根据网络结构图自搭了个InceptionV3网络，算是熟悉网络的搭建，先上代码：
# coding:utf-8
# Author : hiicy redldw
# Date : 2019/04/17
import os
from sklearn.model_selection import train_test_split
import tensorflow as tf
import matplotlib.pyplot as plt
def Conv2dBn(input, kernel_size=(3, 3), nfilter=32, stride=(1, 1), padding="valid", activation='relu', name=None):
"""
卷积层
"""
x = tf.keras.layers.Conv2D(nfilter, kernel_size, stride, padding, activation=activation, name=name)(input)
x = tf.keras.layers.BatchNormalization(axis=3)(x) # axis=3 设置特征轴
return x
def Pool(input, pool_size=(3, 3), stride=(1, 1), pool_type='avg',padding='valid',name=None):
if str.lower(pool_type) == "avg":
x = tf.keras.layers.AveragePooling2D(pool_size, stride,padding=padding, name=name)(input)
elif str.lower(pool_type) == 'max':
x = tf.keras.layers.MaxPooling2D(pool_size, stride,padding=padding, name=name)(input)
return x
class MyInceptionV3(object):
def __init__(self, shape, classes=3,ckptpath=""):
self.shape = shape
self.classes = classes
self.layers = None
self.model = None
self.ckptPath = ckptpath
def build(self):
base_input = tf.keras.layers.Input(shape=self.shape) # 299*299*3
with tf.name_scope("first") as first_scope:
x = Conv2dBn(base_input, stride=(2, 2)) # 149*149*32
with tf.name_scope("second") as second_scope:
x = Conv2dBn(x)#147*147*32
with tf.name_scope("third") as third_scope:
x = Conv2dBn(x, nfilter=64,padding='same') #147*147*64
with tf.name_scope("fourth") as fourth_scope:
x = Pool(x, pool_type="max", stride=(2, 2)) #73*73*64
with tf.name_scope("fifth") as fifth_scope:
x = Conv2dBn(x, kernel_size=(1, 1), nfilter=80) #73*73*80
with tf.name_scope("sixth") as sixth_scope:
x = Conv2dBn(x, nfilter=192) # 71*71*192
with tf.name_scope("seventh") as seventh_scope:
x = Pool(x, pool_type="max", stride=(2, 2)) # 35*35*192
with tf.name_scope("block1") as block1:
with tf.name_scope("module1") as module1:
# FIXME: 利用branch一条一条构通
branch1 = Conv2dBn(x, kernel_size=(1, 1), nfilter=64) #35*35*64
branch2 = Conv2dBn(x, kernel_size=(1, 1), nfilter=48) #35*35*48
branch2 = Conv2dBn(branch2, kernel_size=(5, 5), nfilter=64,padding='same') #35*35*64
branch3 = Conv2dBn(x, kernel_size=(1, 1), nfilter=64)#35*35*64 # 其实可以借用branch1，但这里说明清楚就多写一个
branch3 = Conv2dBn(branch3, kernel_size=(3, 3), nfilter=96,padding='same') #35*35*96
branch3 = Conv2dBn(branch3, kernel_size=(3, 3), nfilter=96,padding='same') #35*35*96
branch4 = Pool(x, pool_type='avg',padding='same')#35*35*192
branch4 = Conv2dBn(branch4, kernel_size=(1,1), nfilter=32) #35*35*32
x = tf.keras.layers.concatenate(
[branch1, branch2, branch3, branch4],
axis=3,
) # 35*35*256
print(x)
with tf.name_scope("module2") as module2:
branch1 = Conv2dBn(x, kernel_size=(1, 1), nfilter=64) #35*35*64
branch2 = Conv2dBn(x, kernel_size=(1, 1), nfilter=48) # 35*35*48
branch2 = Conv2dBn(branch2, kernel_size=(5, 5), nfilter=64,padding='same') #35*35*64
branch3 = Conv2dBn(x, kernel_size=(1, 1), nfilter=64)#35*35*64 其实可以借用blockModule1_1，但这里说明清楚就多写一个
branch3 = Conv2dBn(branch3, kernel_size=(3, 3), nfilter=96,padding='same')#35*35*96
branch3 = Conv2dBn(branch3, kernel_size=(3, 3), nfilter=96,padding='same')#35*35*96
branch4 = Pool(x, pool_type='avg',padding='same') # 35*35*256
branch4 = Conv2dBn(branch4, kernel_size=(1, 1), nfilter=64) #35*35*64
x = tf.keras.layers.concatenate(
[branch1, branch2, branch3, branch4],
axis=3
) # 35*35*288
with tf.name_scope("module3") as module3:
branch1 = Conv2dBn(x, kernel_size=(1, 1), nfilter=64) #35*35*64
branch2 = Conv2dBn(x, kernel_size=(1, 1), nfilter=48) # 35*35*48
branch2 = Conv2dBn(branch2, kernel_size=(5, 5), nfilter=64,padding='same') #35*35*64
branch3 = Conv2dBn(x, kernel_size=(1, 1), nfilter=64)# 35*35*64 其实可以借用blockModule1_1，但这里说明清楚就多写一个
branch3 = Conv2dBn(branch3, kernel_size=(3, 3), nfilter=96,padding='same') #35*35*96
branch3 = Conv2dBn(branch3, kernel_size=(3, 3), nfilter=96,padding='same') # 35*35*96
branch4 = Pool(x, pool_type='avg',padding='same') # 35*35*288
branch4 = Conv2dBn(branch4, kernel_size=(1, 1), nfilter=64) # 35*35*64
x = tf.keras.layers.concatenate(
[branch1, branch2, branch3, branch4],
axis=3
) # 35*35*288
with tf.name_scope("block2") as block2:
with tf.name_scope("module1") as module1:
branch1 = Conv2dBn(x, nfilter=384, stride=(2, 2)) # 17*17*384
branch2 = Conv2dBn(x, kernel_size=(1, 1), nfilter=64) # 35*35*64
branch2 = Conv2dBn(branch2, nfilter=96,padding='same') # 35*35*96
branch2 = Conv2dBn(branch2, nfilter=96, stride=(2, 2)) # 17*17*96
branch3 = Pool(x,pool_type='max', stride=(2, 2)) # 17*17*288
x = tf.keras.layers.concatenate(
[branch1, branch2, branch3],
axis=3,
) # 17*17*768
with tf.name_scope('module2') as module2:
branch1 = Conv2dBn(x, kernel_size=(1, 1), nfilter=192) # 17*17*192
branch2 = Conv2dBn(x, kernel_size=(1, 1), nfilter=128) # 17*17*128
branch2 = Conv2dBn(branch2, kernel_size=(1, 7), nfilter=128,padding='same') # 17*17*128
branch2 = Conv2dBn(branch2, kernel_size=(7, 1), nfilter=192,padding='same') # 17*17*192
branch3 = Conv2dBn(x, kernel_size=(1, 1), nfilter=128) # 17*17*768
# pad:kernel_size // 2
branch3 = Conv2dBn(branch3, kernel_size=(7, 1), nfilter=128,padding='same') # 17*17*128
branch3 = Conv2dBn(branch3, kernel_size=(1, 7), nfilter=128,padding='same') # 17*17*128
branch3 = Conv2dBn(branch3, kernel_size=(7, 1), nfilter=128,padding='same') # 17*17*128
branch3 = Conv2dBn(branch3, kernel_size=(1, 7), nfilter=192,padding='same') # 17*17*192
branch4 = Pool(x, pool_type='avg',padding='same') # 17*17*768
branch4 = Conv2dBn(branch4, kernel_size=(1, 1), nfilter=192) # 17*17*192
x = tf.keras.layers.concatenate(
[branch1, branch2, branch3,branch4],
axis=3,
) # 17*17*768
for i in [3,4]:
with tf.name_scope(f'module{i}'):
branch1 = Conv2dBn(x, kernel_size=(1, 1), nfilter=192) # 17*17*192
branch2 = Conv2dBn(x, kernel_size=(1, 1), nfilter=160) # 17*17*160
branch2 = Conv2dBn(branch2, kernel_size=(1, 7), nfilter=160,padding='same') # 17*17*160
branch2 = Conv2dBn(branch2, kernel_size=(7, 1), nfilter=192,padding='same') # 17*17*192
branch3 = Conv2dBn(x, kernel_size=(1, 1), nfilter=160) # 17*17*160
branch3 = Conv2dBn(branch3, kernel_size=(7, 1), nfilter=160,padding='same') # 17*17*160
branch3 = Conv2dBn(branch3, kernel_size=(1, 7), nfilter=160,padding='same') # 17*17*160
branch3 = Conv2dBn(branch3, kernel_size=(7, 1), nfilter=160,padding='same') # 17*17*160
branch3 = Conv2dBn(branch3, kernel_size=(1, 7), nfilter=192,padding='same') # 17*17*192
branch4 = Pool(x,pool_type='avg',padding='same') # 17*17*768
branch4 = Conv2dBn(branch4, kernel_size=(1, 1), nfilter=192) # 17*17*192
x = tf.keras.layers.concatenate(
[branch1, branch2, branch3, branch4],
axis=3
) # 17*17*768
with tf.name_scope("module5") as module5:
branch1 = Conv2dBn(x, kernel_size=(1, 1), nfilter=192) # 17*17*192
branch2 = Conv2dBn(x, kernel_size=(1, 1), nfilter=192) # 17*17*192
branch2 = Conv2dBn(branch2, kernel_size=(1, 7), nfilter=192,padding='same')# 17*17*192
branch2 = Conv2dBn(branch2, kernel_size=(7, 1), nfilter=192,padding='same')# 17*17*192
branch3 = Conv2dBn(x, kernel_size=(1, 1), nfilter=192)# 17*17*192
branch3 = Conv2dBn(branch3, kernel_size=(7, 1), nfilter=192,padding='same')# 17*17*192
branch3 = Conv2dBn(branch3, kernel_size=(1, 7), nfilter=192,padding='same')# 17*17*192
branch3 = Conv2dBn(branch3, kernel_size=(7, 1), nfilter=192,padding='same')# 17*17*192
branch3 = Conv2dBn(branch3, kernel_size=(1, 7), nfilter=192,padding='same')# 17*17*192
branch4 = Pool(x,pool_type='avg',padding='same') # 17*17*192
branch4 = Conv2dBn(branch4, kernel_size=(1, 1), nfilter=192) # 17*17*192
x = tf.keras.layers.concatenate(
[branch1, branch2, branch3, branch4],
axis=3
) # 17*17*768
with tf.name_scope("block3") as block3:
with tf.name_scope('module1'):
branch1 = Conv2dBn(x, kernel_size=(1, 1), nfilter=192) # 17*17*192
branch1 = Conv2dBn(branch1, nfilter=320, stride=(2, 2)) # 8*8*320
branch2 = Conv2dBn(x, kernel_size=(1, 1), nfilter=192) # 17*17*192
branch2 = Conv2dBn(branch2, kernel_size=(1, 7), nfilter=192,padding='same') # 17*17*192
branch2 = Conv2dBn(branch2, kernel_size=(7, 1), nfilter=192,padding='same') # 17*17*192
branch2 = Conv2dBn(branch2, nfilter=192, stride=(2, 2)) # 8*8*192
branch3 = Pool(x,pool_type='max',stride=(2,2)) # 8*8*768
x = tf.keras.layers.concatenate(
[branch1,branch2,branch3],axis=3
) #8*8*1280
for i in [2,3]:
with tf.name_scope(f'module{i}'):
branch1 = Conv2dBn(x, kernel_size=(1, 1), nfilter=320) # 8*8*320
branch2 = Conv2dBn(x, kernel_size=(1, 1), nfilter=384) # 8*8*384
branch2_1 = Conv2dBn(branch2, kernel_size=(1, 3), nfilter=384,padding='same') # 8*8*384
branch2_2 = Conv2dBn(branch2, kernel_size=(3, 1), nfilter=384,padding='same') # 8*8*384
branch2 = tf.keras.layers.concatenate([branch2_1,branch2_2],axis=3) #8*8*768
branch3 = Conv2dBn(x, kernel_size=(1, 1), nfilter=448) # 8*8*448
branch3 = Conv2dBn(branch3,nfilter=384,padding='same') #8*8*384
branch3_1 = Conv2dBn(branch3, kernel_size=(1, 3), nfilter=384,padding='same') # 8*8*384
branch3_2 = Conv2dBn(branch3, kernel_size=(3, 1), nfilter=384,padding='same') # 8*8*384
branch3 = tf.keras.layers.concatenate([branch3_1,branch3_2],axis=3) # 8*8*768
branch4 = Pool(x,pool_type='avg',padding='same') #6*6*1280
branch4 = Conv2dBn(branch4, kernel_size=(1, 1), nfilter=192) # 8*8*192
x = tf.keras.layers.concatenate(
[branch1,branch2,branch3,branch4],
axis=3
) # 8*8*2048
with tf.name_scope('output'):
x = tf.keras.layers.GlobalAveragePooling2D()(x) # 2048
pred = tf.keras.layers.Dense(self.classes, activation='softmax')(x) # 8
self.model = tf.keras.Model(inputs=base_input,outputs=pred)
optimizer = tf.keras.optimizers.RMSprop()
self.model.compile(loss="categorical_crossentropy",
metrics=['acc'],
optimizer=optimizer)
print(self.model.summary())
return
def predict(self,ModelPath):
if self.model is None:
self.model = tf.keras.models.load_model(ModelPath)
def train(self,x,y,batch=32,epoch=10,is_save=True):
assert self.model is not None
train_X, test_X, train_y, test_y = train_test_split(x,
y,
test_size=0.2,
random_state=20)
self.ckptDir = os.path.dirname(self.ckptPath)
# 保存检查点
# cpCallBack = tf.keras.callbacks.ModelCheckpoint(self.ckptPath,
# save_weights_only=False,
# period=6)
# FIXME: 用生成器方式
history = self.model.fit(x,y,batch,epoch,verbose=1,
shuffle=True,validation_split=0.2)
predictions = self.model.predict(test_X, verbose=0)
loss, accuracy = self.model.evaluate(test_X,test_y, batch_size=batch)
print('test loss: ', loss)
print('test accuracy: ', accuracy)
self._plotLoss(history)
if is_save:
self.model.save(self.ckptPath)
def _plotLoss(self,history):
acc = history.history['acc']
val_acc = history.history['val_acc']
loss = history.history['loss']
val_loss = history.history['val_loss']
x = [i+1 for i in range(len(loss))]
y = [i+1 for i in range(len(val_loss))]
plt.clf()
fig:plt.Figure = plt.figure()
ax1 = fig.add_subplot(121)
plt.plot(x,loss,marker='o',c='c',mc='g',lw=2,label="训练损失")
plt.plot(y,val_loss,marker='o',c='r',mc='#D64474',label="验证损失")
plt.sca(ax1)
ax2 = fig.add_subplot(122)
plt.plot(x, acc, marker='o', c='c', mc='g', lw=2, label="训练精度")
plt.plot(y, val_acc, marker='o', c='r', mc='#D64474', label="验证精度")
plt.sca(ax2)
plt.savefig("imshow.jpg")
def __str__(self):
return "my inception v3!"
参考:
经典网络结构GoogleNet之Inception-v1 v2 v3 v4 Resnet
Google Inception Net介绍及Inception V3结构分析
深度神经网络Google Inception Net-V3结构图
对比它们的网络图发现，与keras的网络结构有些许冲突，我还是决定以keras的作主参考,
后面待续...