devbox
不正经
这个屌丝很懒，什么也没留下！
关注作者
热门标签
热门文章
当前位置:   article > 正文

[深度学习从入门到女装]keras实战-UNet2d（VOC2012）_unet2dmodel

作者：不正经 | 2024-02-17 00:04:10

赞

踩

unet2dmodel

本文实现使用keras实现U-Net2d网络，对VOC2012进行分割

 

UNet2d

首先是Unet2d网络的搭建

  1. import keras.backend as K
  2. from keras.engine import Input,Model
  3. import keras
  4. from keras.optimizers import Adam
  5. from keras.layers import BatchNormalization,Activation,Conv2D,MaxPooling2D,Conv2DTranspose,UpSampling2D
  6. import metrics as m
  7. from keras.layers.core import Lambda
  8. import numpy as np
  9.  
  10.  
  11. def up_and_concate(down_layer, layer):
  12.     in_channel = down_layer.get_shape().as_list()[3]
  13.     out_channel = in_channel // 2
  14.     up = Conv2DTranspose(out_channel,[2,2],strides=[2,2])(down_layer)
  15.     print("--------------")
  16.     print(str(up.get_shape()))
  17.     print(str(layer.get_shape()))
  18.     print("--------------")
  19.     my_concat = Lambda(lambda x: K.concatenate([x[0], x[1]], axis=3))
  20.     concate = my_concat([up, layer])
  21.     # must use lambda 
  22.     #concate=K.concatenate([up, layer], 3)
  23.     return concate
  24.  
  25.  
  26. def attention_block_2d(x, g, inter_channel):
  27.     '''
  29.     :param x: x input from down_sampling same layer output (?,x_height,x_width,x_channel)
  30.     :param g: gate input from up_sampling layer last output (?,g_height,g_width,g_channel)
  31.     g_height,g_width=x_height/2,x_width/2
  32.     :return:
  33.     '''
  34.     print('attention_block:')
  35.     # theta_x(?,g_height,g_width,inter_channel)
  36.     theta_x = Conv2D(inter_channel, [2, 2], strides=[2, 2])(x)
  37.     print(str(theta_x.get_shape()))
  38.     # phi_g(?,g_height,g_width,inter_channel)
  39.     phi_g = Conv2D(inter_channel, [1, 1], strides=[1, 1])(g)
  40.     print(str(phi_g.get_shape()))
  41.     # f(?,g_height,g_width,inter_channel)
  42.     f = Activation('relu')(keras.layers.add([theta_x, phi_g]))
  43.     print(str(f.get_shape()))
  44.     # psi_f(?,g_height,g_width,1)
  45.     psi_f = Conv2D(1, [1, 1], strides=[1, 1])(f)
  46.     print(str(psi_f.get_shape()))
  47.     # sigm_psi_f(?,g_height,g_width)
  48.     sigm_psi_f = Activation('sigmoid')(psi_f)
  49.     print(str(sigm_psi_f.get_shape()))
  50.     # rate(?,x_height,x_width)
  51.     rate = UpSampling2D(size=[2, 2])(sigm_psi_f)
  52.     print(str(rate.get_shape()))
  53.     # att_x(?,x_height,x_width,x_channel)
  54.     att_x = keras.layers.multiply([x, rate])
  55.     print(str(att_x.get_shape()))
  56.     print('-----------------')
  57.     return att_x
  58.  
  59.  
  60. def unet_model_2d_attention(input_shape,n_labels,batch_normalization=False,initial_learning_rate=0.00001,metrics=m.dice_coefficient):
  61.     """
  62.     input_shape:without batch_size,(img_height,img_width,img_depth)
  63.     metrics:
  64.     """
  65.  
  66.  
  67.     inputs=Input(input_shape)
  68.  
  69.     down_layer=[]
  70.  
  71.     layer=inputs
  72.  
  73.     #down_layer_1
  74.     layer=res_block_v2(layer,64,batch_normalization=batch_normalization)
  75.     down_layer.append(layer)
  76.     layer=MaxPooling2D(pool_size=[2,2],strides=[2,2])(layer)
  77.  
  78.     print(str(layer.get_shape()))
  79.  
  80.     # down_layer_2
  81.     layer = res_block_v2(layer, 128, batch_normalization=batch_normalization)
  82.     down_layer.append(layer)
  83.     layer = MaxPooling2D(pool_size=[2, 2], strides=[2, 2])(layer)
  84.  
  85.     print(str(layer.get_shape()))
  86.  
  87.     # down_layer_3
  88.     layer = res_block_v2(layer, 256, batch_normalization=batch_normalization)
  89.     down_layer.append(layer)
  90.     layer = MaxPooling2D(pool_size=[2, 2], strides=[2, 2])(layer)
  91.  
  92.     print(str(layer.get_shape()))
  93.  
  94.     # down_layer_4
  95.     layer = res_block_v2(layer, 512, batch_normalization=batch_normalization)
  96.     down_layer.append(layer)
  97.     layer = MaxPooling2D(pool_size=[2, 2], strides=[2, 2])(layer)
  98.  
  99.     print(str(layer.get_shape()))
  100.  
  101.     # bottle_layer
  102.     layer = res_block_v2(layer, 1024, batch_normalization=batch_normalization)
  103.     print(str(layer.get_shape()))
  104.  
  105.     # up_layer_4
  106.     layer = attention_block_2d( down_layer[3],layer,256)
  107.     layer = res_block_v2(layer, 512,batch_normalization=batch_normalization)
  108.     print(str(layer.get_shape()))
  109.  
  110.     # up_layer_3
  111.     layer =  attention_block_2d( down_layer[2],layer,128)
  112.     layer = res_block_v2(layer, 256, batch_normalization=batch_normalization)
  113.     print(str(layer.get_shape()))
  114.  
  115.     # up_layer_2
  116.     layer =  attention_block_2d( down_layer[1],layer,64)
  117.     layer = res_block_v2(layer, 128, batch_normalization=batch_normalization)
  118.     print(str(layer.get_shape()))
  119.  
  120.     # up_layer_1
  121.     layer =  attention_block_2d( down_layer[0],layer,32)
  122.     layer = res_block_v2(layer, 64, batch_normalization=batch_normalization)
  123.     print(str(layer.get_shape()))
  124.  
  125.     # score_layer
  126.     layer = Conv2D(n_labels,[1,1],strides=[1,1])(layer)
  127.     print(str(layer.get_shape()))
  128.  
  129.     # softmax
  130.     layer = Activation('softmax')(layer)
  131.     print(str(layer.get_shape()))
  132.  
  133.     outputs=layer
  134.  
  135.     model=Model(inputs=inputs,outputs=outputs)
  136.  
  137.     metrics=[metrics]
  138.  
  139.     model.compile(optimizer=Adam(lr=initial_learning_rate), loss=m.dice_coefficient_loss, metrics=metrics)
  140.  
  141.  
  142.     return model
  143.  
  144.  
  145. def unet_model_2d(input_shape,n_labels,batch_normalization=False,initial_learning_rate=0.00001,metrics=m.dice_coefficient):
  146.     """
  147.     input_shape:without batch_size,(img_height,img_width,img_depth)
  148.     metrics:
  149.     """
  150.  
  151.  
  152.     inputs=Input(input_shape)
  153.  
  154.     down_layer=[]
  155.  
  156.     layer=inputs
  157.  
  158.     #down_layer_1
  159.     layer=res_block_v2(layer,64,batch_normalization=batch_normalization)
  160.     down_layer.append(layer)
  161.     layer=MaxPooling2D(pool_size=[2,2],strides=[2,2])(layer)
  162.  
  163.     print(str(layer.get_shape()))
  164.  
  165.     # down_layer_2
  166.     layer = res_block_v2(layer, 128, batch_normalization=batch_normalization)
  167.     down_layer.append(layer)
  168.     layer = MaxPooling2D(pool_size=[2, 2], strides=[2, 2])(layer)
  169.  
  170.     print(str(layer.get_shape()))
  171.  
  172.     # down_layer_3
  173.     layer = res_block_v2(layer, 256, batch_normalization=batch_normalization)
  174.     down_layer.append(layer)
  175.     layer = MaxPooling2D(pool_size=[2, 2], strides=[2, 2])(layer)
  176.  
  177.     print(str(layer.get_shape()))
  178.  
  179.     # down_layer_4
  180.     layer = res_block_v2(layer, 512, batch_normalization=batch_normalization)
  181.     down_layer.append(layer)
  182.     layer = MaxPooling2D(pool_size=[2, 2], strides=[2, 2])(layer)
  183.  
  184.     print(str(layer.get_shape()))
  185.  
  186.     # bottle_layer
  187.     layer = res_block_v2(layer, 1024, batch_normalization=batch_normalization)
  188.     print(str(layer.get_shape()))
  189.  
  190.     # up_layer_4
  191.     layer = up_and_concate(layer, down_layer[3])
  192.     layer = res_block_v2(layer, 512,batch_normalization=batch_normalization)
  193.     print(str(layer.get_shape()))
  194.  
  195.     # up_layer_3
  196.     layer = up_and_concate(layer, down_layer[2])
  197.     layer = res_block_v2(layer, 256, batch_normalization=batch_normalization)
  198.     print(str(layer.get_shape()))
  199.  
  200.     # up_layer_2
  201.     layer = up_and_concate(layer, down_layer[1])
  202.     layer = res_block_v2(layer, 128, batch_normalization=batch_normalization)
  203.     print(str(layer.get_shape()))
  204.  
  205.     # up_layer_1
  206.     layer = up_and_concate(layer, down_layer[0])
  207.     layer = res_block_v2(layer, 64, batch_normalization=batch_normalization)
  208.     print(str(layer.get_shape()))
  209.  
  210.     # score_layer
  211.     layer = Conv2D(n_labels,[1,1],strides=[1,1])(layer)
  212.     print(str(layer.get_shape()))
  213.  
  214.     # softmax
  215.     layer = Activation('softmax')(layer)
  216.     print(str(layer.get_shape()))
  217.  
  218.     outputs=layer
  219.  
  220.     model=Model(inputs=inputs,outputs=outputs)
  221.  
  222.     metrics=[metrics]
  223.  
  224.     model.compile(optimizer=Adam(lr=initial_learning_rate), loss=m.dice_coefficient_loss, metrics=metrics)
  225.  
  226.  
  227.     return model
  228.  
  229.  
  230.  
  231. def res_block_v2(input_layer,out_n_filters,batch_normalization=False,kernel_size=[3,3],stride=[1,1],padding='same'):
  232.  
  233.     input_n_filters = input_layer.get_shape().as_list()[3]
  234.     print(str(input_layer.get_shape()))
  235.     layer=input_layer
  236.  
  237.     for i in range(2):
  238.         if batch_normalization:
  239.             layer=BatchNormalization()(layer)
  240.         layer=Activation('relu')(layer)
  241.         layer=Conv2D(out_n_filters,kernel_size,strides=stride,padding=padding)(layer)
  242.  
  243.     if out_n_filters!=input_n_filters:
  244.         skip_layer=Conv2D(out_n_filters,[1,1],strides=stride,padding=padding)(input_layer)
  245.     else:
  246.         skip_layer=input_layer
  247.  
  248.     out_layer=keras.layers.add([layer,skip_layer])
  249.  
  250.  
  251.     return out_layer
  252.  
  253.  
  254.  
  255.

使用Keras中的Model类，首先使用Input(input_shape)，注意这里的input_shape是不带batch_size这一维的，在这里就是(img_height,img_width,img_depth)

metrics为评判标准

  1. def up_and_concate(down_layer, layer):
  2.     in_channel = down_layer.get_shape().as_list()[3]
  3.     out_channel = in_channel // 2
  4.     up = Conv2DTranspose(out_channel,[2,2],strides=[2,2])(down_layer)
  5.     print("--------------")
  6.     print(str(up.get_shape()))
  7.     print(str(layer.get_shape()))
  8.     print("--------------")
  9.     my_concat = Lambda(lambda x: K.concatenate([x[0], x[1]], axis=3))
  10.     concate = my_concat([up, layer])
  11.     # must use lambda 
  12.     #concate=K.concatenate([up, layer], 3)
  13.     return concate

以上为skip_connection的函数，down_layer是上一层上采样层的输出，layer为同层下采样层的输出

注意这里不能直接用K.concatenate，会报错说使用的tensor不是keras里边的tensor，必须使用Lambda

  1. def res_block_v2(input_layer,out_n_filters,batch_normalization=False,kernel_size=[3,3],stride=[1,1],padding='same'):
  2.  
  3.     input_n_filters = input_layer.get_shape().as_list()[3]
  4.     print(str(input_layer.get_shape()))
  5.     layer=input_layer
  6.  
  7.     for i in range(2):
  8.         if batch_normalization:
  9.             layer=BatchNormalization()(layer)
  10.         layer=Activation('relu')(layer)
  11.         layer=Conv2D(out_n_filters,kernel_size,strides=stride,padding=padding)(layer)
  12.  
  13.     if out_n_filters!=input_n_filters:
  14.         skip_layer=Conv2D(out_n_filters,[1,1],strides=stride,padding=padding)(input_layer)
  15.     else:
  16.         skip_layer=input_layer
  17.  
  18.     out_layer=keras.layers.add([layer,skip_layer])
  19.  
  20.  
  21.     return out_layer

 

VOC读取

接下来是voc数据集的读取，voc数据集的目录如下：

Annotations为存放每个图片的描述文件（.xml），类别，检测框什么的，在分割上用不到

ImageSets为存放了各个任务的trian val所需图片的名称

JPEGImages为存放所有图片（.jpg）

SegmentationClass是用于语义分割的标签（.png）

SegmentationObject是用于实例分割的标签（.png）

在本文只用了ImageSets中的Segmentation和SegmentionClass中的文件

文件的读取如下：

  1. import tensorflow as tf
  2. from PIL import Image
  3. import PIL
  4. import scipy.misc as misc
  5. import numpy as np
  6.  
  7. def make_one_hot(x,n):
  8.     '''
  9.     print(x.shape)
  10.     one_hot=np.zeros([x.shape[0],x.shape[1],n])
  11.     print(one_hot.shape)
  12.     for i in range(n):
  13.         #print(x==i)
  14.         print(one_hot[x==i])
  15.         one_hot[x==i][i]=1
  16.     '''
  17.     one_hot = np.zeros([x.shape[0], x.shape[1], n])
  18.     for i in range(x.shape[0]):
  19.         for j in range(x.shape[1]):
  20.             one_hot[i,j,x[i,j]]=1
  21.     return one_hot
  22.  
  23.  
  24. class voc_reader:
  25.     def __init__(self,resize_width,resize_height,train_batch_size,val_batch_size):
  26.         self.train_file_name_list=self.load_file_name_list(file_path="D:\\pyproject\\data\\VOCtrainval_11-May-2012\\VOCdevkit\\VOC2012\\ImageSets\\Segmentation\\train.txt")
  27.         self.val_file_name_list=self.load_file_name_list(file_path="D:\\pyproject\\data\\VOCtrainval_11-May-2012\\VOCdevkit\\VOC2012\\ImageSets\\Segmentation\\val.txt")
  28.         self.row_file_path="D:\\pyproject\\data\\VOCtrainval_11-May-2012\\VOCdevkit\\VOC2012\\JPEGImages\\"
  29.         self.label_file_path="D:\\pyproject\\data\\VOCtrainval_11-May-2012\\VOCdevkit\\VOC2012\\SegmentationClass\\"
  30.         self.train_batch_index=0
  31.         self.val_batch_index=0
  32.         self.resize_width=resize_width
  33.         self.resize_height=resize_height
  34.         self.n_train_file=len(self.train_file_name_list)
  35.         self.n_val_file=len(self.val_file_name_list)
  36.         self.train_batch_size=train_batch_size
  37.         self.val_batch_size=val_batch_size
  38.         print(self.n_train_file)
  39.         print(self.n_val_file)
  40.         self.n_train_steps_per_epoch=self.n_train_file//self.train_batch_size
  41.         self.n_val_steps_per_epoch=self.n_val_file//self.val_batch_size
  42.  
  43.  
  44.     def load_file_name_list(self,file_path):
  45.         file_name_list=[]
  46.         with open(file_path, 'r') as file_to_read:
  47.             while True:
  48.                 lines = file_to_read.readline().strip()  # 整行读取数据
  49.                 if not lines:
  50.                     break
  51.                     pass
  52.                 file_name_list.append(lines)
  53.                 pass
  54.         return file_name_list
  55.  
  56.     def next_train_batch(self):
  57.         train_imgs=np.zeros((self.train_batch_size,self.resize_height,self.resize_width,3))
  58.         train_labels=np.zeros([self.train_batch_size,self.resize_height,self.resize_width,21])
  59.         if self.train_batch_index>=self.n_train_steps_per_epoch:
  60.             print("next epoch")
  61.             self.train_batch_index=0
  62.         print('------------------')
  63.         print(self.train_batch_index)
  64.         for i in range(self.train_batch_size):
  65.             index=self.train_batch_size*self.train_batch_index+i
  66.             print('index'+str(index))
  67.             img = Image.open(self.row_file_path+self.train_file_name_list[index]+'.jpg')
  68.             img=img.resize((self.resize_height,self.resize_width),Image.NEAREST)
  69.             img=np.array(img)
  70.             train_imgs[i]=img
  71.             #print(img.shape)
  72.             np.set_printoptions(threshold=np.inf)
  73.  
  74.             label=Image.open(self.label_file_path+self.train_file_name_list[index]+'.png')
  75.             label=label.resize((self.resize_height,self.resize_width),Image.NEAREST)
  76.             label=np.array(label, dtype=np.int32)
  77.             #print(label[label>20])
  78.             #label[label == 255] = -1
  79.             label[label==255]=0
  80.             #print(label)
  81.             #print(label.shape)
  82.             one_hot_label=make_one_hot(label,21)
  83.             train_labels[i]=one_hot_label
  84.             #print(one_hot_label.shape)
  85.             #print(label)
  86.             #print(label)
  87.  
  88.         self.train_batch_index+=1
  89.         print('------------------')
  90.  
  91.  
  92.         return train_imgs,train_labels
  93.  
  94.  
  95.     def next_val_batch(self):
  96.         val_imgs = np.zeros((self.val_batch_size, self.resize_height, self.resize_width, 3))
  97.         val_labels = np.zeros([self.val_batch_size, self.resize_height, self.resize_width, 21])
  98.         if self.val_batch_index>=self.n_val_steps_per_epoch:
  99.             print("next epoch")
  100.             self.val_batch_index=0
  101.         print('------------------')
  102.         print(self.val_batch_index)
  103.  
  104.  
  105.         for i in range(self.val_batch_size):
  106.             index=self.val_batch_size*self.val_batch_index+i
  107.             print('index'+str(index))
  108.             img=Image.open(self.row_file_path+self.val_file_name_list[index]+'.jpg')
  109.             img = img.resize((self.resize_height, self.resize_width), Image.NEAREST)
  110.             img = np.array(img)
  111.             val_imgs[i]=img
  112.  
  113.             label = Image.open(self.label_file_path + self.val_file_name_list[index] + '.png')
  114.             label = label.resize((self.resize_height, self.resize_width), Image.NEAREST)
  115.             label = np.array(label, dtype=np.int32)
  116.             # print(label[label>20])
  117.             # label[label == 255] = -1
  118.             label[label == 255] = 0
  119.             # print(label)
  120.             # print(label.shape)
  121.             one_hot_label = make_one_hot(label, 21)
  122.             val_labels[i]=one_hot_label
  123.         print('------------------')
  124.         self.val_batch_index+=1
  125.  
  126.  
  127.  
  128.         return val_imgs,val_labels

 

Train

先构造train和val的generator，用于等会的fit_generator

  1. def train_generator_data(batch_size,voc_reader):
  2.     while True:
  3.         x,y=voc_reader.next_train_batch(batch_size)
  4.         yield (x,y)
  5.  
  6. def val_generator_data(batch_size,voc_reader):
  7.     while True:
  8.         x,y=voc_reader.next_val_batch(batch_size)
  9.         yield (x,y)

随后定义callback

  1. def get_callbacks(model_file,initial_learning_rate=0.0001,learning_rate_drop=0.5,learning_rate_epochs=None,
  2.                   learning_rate_patience=50,logging_file="training.log",verbosity=1,early_stopping_patience=None):
  3.     callbacks=list()
  4.  
  5.     callbacks.append(ModelCheckpoint(model_file,save_best_only=True))
  6.     callbacks.append(CSVLogger(logging_file,append=True))
  7.     callbacks.append(TensorBoard())
  8.  
  9.     if learning_rate_epochs:
  10.         callbacks.append(LearningRateScheduler(partial(step_decay, initial_lrate=initial_learning_rate,
  11.                                                        drop=learning_rate_drop, epochs_drop=learning_rate_epochs)))
  12.     else:
  13.         callbacks.append(ReduceLROnPlateau(factor=learning_rate_drop, patience=learning_rate_patience,
  14.                                            verbose=verbosity))
  15.     if early_stopping_patience:
  16.         callbacks.append(EarlyStopping(verbose=verbosity, patience=early_stopping_patience))
  17.     return callbacks

加载已经训练过的模型

  1. def load_old_model(model_file):
  2.     print("Loading pre-trained model")
  3.     custom_objects = {'dice_coefficient_loss': dice_coefficient_loss, 'dice_coefficient': dice_coefficient}
  4.  
  5.     try:
  6.         from keras_contrib.layers import InstanceNormalization
  7.         custom_objects["InstanceNormalization"] = InstanceNormalization
  8.     except ImportError:
  9.         pass
  10.     try:
  11.         return load_model(model_file, custom_objects=custom_objects)
  12.     except ValueError as error:
  13.         if 'InstanceNormalization' in str(error):
  14.             raise ValueError(str(error) + "\n\nPlease install keras-contrib to use InstanceNormalization:\n"
  15.                                           "'pip install git+https://www.github.com/keras-team/keras-contrib.git'")
  16.         else:
  17.             raise error

train_model函数

  1. def train_model(model, model_file, training_generator, validation_generator, steps_per_epoch, validation_steps,
  2.                 initial_learning_rate=0.001, learning_rate_drop=0.5, learning_rate_epochs=None, n_epochs=500,
  3.                 learning_rate_patience=20, early_stopping_patience=None):
  4.     """
  5.     Train a Keras model.
  6.     :param early_stopping_patience: If set, training will end early if the validation loss does not improve after the
  7.     specified number of epochs.
  8.     :param learning_rate_patience: If learning_rate_epochs is not set, the learning rate will decrease if the validation
  9.     loss does not improve after the specified number of epochs. (default is 20)
  10.     :param model: Keras model that will be trained.
  11.     :param model_file: Where to save the Keras model.
  12.     :param training_generator: Generator that iterates through the training data.
  13.     :param validation_generator: Generator that iterates through the validation data.
  14.     :param steps_per_epoch: Number of batches that the training generator will provide during a given epoch.
  15.     :param validation_steps: Number of batches that the validation generator will provide during a given epoch.
  16.     :param initial_learning_rate: Learning rate at the beginning of training.
  17.     :param learning_rate_drop: How much at which to the learning rate will decay.
  18.     :param learning_rate_epochs: Number of epochs after which the learning rate will drop.
  19.     :param n_epochs: Total number of epochs to train the model.
  20.     :return:
  21.     """
  22.     model.fit_generator(generator=training_generator,
  23.                         steps_per_epoch=steps_per_epoch,
  24.                         epochs=n_epochs,
  25.                         validation_data=validation_generator,
  26.                         validation_steps=validation_steps,
  27.                         callbacks=get_callbacks(model_file,
  28.                                                 initial_learning_rate=initial_learning_rate,
  29.                                                 learning_rate_drop=learning_rate_drop,
  30.                                                 learning_rate_epochs=learning_rate_epochs,
  31.                                                 learning_rate_patience=learning_rate_patience,
  32.                                                 early_stopping_patience=early_stopping_patience))
  33.

 

github地址：https://github.com/panxiaobai/voc_keras

本文内容由网友自发贡献，转载请注明出处：【wpsshop博客】
推荐阅读
相关标签

Copyright © 2003-2013 www.wpsshop.cn 版权所有，并保留所有权利。

  

闽ICP备14008679号