Ver en TensorFlow.org

Ejecutar en Google Colab

Ver fuente en GitHub

Descargar cuaderno

Visión general

Este cuaderno demostrará cómo utilizar la capa de Normalización de peso y cómo puede mejorar la convergencia.

Peso Normalización

Una simple reparametrización para acelerar el entrenamiento de redes neuronales profundas:

Tim Salimans, Diederik P. Kingma (2016)

Al volver a parametrizar los pesos de esta manera, se mejora el acondicionamiento del problema de optimización y se acelera la convergencia del descenso de gradiente estocástico. Nuestra reparametrización se inspira en la normalización por lotes, pero no introduce ninguna dependencia entre los ejemplos de un minibatch. Esto significa que nuestro método también se puede aplicar con éxito a modelos recurrentes como LSTM y a aplicaciones sensibles al ruido, como el aprendizaje por refuerzo profundo o los modelos generativos, para los que la normalización por lotes es menos adecuada. Aunque nuestro método es mucho más simple, aún proporciona gran parte de la aceleración de la normalización por lotes completa. Además, la sobrecarga computacional de nuestro método es menor, lo que permite tomar más pasos de optimización en la misma cantidad de tiempo.

https://arxiv.org/abs/1602.07868

Configuración

pip install -q -U tensorflow-addons

import tensorflow as tf
import tensorflow_addons as tfa

import numpy as np
from matplotlib import pyplot as plt

# Hyper Parameters
batch_size = 32
epochs = 10
num_classes=10

Construir modelos

# Standard ConvNet
reg_model = tf.keras.Sequential([
    tf.keras.layers.Conv2D(6, 5, activation='relu'),
    tf.keras.layers.MaxPooling2D(2, 2),
    tf.keras.layers.Conv2D(16, 5, activation='relu'),
    tf.keras.layers.MaxPooling2D(2, 2),
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(120, activation='relu'),
    tf.keras.layers.Dense(84, activation='relu'),
    tf.keras.layers.Dense(num_classes, activation='softmax'),
])

# WeightNorm ConvNet
wn_model = tf.keras.Sequential([
    tfa.layers.WeightNormalization(tf.keras.layers.Conv2D(6, 5, activation='relu')),
    tf.keras.layers.MaxPooling2D(2, 2),
    tfa.layers.WeightNormalization(tf.keras.layers.Conv2D(16, 5, activation='relu')),
    tf.keras.layers.MaxPooling2D(2, 2),
    tf.keras.layers.Flatten(),
    tfa.layers.WeightNormalization(tf.keras.layers.Dense(120, activation='relu')),
    tfa.layers.WeightNormalization(tf.keras.layers.Dense(84, activation='relu')),
    tfa.layers.WeightNormalization(tf.keras.layers.Dense(num_classes, activation='softmax')),
])

Cargar datos

(x_train, y_train), (x_test, y_test) = tf.keras.datasets.cifar10.load_data()

# Convert class vectors to binary class matrices.
y_train = tf.keras.utils.to_categorical(y_train, num_classes)
y_test = tf.keras.utils.to_categorical(y_test, num_classes)

x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train /= 255
x_test /= 255

Downloading data from https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
170500096/170498071 [==============================] - 11s 0us/step

Modelos de trenes

reg_model.compile(optimizer='adam', 
                  loss='categorical_crossentropy',
                  metrics=['accuracy'])

reg_history = reg_model.fit(x_train, y_train,
                            batch_size=batch_size,
                            epochs=epochs,
                            validation_data=(x_test, y_test),
                            shuffle=True)

Epoch 1/10
1563/1563 [==============================] - 9s 4ms/step - loss: 1.8336 - accuracy: 0.3253 - val_loss: 1.4039 - val_accuracy: 0.4957
Epoch 2/10
1563/1563 [==============================] - 5s 3ms/step - loss: 1.3773 - accuracy: 0.5039 - val_loss: 1.3419 - val_accuracy: 0.5309
Epoch 3/10
1563/1563 [==============================] - 5s 3ms/step - loss: 1.2510 - accuracy: 0.5497 - val_loss: 1.2108 - val_accuracy: 0.5710
Epoch 4/10
1563/1563 [==============================] - 5s 3ms/step - loss: 1.1606 - accuracy: 0.5858 - val_loss: 1.2134 - val_accuracy: 0.5687
Epoch 5/10
1563/1563 [==============================] - 5s 3ms/step - loss: 1.0971 - accuracy: 0.6100 - val_loss: 1.1534 - val_accuracy: 0.5880
Epoch 6/10
1563/1563 [==============================] - 5s 3ms/step - loss: 1.0420 - accuracy: 0.6296 - val_loss: 1.1944 - val_accuracy: 0.5865
Epoch 7/10
1563/1563 [==============================] - 5s 3ms/step - loss: 1.0014 - accuracy: 0.6445 - val_loss: 1.1386 - val_accuracy: 0.6012
Epoch 8/10
1563/1563 [==============================] - 5s 3ms/step - loss: 0.9550 - accuracy: 0.6623 - val_loss: 1.1659 - val_accuracy: 0.6020
Epoch 9/10
1563/1563 [==============================] - 5s 3ms/step - loss: 0.9196 - accuracy: 0.6737 - val_loss: 1.1539 - val_accuracy: 0.6027
Epoch 10/10
1563/1563 [==============================] - 5s 3ms/step - loss: 0.8768 - accuracy: 0.6889 - val_loss: 1.1509 - val_accuracy: 0.6029

wn_model.compile(optimizer='adam', 
                 loss='categorical_crossentropy',
                 metrics=['accuracy'])

wn_history = wn_model.fit(x_train, y_train,
                          batch_size=batch_size,
                          epochs=epochs,
                          validation_data=(x_test, y_test),
                          shuffle=True)

Epoch 1/10
1563/1563 [==============================] - 14s 8ms/step - loss: 1.8195 - accuracy: 0.3319 - val_loss: 1.4563 - val_accuracy: 0.4721
Epoch 2/10
1563/1563 [==============================] - 10s 7ms/step - loss: 1.4049 - accuracy: 0.4937 - val_loss: 1.3051 - val_accuracy: 0.5301
Epoch 3/10
1563/1563 [==============================] - 10s 6ms/step - loss: 1.2669 - accuracy: 0.5461 - val_loss: 1.2858 - val_accuracy: 0.5425
Epoch 4/10
1563/1563 [==============================] - 10s 6ms/step - loss: 1.1622 - accuracy: 0.5868 - val_loss: 1.2278 - val_accuracy: 0.5587
Epoch 5/10
1563/1563 [==============================] - 10s 6ms/step - loss: 1.0782 - accuracy: 0.6175 - val_loss: 1.1755 - val_accuracy: 0.5825
Epoch 6/10
1563/1563 [==============================] - 10s 6ms/step - loss: 1.0280 - accuracy: 0.6383 - val_loss: 1.1772 - val_accuracy: 0.5827
Epoch 7/10
1563/1563 [==============================] - 10s 6ms/step - loss: 0.9705 - accuracy: 0.6527 - val_loss: 1.1542 - val_accuracy: 0.5895
Epoch 8/10
1563/1563 [==============================] - 10s 6ms/step - loss: 0.9291 - accuracy: 0.6695 - val_loss: 1.1680 - val_accuracy: 0.5924
Epoch 9/10
1563/1563 [==============================] - 10s 6ms/step - loss: 0.8837 - accuracy: 0.6884 - val_loss: 1.1302 - val_accuracy: 0.6039
Epoch 10/10
1563/1563 [==============================] - 10s 6ms/step - loss: 0.8437 - accuracy: 0.7029 - val_loss: 1.1593 - val_accuracy: 0.6018

reg_accuracy = reg_history.history['accuracy']
wn_accuracy = wn_history.history['accuracy']

plt.plot(np.linspace(0, epochs,  epochs), reg_accuracy,
             color='red', label='Regular ConvNet')

plt.plot(np.linspace(0, epochs, epochs), wn_accuracy,
         color='blue', label='WeightNorm ConvNet')

plt.title('WeightNorm Accuracy Comparison')
plt.legend()
plt.grid(True)
plt.show()