作者:chen_h
微信号 & QQ:862251340
微信公众号:coderpai
简书地址:https://www.jianshu.com/p/d6ac55e7b076
当我们要使用神经网络来构建一个多分类模型时,我们一般都会采用 softmax 函数来作为最后的分类函数。softmax 函数对每一个分类结果都会分配一个概率,我们把比较高的那个概率对应的类别作为模型的输出。这就是为什么我们能从模型中推导出具体分类结果。为了训练模型,我们使用 softmax 函数进行反向传播,进行训练。我们最后输出的就是一个 0-1 向量。
在这篇文章中,我们不会去解释什么是 softmax 回归或者什么是 CNN。这篇文章的主要工作是如何在 TensorFlow 上面设计一个 L2 约束的 softmax 函数,我们使用的数据集是 MNIST。完整的理论分析可以查看这篇论文。
在具体实现之前,我们先来弄清楚一些概念。
softmax 损失函数可以定义如下:
其中各个参数定义如下:
带约束的损失函数定义几乎和之前的一样,我们的目的还是最小化这个损失函数。
但是,我们需要对 f(x) 函数进行修改。
我们不是直接计算最后层权重与前一层网络输出 f(x) 之间的乘积,而是对前一层的 f(x) 先做一次归一化,然后对这个归一化的值进行 α 倍数的放大,最后我们进行常规的 softmax 函数进行计算。
也就是说,损失函数是受到如下约束:
所以,我们的架构看起来是如下图(这也是我想要实现的架构图):
C 表示卷积层,P 表示池化层,FC 表示全连接层,L2-Norm 层和Scale 层是我们重点要实现的层。
为了实现这个模型,我们使用这个代码库 进行学习。
在应用 dropout 之前,我们先对 N-1 层的输出进行正则化,然后把正则化之后的结果乘以参数 alpha,然后进行 softmax 函数计算。下面是具体的代码展示:
fc1 = alpha * tf.divide(fc1, tf.norm(fc1, ord='euclidean'))
如果我们把 alpha 设置为 0,那么这就是常规的 softmax 函数,否则就是一个 L2 约束。
完整代码如下:
# Actual Code : https://github.com/aymericdamien/TensorFlow-Examples/blob/master/notebooks/3_NeuralNetworks/convolutional_network.ipynb
# Modified By: Manash
from __future__ import division, print_function, absolute_import
# Import MNIST data
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("/tmp/data/", one_hot=False)
import tensorflow as tf
import matplotlib.pyplot as plt
import numpy as np
# Training Parameters
learning_rate = 0.001
num_steps = 100
batch_size = 20
# Network Parameters
num_input = 784 # MNIST data input (img shape: 28*28)
num_classes = 10 # MNIST total classes (0-9 digits)
dropout = 0.75 # Dropout, probability to keep units
# Create the neural network
def conv_net(x_dict, n_classes, dropout, reuse, is_training, alpha=5):
# Define a scope for reusing the variables
with tf.variable_scope('ConvNet', reuse=reuse):
# TF Estimator input is a dict, in case of multiple inputs
x = x_dict['images']
# MNIST data input is a 1-D vector of 784 features (28*28 pixels)
# Reshape to match picture format [Height x Width x Channel]
# Tensor input become 4-D: [Batch Size, Height, Width, Channel]
x = tf.reshape(x, shape=[-1, 28, 28, 1])
# Convolution Layer with 32 filters and a kernel size of 5
conv1 = tf.layers.conv2d(x, 32, 5, activation=tf.nn.relu)
# Max Pooling (down-sampling) with strides of 2 and kernel size of 2
conv1 = tf.layers.max_pooling2d(conv1, 2, 2)
# Convolution Layer with 32 filters and a kernel size of 5
conv2 = tf.layers.conv2d(conv1, 64, 3, activation=tf.nn.relu)
# Max Pooling (down-sampling) with strides of 2 and kernel size of 2
conv2 = tf.layers.max_pooling2d(conv2, 2, 2)
# Flatten the data to a 1-D vector for the fully connected layer
fc1 = tf.contrib.layers.flatten(conv2)
# Fully connected layer (in tf contrib folder for now)
fc1 = tf.layers.dense(fc1, 1024)
# If alpha is not zero then perform the l2-Normalization then scaling up
if alpha != 0:
fc1 = alpha * tf.divide(fc1, tf.norm(fc1, ord='euclidean'))
# Apply Dropout (if is_training is False, dropout is not applied)
fc1 = tf.layers.dropout(fc1, rate=dropout, training=is_training)
# Output layer, class prediction
out = tf.layers.dense(fc1, n_classes)
return out
# Define the model function (following TF Estimator Template)
def model_fn(features, labels, mode):
# Set alpha
alph = 50
# Build the neural network
# Because Dropout have different behavior at training and prediction time, we
# need to create 2 distinct computation graphs that still share the same weights.
logits_train = conv_net(features, num_classes, dropout, reuse=False, is_training=True, alpha=alph)
# At test time we don't need to normalize or scale, it's redundant as per paper : https://arxiv.org/abs/1703.09507
logits_test = conv_net(features, num_classes, dropout, reuse=True, is_training=False, alpha=0)
# Predictions
pred_classes = tf.argmax(logits_test, axis=1)
pred_probas = tf.nn.softmax(logits_test)
# If prediction mode, early return
if mode == tf.estimator.ModeKeys.PREDICT:
return tf.estimator.EstimatorSpec(mode, predictions=pred_classes)
# Define loss and optimizer
loss_op = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=logits_train, labels=tf.cast(labels, dtype=tf.int32)))
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
train_op = optimizer.minimize(loss_op, global_step=tf.train.get_global_step())
# Evaluate the accuracy of the model
acc_op = tf.metrics.accuracy(labels=labels, predictions=pred_classes)
# TF Estimators requires to return a EstimatorSpec, that specify
# the different ops for training, evaluating, ...
estim_specs = tf.estimator.EstimatorSpec(
mode=mode,
predictions=pred_classes,
loss=loss_op,
train_op=train_op,
eval_metric_ops={'accuracy': acc_op})
return estim_specs
# Build the Estimator
model = tf.estimator.Estimator(model_fn)
# Define the input function for training
input_fn = tf.estimator.inputs.numpy_input_fn(
x={'images': mnist.train.images}, y=mnist.train.labels,
batch_size=batch_size, num_epochs=None, shuffle=False)
# Train the Model
model.train(input_fn, steps=num_steps)
# Evaluate the Model
# Define the input function for evaluating
input_fn = tf.estimator.inputs.numpy_input_fn(
x={'images': mnist.test.images}, y=mnist.test.labels,
batch_size=batch_size, shuffle=False)
# Use the Estimator 'evaluate' method
model.evaluate(input_fn)
# Predict single images
n_images = 4
# Get images from test set
test_images = mnist.test.images[:n_images]
# Prepare the input data
input_fn = tf.estimator.inputs.numpy_input_fn(
x={'images': test_images}, shuffle=False)
# Use the model to predict the images class
preds = list(model.predict(input_fn))
# Display
for i in range(n_images):
plt.imshow(np.reshape(test_images[i], [28, 28]), cmap='gray')
plt.show()
print("Model prediction:", preds[i])
这个真的能提高性能吗?是的,而且效果非常好,它能提高大约 1% 的性能。我没有计算很多的迭代,主要是我没有很好的电脑。如果你对这个性能有你疑惑,你可以自己试试看。
以下是不同 alpha 值对应的模型性能:
橘黄色的线表示用常规的 softmax 函数,蓝色的线是用 L2 约束的 softmax 函数。
作者:chen_h
微信号 & QQ:862251340
简书地址:https://www.jianshu.com/p/d6ac55e7b076
CoderPai 是一个专注于算法实战的平台,从基础的算法到人工智能算法都有设计。如果你对算法实战感兴趣,请快快关注我们吧。加入AI实战微信群,AI实战QQ群,ACM算法微信群,ACM算法QQ群。长按或者扫描如下二维码,关注 “CoderPai” 微信号(coderpai)