una poda de 3 épocas.
# Fine-tune model
pruned_model.fit(
train_images,
train_labels,
epochs=3,
validation_split=0.1,
callbacks=callbacks)
2021-09-02 11:15:31.836903: W tensorflow/python/util/util.cc:348] Sets are not currently considered sequences, but this may change in the future, so consider avoiding using them. Epoch 1/3 1688/1688 [==============================] - 9s 5ms/step - loss: 0.2095 - accuracy: 0.9305 - val_loss: 0.1440 - val_accuracy: 0.9528 Epoch 2/3 1688/1688 [==============================] - 8s 4ms/step - loss: 0.1042 - accuracy: 0.9671 - val_loss: 0.0947 - val_accuracy: 0.9715 Epoch 3/3 1688/1688 [==============================] - 8s 4ms/step - loss: 0.0743 - accuracy: 0.9782 - val_loss: 0.0829 - val_accuracy: 0.9770 <keras.callbacks.History at 0x7f6a81f94250>
Defina funciones auxiliares para calcular e imprimir la escasez y los grupos del modelo.
def print_model_weights_sparsity(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:
if "kernel" not in weight.name or "centroid" in weight.name:
continue
weight_size = weight.numpy().size
zero_num = np.count_nonzero(weight == 0)
print(
f"{weight.name}: {zero_num/weight_size:.2%} sparsity ",
f"({zero_num}/{weight_size})",
)
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 "
)
Primero quitemos la envoltura de poda, luego verifiquemos que los granos del modelo se podaron correctamente.
stripped_pruned_model = tfmot.sparsity.keras.strip_pruning(pruned_model)
print_model_weights_sparsity(stripped_pruned_model)
conv2d/kernel:0: 50.00% sparsity (54/108) dense/kernel:0: 50.00% sparsity (10140/20280)
Aplique la dispersión preservando la agrupación en clústeres y verifique su efecto sobre la dispersión del modelo en ambos casos
A continuación, aplique la dispersión preservando la agrupación en el modelo podado y observe el número de agrupaciones y verifique que se conserve la escasez.
import tensorflow_model_optimization as tfmot
from tensorflow_model_optimization.python.core.clustering.keras.experimental import (
cluster,
)
cluster_weights = tfmot.clustering.keras.cluster_weights
CentroidInitialization = tfmot.clustering.keras.CentroidInitialization
cluster_weights = cluster.cluster_weights
clustering_params = {
'number_of_clusters': 8,
'cluster_centroids_init': CentroidInitialization.KMEANS_PLUS_PLUS,
'preserve_sparsity': True
}
sparsity_clustered_model = cluster_weights(stripped_pruned_model, **clustering_params)
sparsity_clustered_model.compile(optimizer='adam',
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=['accuracy'])
print('Train sparsity preserving clustering model:')
sparsity_clustered_model.fit(train_images, train_labels,epochs=3, validation_split=0.1)
Train sparsity preserving clustering model: Epoch 1/3 1688/1688 [==============================] - 9s 5ms/step - loss: 0.0495 - accuracy: 0.9847 - val_loss: 0.0611 - val_accuracy: 0.9843 Epoch 2/3 1688/1688 [==============================] - 8s 5ms/step - loss: 0.0472 - accuracy: 0.9855 - val_loss: 0.0705 - val_accuracy: 0.9812 Epoch 3/3 1688/1688 [==============================] - 8s 5ms/step - loss: 0.0463 - accuracy: 0.9846 - val_loss: 0.0796 - val_accuracy: 0.9780 <keras.callbacks.History at 0x7f6a81c10250>
Quite la envoltura de agrupamiento primero, luego verifique que el modelo esté correctamente podado y agrupado.
stripped_clustered_model = tfmot.clustering.keras.strip_clustering(sparsity_clustered_model)
print("Model sparsity:\n")
print_model_weights_sparsity(stripped_clustered_model)
print("\nModel clusters:\n")
print_model_weight_clusters(stripped_clustered_model)
Model sparsity: kernel:0: 51.85% sparsity (56/108) kernel:0: 60.83% sparsity (12337/20280) Model clusters: conv2d/kernel:0: 8 clusters dense/kernel:0: 8 clusters
Aplique QAT y PCQAT y verifique el efecto en los grupos de modelos y la dispersión
A continuación, aplique QAT y PCQAT en el modelo de clúster disperso y observe que PCQAT conserva la dispersión de peso y los clústeres en su modelo. Tenga en cuenta que el modelo despojado se pasa a la API de QAT y PCQAT.
# 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)
# PCQAT
quant_aware_annotate_model = tfmot.quantization.keras.quantize_annotate_model(
stripped_clustered_model)
pcqat_model = tfmot.quantization.keras.quantize_apply(
quant_aware_annotate_model,
tfmot.experimental.combine.Default8BitClusterPreserveQuantizeScheme(preserve_sparsity=True))
pcqat_model.compile(optimizer='adam',
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=['accuracy'])
print('Train pcqat model:')
pcqat_model.fit(train_images, train_labels, batch_size=128, epochs=1, validation_split=0.1)
Train qat model: 422/422 [==============================] - 4s 8ms/step - loss: 0.0343 - accuracy: 0.9892 - val_loss: 0.0600 - val_accuracy: 0.9858 Train pcqat model: WARNING:tensorflow:Gradients do not exist for variables ['conv2d/kernel:0', 'dense/kernel:0'] when minimizing the loss. WARNING:tensorflow:Gradients do not exist for variables ['conv2d/kernel:0', 'dense/kernel:0'] when minimizing the loss. 422/422 [==============================] - 4s 8ms/step - loss: 0.0371 - accuracy: 0.9880 - val_loss: 0.0664 - val_accuracy: 0.9832 <keras.callbacks.History at 0x7f6a81792910>
print("QAT Model clusters:")
print_model_weight_clusters(qat_model)
print("\nQAT Model sparsity:")
print_model_weights_sparsity(qat_model)
print("\nPCQAT Model clusters:")
print_model_weight_clusters(pcqat_model)
print("\nPCQAT Model sparsity:")
print_model_weights_sparsity(pcqat_model)
QAT Model clusters: quant_conv2d/conv2d/kernel:0: 101 clusters quant_dense/dense/kernel:0: 18285 clusters QAT Model sparsity: conv2d/kernel:0: 7.41% sparsity (8/108) dense/kernel:0: 7.64% sparsity (1549/20280) PCQAT Model clusters: quant_conv2d/conv2d/kernel:0: 8 clusters quant_dense/dense/kernel:0: 8 clusters PCQAT Model sparsity: conv2d/kernel:0: 51.85% sparsity (56/108) dense/kernel:0: 60.84% sparsity (12338/20280)
Vea los beneficios de compresión del modelo PCQAT
Defina la función auxiliar para obtener el archivo de modelo comprimido.
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
Observe que la aplicación de escasez, agrupamiento y PCQAT a un modelo produce importantes beneficios de compresión.
# 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)
# PCQAT model
converter = tf.lite.TFLiteConverter.from_keras_model(pcqat_model)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
pcqat_tflite_model = converter.convert()
pcqat_model_file = 'pcqat_model.tflite'
# Save the model.
with open(pcqat_model_file, 'wb') as f:
f.write(pcqat_tflite_model)
print("QAT model size: ", get_gzipped_model_size(qat_model_file), ' KB')
print("PCQAT model size: ", get_gzipped_model_size(pcqat_model_file), ' KB')
WARNING:absl:Found untraced functions such as reshape_layer_call_and_return_conditional_losses, reshape_layer_call_fn, conv2d_layer_call_and_return_conditional_losses, conv2d_layer_call_fn, flatten_layer_call_and_return_conditional_losses while saving (showing 5 of 20). These functions will not be directly callable after loading. INFO:tensorflow:Assets written to: /tmp/tmp6_obh00g/assets INFO:tensorflow:Assets written to: /tmp/tmp6_obh00g/assets 2021-09-02 11:16:32.221664: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:351] Ignored output_format. 2021-09-02 11:16:32.221712: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:354] Ignored drop_control_dependency. WARNING:absl:Found untraced functions such as reshape_layer_call_and_return_conditional_losses, reshape_layer_call_fn, conv2d_layer_call_and_return_conditional_losses, conv2d_layer_call_fn, flatten_layer_call_and_return_conditional_losses while saving (showing 5 of 20). These functions will not be directly callable after loading. INFO:tensorflow:Assets written to: /tmp/tmpuqqwyk0s/assets INFO:tensorflow:Assets written to: /tmp/tmpuqqwyk0s/assets QAT model size: 13.723 KB PCQAT model size: 7.352 KB 2021-09-02 11:16:33.766310: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:351] Ignored output_format. 2021-09-02 11:16:33.766350: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:354] Ignored drop_control_dependency.
Vea la persistencia de la precisión de TF a TFLite
Defina una función auxiliar para evaluar el modelo TFLite en el conjunto de datos de prueba.
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
Evalúe el modelo, que se ha podado, agrupado y cuantificado, y luego observe que la precisión de TensorFlow persiste en el backend de TFLite.
interpreter = tf.lite.Interpreter(pcqat_model_file)
interpreter.allocate_tensors()
pcqat_test_accuracy = eval_model(interpreter)
print('Pruned, clustered and quantized TFLite test_accuracy:', pcqat_test_accuracy)
print('Baseline TF test accuracy:', baseline_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. Pruned, clustered and quantized TFLite test_accuracy: 0.9803 Baseline TF test accuracy: 0.9811000227928162
Conclusión
En este tutorial, aprendió a crear un modelo, pode mediante la prune_low_magnitude() de la API, y aplicar escasez preservar la agrupación utilizando los cluster_weights() API para preservar la escasez, mientras que la agrupación de los pesos.
A continuación, se aplicó el entrenamiento consciente de cuantificación de preservación de agrupaciones y escasez (PCQAT) para preservar la escasez de modelos y las agrupaciones mientras se usaba QAT. El modelo PCQAT final se comparó con el modelo QAT para mostrar que la escasez y los conglomerados se conservan en el primero y se pierden en el segundo.
A continuación, los modelos se convirtieron a TFLite para mostrar los beneficios de compresión de encadenar la escasez, la agrupación y las técnicas de optimización del modelo PCQAT, y se evaluó el modelo TFLite para garantizar que la precisión persista en el backend de TFLite.
Finalmente, la precisión del modelo PCQAT TFLite se comparó con la precisión del modelo de referencia previa a la optimización para mostrar que las técnicas de optimización colaborativa lograron lograr los beneficios de la compresión manteniendo una precisión similar en comparación con el modelo original.