- 2014年，ImageNet挑戰(zhàn)賽(ILSVRC14)中，GoogLeNet(InceptionNet v1)獲得了第一名、VGG獲得了第二名，這兩類模型結(jié)構(gòu)的共同特點(diǎn)是層次更深了
InceptionNet 優(yōu)點(diǎn)
-
InceptionNet通過分支的方式增大網(wǎng)絡(luò)的寬度和深度能夠很好的提高網(wǎng)絡(luò)的性能，避免過擬合。
- 一般來說卷積網(wǎng)絡(luò)要提升表達(dá)能力，主要依靠增加輸出通道的數(shù)量，但是帶來的副作用是計(jì)算量增大和過擬合。因此InceptionNet采用分支網(wǎng)絡(luò)堆疊在一起產(chǎn)生較大通道的輸出。InceptionNet中產(chǎn)生4個(gè)網(wǎng)絡(luò)(3個(gè)卷積和1個(gè)池化)對(duì)前一層的輸出做計(jì)算，然后將不同卷積大小的輸出在通道上堆疊在一起。卷積、池化后的尺寸相同，將通道相加，一方面增加了網(wǎng)絡(luò)的寬度，另一方面同時(shí)網(wǎng)絡(luò)中的卷積和大小也不一樣，可以增加網(wǎng)絡(luò)對(duì)不同尺度的適應(yīng)性。
InceptionNet分支采用了
的卷積網(wǎng)絡(luò)，
，經(jīng)過具有256個(gè)通道的
卷積層之后(stride=1，pad=2)，輸出數(shù)據(jù)為
，其中，卷積層的參數(shù)為
。而假如上一層輸出先經(jīng)過具有32個(gè)通道的
，大約減少了4倍。
InceptionNet為了避免梯度消失，網(wǎng)絡(luò)額外增加了2個(gè)輔助的softmax用于向前傳導(dǎo)梯度（輔助分類器）。輔助分類器是將中間某一層的輸出用作分類，并按一個(gè)較小的權(quán)重（0.3）加到最終分類結(jié)果中，這樣相當(dāng)于做了模型融合，同時(shí)給網(wǎng)絡(luò)增加了反向傳播的梯度信號(hào)，也提供了額外的正則化，對(duì)于整個(gè)網(wǎng)絡(luò)的訓(xùn)練很有裨益。而在實(shí)際測(cè)試的時(shí)候，這兩個(gè)額外的softmax會(huì)被去掉。
提出并采用了 BatchNormalization(BN)。BN是一種非常有用的正則化方法，可以讓卷積網(wǎng)絡(luò)的訓(xùn)練速度加快很多，同時(shí)收斂后的分類準(zhǔn)確率也可以提高很多。BN就是神經(jīng)網(wǎng)絡(luò)在訓(xùn)練的時(shí)候?qū)γ總€(gè)Batch數(shù)據(jù)的內(nèi)部進(jìn)行標(biāo)準(zhǔn)化。BN在某種程度上還起到正則的作用，所以可以減小或者取消Dropout，優(yōu)化網(wǎng)絡(luò)結(jié)構(gòu)。同時(shí)添加BN后，還需要對(duì)超參數(shù)做一些調(diào)整，增大學(xué)速率，加快學(xué)習(xí)速率的衰減，去除Drought，減輕正則，去除LRN，對(duì)樣本進(jìn)行洗牌等等。
網(wǎng)絡(luò)最后采用了AveragePool來代替全連接層，該想法來自NIN（Network in Network），事實(shí)證明這樣可以將準(zhǔn)確率提高0.6%。但是，實(shí)際在最后還是加了一個(gè)全連接層，主要是為了方便對(duì)輸出進(jìn)行靈活調(diào)整。
InceptionNet將一個(gè)較大的卷積網(wǎng)絡(luò)拆分成兩個(gè)小的卷積網(wǎng)絡(luò)（VGGNet也使用了此思想）。比如將一個(gè) 5?5 的卷積核分成兩個(gè) 3?3 的卷積核（作用是一樣），這樣可以節(jié)約大量參數(shù)，加速運(yùn)算并減輕過擬合，同時(shí)增加了一層非線性變換拓展了模型的表達(dá)能力。
借鑒inceptionNet的思想，實(shí)現(xiàn)圖片的分類
- inception_net：原尺寸
- 卷積層：filters=32,kernel_size=(3,3),strides=(1, 1),padding='same'，卷積后尺寸:
- 池化層： pool_size=(2,2),strides=(2, 2),padding='same', 池化后尺寸：
- inception_block：
inception后的尺寸：
- inception_block：
- 池化層： pool_size=(2,2),strides=(2, 2),padding='valid', 池化后尺寸：
- inception_block：
- inception_block：
- 池化層： pool_size=(2,2),strides=(2, 2),padding='valid', 池化后尺寸：
- 卷積層：filters=32,kernel_size=(3,3),strides=(1, 1),padding='same'，卷積后尺寸:
import tensorflow as tf
import os
import pickle
import numpy as np
CIFAR_DIR = "./../cifar-10-batches-py"
print (os.listdir(CIFAR_DIR))
print(tf.__version__)
def load_data(filename):
"""read data from data file."""
with open(filename, 'rb') as f:
data = pickle.load(f, encoding='ISO8859-1')
return data['data'], data['labels']
# tensorflow.Dataset.
class CifarData:
def __init__(self, filenames, need_shuffle):
all_data = []
all_labels = []
for filename in filenames:
data, labels = load_data(filename)
all_data.append(data)
all_labels.append(labels)
self._data = np.vstack(all_data)
self._data = self._data / 127.5 - 1
self._labels = np.hstack(all_labels)
print (self._data.shape)
print (self._labels.shape)
self._num_examples = self._data.shape[0]
self._need_shuffle = need_shuffle
self._indicator = 0
if self._need_shuffle:
self._shuffle_data()
def _shuffle_data(self):
# [0,1,2,3,4,5] -> [5,3,2,4,0,1]
p = np.random.permutation(self._num_examples)
self._data = self._data[p]
self._labels = self._labels[p]
def next_batch(self, batch_size):
"""return batch_size examples as a batch."""
end_indicator = self._indicator + batch_size
if end_indicator > self._num_examples:
if self._need_shuffle:
self._shuffle_data()
self._indicator = 0
end_indicator = batch_size
else:
raise Exception("have no more examples")
if end_indicator > self._num_examples:
raise Exception("batch size is larger than all examples")
batch_data = self._data[self._indicator: end_indicator]
batch_labels = self._labels[self._indicator: end_indicator]
self._indicator = end_indicator
return batch_data, batch_labels
train_filenames = [os.path.join(CIFAR_DIR, 'data_batch_%d' % i) for i in range(1, 6)]
test_filenames = [os.path.join(CIFAR_DIR, 'test_batch')]
train_data = CifarData(train_filenames, True)
test_data = CifarData(test_filenames, False)
def inception_block(x,
output_channel_for_each_path,
name):
"""inception block implementation"""
"""
Args:
- x:
- output_channel_for_each_path: eg: [10, 20, 5]
- name:
- 圖片的寬和高沒有改變，通道數(shù)改變了(卷積的通道加池化后的通道,池化不改變通道數(shù)目)
"""
with tf.variable_scope(name):
conv1_1 = tf.layers.conv2d(x,
output_channel_for_each_path[0],
(1, 1),
strides = (1,1),
padding = 'same',
activation = tf.nn.relu,
name = 'conv1_1')
conv3_3 = tf.layers.conv2d(x,
output_channel_for_each_path[1],
(3, 3),
strides = (1,1),
padding = 'same',
activation = tf.nn.relu,
name = 'conv3_3')
conv5_5 = tf.layers.conv2d(x,
output_channel_for_each_path[2],
(5, 5),
strides = (1,1),
padding = 'same',
activation = tf.nn.relu,
name = 'conv5_5')
max_pooling = tf.layers.max_pooling2d(x,
pool_size = (2, 2),
strides = (2, 2),
padding = 'valid',
name = 'max_pooling')
max_pooling_shape = max_pooling.get_shape().as_list()[1:]
input_shape = x.get_shape().as_list()[1:]
width_padding = (input_shape[0] - max_pooling_shape[0]) // 2
height_padding = (input_shape[1] - max_pooling_shape[1]) // 2
padded_pooling = tf.pad(max_pooling,
[[0, 0],
[width_padding, width_padding],
[height_padding, height_padding],
[0, 0]])
# concat ： 指定軸拼接，其他軸的形狀要相同
concat_layer = tf.concat([conv1_1, conv3_3, conv5_5, padded_pooling],
axis = 3)
# print(conv1_1.get_shape().as_list(),
# conv3_3.get_shape().as_list(),
# conv5_5.get_shape().as_list(),
# padded_pooling.get_shape().as_list(),
# concat_layer.get_shape().as_list())
return concat_layer
x = tf.placeholder(tf.float32, [None, 3072])
y = tf.placeholder(tf.int64, [None])
# [None], eg: [0,5,6,3]
x_image = tf.reshape(x, [-1, 3, 32, 32])
# 32*32
x_image = tf.transpose(x_image, perm=[0, 2, 3, 1])
# conv1: 神經(jīng)元圖， feature_map, 輸出圖像
conv1 = tf.layers.conv2d(x_image,
32, # output channel number
(3,3), # kernel size
strides = (1, 1),
padding = 'same',
activation = tf.nn.relu,
name = 'conv1')
pooling1 = tf.layers.max_pooling2d(conv1,
pool_size = (2, 2), # kernel size
strides = (2, 2), # stride
padding = 'same',
name = 'pool1')
inception_2a = inception_block(pooling1,
[16, 16, 16],
name = 'inception_2a')
inception_2b = inception_block(inception_2a,
[16, 16, 16],
name = 'inception_2b')
pooling2 = tf.layers.max_pooling2d(inception_2b,
pool_size = (2, 2), # kernel size
strides = (2, 2), # stride
padding='valid',
name = 'pool2')
inception_3a = inception_block(pooling2,
[16, 16, 16],
name = 'inception_3a')
inception_3b = inception_block(inception_3a,
[16, 16, 16],
name = 'inception_3b')
pooling3 = tf.layers.max_pooling2d(inception_3b,
(2, 2), # kernel size
(2, 2), # stride
name = 'pool3')
flatten = tf.layers.flatten(pooling3)
y_ = tf.layers.dense(flatten, 10)
loss = tf.losses.sparse_softmax_cross_entropy(labels=y, logits=y_)
# indices
predict = tf.argmax(y_, 1)
correct_prediction = tf.equal(predict, y)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float64))
with tf.name_scope('train_op'):
train_op = tf.train.AdamOptimizer(1e-3).minimize(loss)
init = tf.global_variables_initializer()
batch_size = 20
train_steps = 1000
test_steps = 100
with tf.Session() as sess:
sess.run(init)
for i in range(train_steps):
batch_data, batch_labels = train_data.next_batch(batch_size)
loss_val, acc_val, _ = sess.run(
[loss, accuracy, train_op],
feed_dict={
x: batch_data,
y: batch_labels})
if (i+1) % 100 == 0:
print ('[Train] Step: %d, loss: %4.5f, acc: %4.5f'% (i+1, loss_val, acc_val))
if (i+1) % 1000 == 0:
test_data = CifarData(test_filenames, False)
all_test_acc_val = []
for j in range(test_steps):
test_batch_data, test_batch_labels \
= test_data.next_batch(batch_size)
test_acc_val = sess.run(
[accuracy],
feed_dict = {
x: test_batch_data,
y: test_batch_labels
})
all_test_acc_val.append(test_acc_val)
test_acc = np.mean(all_test_acc_val)
print ('[Test ] Step: %d, acc: %4.5f' % (i+1, test_acc))
