Chapter 11 – Training Deep Neural Networks
This notebook contains all the sample code and solutions to the exercises in chapter 11.
Run in Google ColabSetup¶
First, let's import a few common modules, ensure MatplotLib plots figures inline and prepare a function to save the figures. We also check that Python 3.5 or later is installed (although Python 2.x may work, it is deprecated so we strongly recommend you use Python 3 instead), as well as Scikit-Learn ≥0.20 and TensorFlow ≥2.0.
# Python ≥3.5 is required
import sys
assert sys.version_info >= (3, 5)
# Scikit-Learn ≥0.20 is required
import sklearn
assert sklearn.__version__ >= "0.20"
try:
# %tensorflow_version only exists in Colab.
%tensorflow_version 2.x
except Exception:
pass
# TensorFlow ≥2.0 is required
import tensorflow as tf
from tensorflow import keras
assert tf.__version__ >= "2.0"
%load_ext tensorboard
# Common imports
import numpy as np
import os
# to make this notebook's output stable across runs
np.random.seed(42)
# To plot pretty figures
%matplotlib inline
import matplotlib as mpl
import matplotlib.pyplot as plt
mpl.rc('axes', labelsize=14)
mpl.rc('xtick', labelsize=12)
mpl.rc('ytick', labelsize=12)
# Where to save the figures
PROJECT_ROOT_DIR = "."
CHAPTER_ID = "deep"
IMAGES_PATH = os.path.join(PROJECT_ROOT_DIR, "images", CHAPTER_ID)
os.makedirs(IMAGES_PATH, exist_ok=True)
def save_fig(fig_id, tight_layout=True, fig_extension="png", resolution=300):
path = os.path.join(IMAGES_PATH, fig_id + "." + fig_extension)
print("Saving figure", fig_id)
if tight_layout:
plt.tight_layout()
plt.savefig(path, format=fig_extension, dpi=resolution)
1. Vanishing/Exploding Gradients Problem¶
The saturating activation functions can be problematic and lead to vanishing/exploding gradients problem.
def logit(z):
return 1 / (1 + np.exp(-z))
z = np.linspace(-5, 5, 200)
plt.plot([-5, 5], [0, 0], 'k-')
plt.plot([-5, 5], [1, 1], 'k--')
plt.plot([0, 0], [-0.2, 1.2], 'k-')
plt.plot([-5, 5], [-3/4, 7/4], 'g--')
plt.plot(z, logit(z), "b-", linewidth=2)
props = dict(facecolor='black', shrink=0.1)
plt.annotate('Saturating', xytext=(3.5, 0.7), xy=(5, 1), arrowprops=props, fontsize=14, ha="center")
plt.annotate('Saturating', xytext=(-3.5, 0.3), xy=(-5, 0), arrowprops=props, fontsize=14, ha="center")
plt.annotate('Linear', xytext=(2, 0.2), xy=(0, 0.5), arrowprops=props, fontsize=14, ha="center")
plt.grid(True)
plt.title("Sigmoid activation function", fontsize=14)
plt.axis([-5, 5, -0.2, 1.2])
save_fig("sigmoid_saturation_plot")
plt.show()
1.1 Xavier and He Initialization¶
Using the Xavier or He initialization for the weights helps prevent the vanishing/exploding gradient problem.
keras.layers.Dense(10, activation="relu", kernel_initializer="he_normal")
1.2 Nonsaturating Activation Functions¶
Using non-saturating activation functions helps with the vanishing/exploding gradient problem.
Leaky ReLU¶
def leaky_relu(z, alpha=0.01):
return np.maximum(alpha*z, z)
plt.plot(z, leaky_relu(z, 0.05), "b-", linewidth=2)
plt.plot([-5, 5], [0, 0], 'k-')
plt.plot([0, 0], [-0.5, 4.2], 'k-')
plt.grid(True)
props = dict(facecolor='black', shrink=0.1)
plt.annotate('Leak', xytext=(-3.5, 0.5), xy=(-5, -0.2), arrowprops=props, fontsize=14, ha="center")
plt.title("Leaky ReLU activation function", fontsize=14)
plt.axis([-5, 5, -0.5, 4.2])
save_fig("leaky_relu_plot")
plt.show()
Let's train a neural network on Fashion MNIST using the Leaky ReLU:
(X_train_full, y_train_full), (X_test, y_test) = keras.datasets.fashion_mnist.load_data()
X_train_full = X_train_full / 255.0
X_test = X_test / 255.0
X_valid, X_train = X_train_full[:5000], X_train_full[5000:]
y_valid, y_train = y_train_full[:5000], y_train_full[5000:]
tf.random.set_seed(42)
np.random.seed(42)
model = keras.models.Sequential([
keras.layers.Flatten(input_shape=[28, 28]),
keras.layers.Dense(300, kernel_initializer="he_normal"),
keras.layers.LeakyReLU(),
keras.layers.Dense(100, kernel_initializer="he_normal"),
keras.layers.LeakyReLU(),
keras.layers.Dense(10, activation="softmax")
])
model.compile(loss="sparse_categorical_crossentropy",
optimizer=keras.optimizers.SGD(lr=1e-3),
metrics=["accuracy"])
history = model.fit(X_train, y_train, epochs=10,
validation_data=(X_valid, y_valid))
Now let's try PReLU:
tf.random.set_seed(42)
np.random.seed(42)
model = keras.models.Sequential([
keras.layers.Flatten(input_shape=[28, 28]),
keras.layers.Dense(300, kernel_initializer="he_normal"),
keras.layers.PReLU(),
keras.layers.Dense(100, kernel_initializer="he_normal"),
keras.layers.PReLU(),
keras.layers.Dense(10, activation="softmax")
])
model.compile(loss="sparse_categorical_crossentropy",
optimizer=keras.optimizers.SGD(lr=1e-3),
metrics=["accuracy"])
history = model.fit(X_train, y_train, epochs=10,
validation_data=(X_valid, y_valid))
ELU¶
def elu(z, alpha=1):
return np.where(z < 0, alpha * (np.exp(z) - 1), z)
plt.plot(z, elu(z), "b-", linewidth=2)
plt.plot([-5, 5], [0, 0], 'k-')
plt.plot([-5, 5], [-1, -1], 'k--')
plt.plot([0, 0], [-2.2, 3.2], 'k-')
plt.grid(True)
plt.title(r"ELU activation function ($\alpha=1$)", fontsize=14)
plt.axis([-5, 5, -2.2, 3.2])
save_fig("elu_plot")
plt.show()
Implementing ELU in TensorFlow is trivial, just specify the activation function when building each layer:
keras.layers.Dense(10, activation="elu")
SELU¶
This activation function was proposed in this great paper by Günter Klambauer, Thomas Unterthiner and Andreas Mayr, published in June 2017. During training, a neural network composed exclusively of a stack of dense layers using the SELU activation function and LeCun initialization will self-normalize: the output of each layer will tend to preserve the same mean and variance during training, which solves the vanishing/exploding gradients problem. As a result, this activation function outperforms the other activation functions very significantly for such neural nets, so you should really try it out. Unfortunately, the self-normalizing property of the SELU activation function is easily broken: you cannot use ℓ1 or ℓ2 regularization, regular dropout, max-norm, skip connections or other non-sequential topologies (so recurrent neural networks won't self-normalize). However, in practice it works quite well with sequential CNNs. If you break self-normalization, SELU will not necessarily outperform other activation functions.
from scipy.special import erfc
# alpha and scale to self normalize with mean 0 and standard deviation 1
# (see equation 14 in the paper):
alpha_0_1 = -np.sqrt(2 / np.pi) / (erfc(1/np.sqrt(2)) * np.exp(1/2) - 1)
scale_0_1 = (1 - erfc(1 / np.sqrt(2)) * np.sqrt(np.e)) * np.sqrt(2 * np.pi) * (2 * erfc(np.sqrt(2))*np.e**2 + np.pi*erfc(1/np.sqrt(2))**2*np.e - 2*(2+np.pi)*erfc(1/np.sqrt(2))*np.sqrt(np.e)+np.pi+2)**(-1/2)
def selu(z, scale=scale_0_1, alpha=alpha_0_1):
return scale * elu(z, alpha)
plt.plot(z, selu(z), "b-", linewidth=2)
plt.plot([-5, 5], [0, 0], 'k-')
plt.plot([-5, 5], [-1.758, -1.758], 'k--')
plt.plot([0, 0], [-2.2, 3.2], 'k-')
plt.grid(True)
plt.title("SELU activation function", fontsize=14)
plt.axis([-5, 5, -2.2, 3.2])
save_fig("selu_plot")
plt.show()
By default, the SELU hyperparameters (scale and alpha) are tuned in such a way that the mean output of each neuron remains close to 0, and the standard deviation remains close to 1 (assuming the inputs are standardized with mean 0 and standard deviation 1 too). Using this activation function, even a 1,000 layer deep neural network preserves roughly mean 0 and standard deviation 1 across all layers, avoiding the exploding/vanishing gradients problem:
np.random.seed(42)
Z = np.random.normal(size=(500, 100)) # standardized inputs
for layer in range(1000):
W = np.random.normal(size=(100, 100), scale=np.sqrt(1 / 100)) # LeCun initialization
Z = selu(np.dot(Z, W))
means = np.mean(Z, axis=0).mean()
stds = np.std(Z, axis=0).mean()
if layer % 100 == 0:
print("Layer {}: mean {:.2f}, std deviation {:.2f}".format(layer, means, stds))
Using SELU is easy:
keras.layers.Dense(10, activation="selu",
kernel_initializer="lecun_normal")
Let's create a neural net for Fashion MNIST with 100 hidden layers, using the SELU activation function:
np.random.seed(42)
tf.random.set_seed(42)
model = keras.models.Sequential()
model.add(keras.layers.Flatten(input_shape=[28, 28]))
model.add(keras.layers.Dense(300, activation="selu",
kernel_initializer="lecun_normal"))
for layer in range(99):
model.add(keras.layers.Dense(100, activation="selu",
kernel_initializer="lecun_normal"))
model.add(keras.layers.Dense(10, activation="softmax"))
model.compile(loss="sparse_categorical_crossentropy",
optimizer=keras.optimizers.SGD(lr=1e-3),
metrics=["accuracy"])
Now let's train it. Do not forget to scale the inputs to mean 0 and standard deviation 1:
pixel_means = X_train.mean(axis=0, keepdims=True)
pixel_stds = X_train.std(axis=0, keepdims=True)
X_train_scaled = (X_train - pixel_means) / pixel_stds
X_valid_scaled = (X_valid - pixel_means) / pixel_stds
X_test_scaled = (X_test - pixel_means) / pixel_stds
history = model.fit(X_train_scaled, y_train, epochs=5,
validation_data=(X_valid_scaled, y_valid))
1.3 Batch Normalization¶
Sometimes applying batch normalization before the activation function works better (there's a debate on this topic). Moreover, the layer before a BatchNormalization layer does not need to have bias terms, since the BatchNormalization layer some as well, it would be a waste of parameters, so you can set use_bias=False when creating those layers:
model = keras.models.Sequential([
keras.layers.Flatten(input_shape=[28, 28]),
keras.layers.BatchNormalization(),
keras.layers.Dense(300, use_bias=False),
keras.layers.BatchNormalization(),
keras.layers.Activation("relu"),
keras.layers.Dense(100, use_bias=False),
keras.layers.BatchNormalization(),
keras.layers.Activation("relu"),
keras.layers.Dense(10, activation="softmax")
])
model.compile(loss="sparse_categorical_crossentropy",
optimizer=keras.optimizers.SGD(lr=1e-3),
metrics=["accuracy"])
history = model.fit(X_train, y_train, epochs=10,
validation_data=(X_valid, y_valid))
1.4 Gradient Clipping¶
All Keras optimizers accept clipnorm or clipvalue arguments:
optimizer = keras.optimizers.SGD(clipvalue=1.0)
optimizer = keras.optimizers.SGD(clipnorm=1.0)
2. Reusing Pretrained Layers¶
Reusing a Keras model¶
Let's split the fashion MNIST training set in two:
X_train_A: all images of all items except for sandals and shirts (classes 5 and 6).X_train_B: a much smaller training set of just the first 200 images of sandals or shirts.
The validation set and the test set are also split this way, but without restricting the number of images.
We will train a model on set A (classification task with 8 classes), and try to reuse it to tackle set B (binary classification). We hope to transfer a little bit of knowledge from task A to task B, since classes in set A (sneakers, ankle boots, coats, t-shirts, etc.) are somewhat similar to classes in set B (sandals and shirts). However, since we are using Dense layers, only patterns that occur at the same location can be reused (in contrast, convolutional layers will transfer much better, since learned patterns can be detected anywhere on the image, as we will see in the CNN chapter).
def split_dataset(X, y):
y_5_or_6 = (y == 5) | (y == 6) # sandals or shirts
y_A = y[~y_5_or_6]
y_A[y_A > 6] -= 2 # class indices 7, 8, 9 should be moved to 5, 6, 7
y_B = (y[y_5_or_6] == 6).astype(np.float32) # binary classification task: is it a shirt (class 6)?
return ((X[~y_5_or_6], y_A),
(X[y_5_or_6], y_B))
(X_train_A, y_train_A), (X_train_B, y_train_B) = split_dataset(X_train, y_train)
(X_valid_A, y_valid_A), (X_valid_B, y_valid_B) = split_dataset(X_valid, y_valid)
(X_test_A, y_test_A), (X_test_B, y_test_B) = split_dataset(X_test, y_test)
X_train_B = X_train_B[:200]
y_train_B = y_train_B[:200]
X_train_A.shape
X_train_B.shape
y_train_A[:30]
y_train_B[:30]
tf.random.set_seed(42)
np.random.seed(42)
model_A = keras.models.Sequential()
model_A.add(keras.layers.Flatten(input_shape=[28, 28]))
for n_hidden in (300, 100, 50, 50, 50):
model_A.add(keras.layers.Dense(n_hidden, activation="selu"))
model_A.add(keras.layers.Dense(8, activation="softmax"))
model_A.compile(loss="sparse_categorical_crossentropy",
optimizer=keras.optimizers.SGD(lr=1e-3),
metrics=["accuracy"])
history = model_A.fit(X_train_A, y_train_A, epochs=20,
validation_data=(X_valid_A, y_valid_A))
model_A.save("my_model_A.h5")
model_B = keras.models.Sequential()
model_B.add(keras.layers.Flatten(input_shape=[28, 28]))
for n_hidden in (300, 100, 50, 50, 50):
model_B.add(keras.layers.Dense(n_hidden, activation="selu"))
model_B.add(keras.layers.Dense(1, activation="sigmoid"))
model_B.compile(loss="binary_crossentropy",
optimizer=keras.optimizers.SGD(lr=1e-3),
metrics=["accuracy"])
history = model_B.fit(X_train_B, y_train_B, epochs=20,
validation_data=(X_valid_B, y_valid_B))
model.summary()
model_A = keras.models.load_model("my_model_A.h5")
model_B_on_A = keras.models.Sequential(model_A.layers[:-1])
model_B_on_A.add(keras.layers.Dense(1, activation="sigmoid"))
model_A_clone = keras.models.clone_model(model_A)
model_A_clone.set_weights(model_A.get_weights())
model_B_on_A.compile(loss="binary_crossentropy",
optimizer=keras.optimizers.SGD(lr=1e-3),
metrics=["accuracy"])
history = model_B_on_A.fit(X_train_B, y_train_B, epochs=16,
validation_data=(X_valid_B, y_valid_B))
So, what's the final verdict?
model_B.evaluate(X_test_B, y_test_B)
model_B_on_A.evaluate(X_test_B, y_test_B)
Great! We got quite a bit of transfer: the error rate dropped by a factor of 4!
(100 - 96.95) / (100 - 99.25)
Faster Optimizers¶
Momentum optimization¶
optimizer = keras.optimizers.SGD(lr=0.001, momentum=0.9)
Nesterov Accelerated Gradient¶
optimizer = keras.optimizers.SGD(lr=0.001, momentum=0.9, nesterov=True)
AdaGrad¶
optimizer = keras.optimizers.Adagrad(lr=0.001)
RMSProp¶
optimizer = keras.optimizers.RMSprop(lr=0.001, rho=0.9)
Adam Optimization¶
optimizer = keras.optimizers.Adam(lr=0.001, beta_1=0.9, beta_2=0.999)
Adamax Optimization¶
optimizer = keras.optimizers.Adamax(lr=0.001, beta_1=0.9, beta_2=0.999)
Nadam Optimization¶
optimizer = keras.optimizers.Nadam(lr=0.001, beta_1=0.9, beta_2=0.999)
4. Learning Rate Scheduling¶
tf.keras schedulers¶
model = keras.models.Sequential([
keras.layers.Flatten(input_shape=[28, 28]),
keras.layers.Dense(300, activation="selu", kernel_initializer="lecun_normal"),
keras.layers.Dense(100, activation="selu", kernel_initializer="lecun_normal"),
keras.layers.Dense(10, activation="softmax")
])
s = 20 * len(X_train) // 32 # number of steps in 20 epochs (batch size = 32)
learning_rate = keras.optimizers.schedules.ExponentialDecay(0.01, s, 0.1)
optimizer = keras.optimizers.SGD(learning_rate)
model.compile(loss="sparse_categorical_crossentropy", optimizer=optimizer, metrics=["accuracy"])
n_epochs = 25
history = model.fit(X_train_scaled, y_train, epochs=n_epochs,
validation_data=(X_valid_scaled, y_valid))
5. Avoiding Overfitting Through Regularization¶
5.1 $\ell_1$ and $\ell_2$ regularization¶
# or L1(0.1) for l1 regularization with a factor of 0.1
# or L1_L2(0.1, 0.01) for both L1 and L2 regularization, with factors 0.1 and 0.01 respectively
model = keras.models.Sequential([
keras.layers.Flatten(input_shape=[28, 28]),
keras.layers.Dense(300, activation="elu",
kernel_initializer="he_normal",
kernel_regularizer=keras.regularizers.l2(0.01)),
keras.layers.Dense(100, activation="elu",
kernel_initializer="he_normal",
kernel_regularizer=keras.regularizers.l2(0.01)),
keras.layers.Dense(10, activation="softmax",
kernel_regularizer=keras.regularizers.l2(0.01))
])
model.compile(loss="sparse_categorical_crossentropy", optimizer="nadam", metrics=["accuracy"])
n_epochs = 2
history = model.fit(X_train_scaled, y_train, epochs=n_epochs,
validation_data=(X_valid_scaled, y_valid))
5.2 Dropout¶
model = keras.models.Sequential([
keras.layers.Flatten(input_shape=[28, 28]),
keras.layers.Dropout(rate=0.2),
keras.layers.Dense(300, activation="elu", kernel_initializer="he_normal"),
keras.layers.Dropout(rate=0.2),
keras.layers.Dense(100, activation="elu", kernel_initializer="he_normal"),
keras.layers.Dropout(rate=0.2),
keras.layers.Dense(10, activation="softmax")
])
model.compile(loss="sparse_categorical_crossentropy", optimizer="nadam", metrics=["accuracy"])
n_epochs = 2
history = model.fit(X_train_scaled, y_train, epochs=n_epochs,
validation_data=(X_valid_scaled, y_valid))
5.3 Max norm¶
MaxNorm constrains the weights incident to each hidden unit to have a norm less than or equal to a desired value.
model = keras.models.Sequential([
keras.layers.Flatten(input_shape=[28, 28]),
keras.layers.Dense(300, activation="selu", kernel_initializer="lecun_normal",
kernel_constraint=keras.constraints.max_norm(1.)),
keras.layers.Dense(100, activation="selu", kernel_initializer="lecun_normal",
kernel_constraint=keras.constraints.max_norm(1.)),
keras.layers.Dense(10, activation="softmax")
])
model.compile(loss="sparse_categorical_crossentropy", optimizer="nadam", metrics=["accuracy"])
n_epochs = 2
history = model.fit(X_train_scaled, y_train, epochs=n_epochs,
validation_data=(X_valid_scaled, y_valid))