https://www.tensorflow.org/tutorials/images/segmentation
这篇教程将重点讨论图像分割任务,使用的是改进版的 U-Net。
每个像素分别属于以下三个类别中的一个:
类别 1:像素是宠物的一部分。
类别 2:像素是宠物的轮廓。
类别 3:以上都不是/外围像素。
数据集的路径:
定义模型
这里用到的模型是一个改版的 U-Net。U-Net 由一个编码器(下采样器(downsampler))和一个解码器(上采样器(upsampler))组成。为了学习到鲁棒的特征,同时减少可训练参数的数量,这里可以使用一个预训练模型作为编码器。因此,这项任务中的编码器将使用一个预训练的 MobileNetV2 模型,它的中间输出值将被使用。解码器将使用在 TensorFlow Examples 中的 Pix2pix tutorial 里实施过的升频取样模块。
输出信道数量为 3 是因为每个像素有三种可能的标签。把这想象成一个多类别分类,每个像素都将被分到三个类别当中。
训练模型
现在,要做的只剩下编译和训练模型了。这里用到的损失函数是 losses.sparse_categorical_crossentropy。使用这个损失函数是因为神经网络试图给每一个像素分配一个标签,和多类别预测是一样的。在正确的分割掩码中,每个像素点的值是 {0,1,2} 中的一个。同时神经网络也输出三个信道。本质上,每个信道都在尝试学习预测一个类别,而 losses.sparse_categorical_crossentropy 正是这一情形下推荐使用的损失函数。根据神经网络的输出值,分配给每个像素的标签为输出值最高的信道所表示的那一类。这就是 create_mask 函数所做的工作。
# -*- coding: utf-8 -*-
"""
Created on Thu Jul 30 18:51:35 2020
@author: YZK
"""
import tensorflow as tf
print(tf.__version__)
import numpy as np
import os
from IPython.display import clear_output
from PIL import Image
import matplotlib.pyplot as plt
import pix2pix
samples_all_number=1000
save_dir='./log'
count=0
if not os.path.exists(save_dir): os.mkdir(save_dir)
#从磁盘上加载原始图片
def load_tensor_from_file(img_file):
img = Image.open(img_file)
sample_image = np.array(img)
#样本图片规格不一致,需要做通道转换,否则抛异常
if len(sample_image.shape) != 3 or sample_image.shape[2] == 4:
img = img.convert("RGB")
sample_image = np.array(img)
sample_image = tf.image.resize(sample_image,[128, 128])
return sample_image
#从磁盘上加载标记图片
def load_ann_from_file(img_file):
sample_image = tf.image.decode_image(tf.io.read_file(img_file))
#样本图片规格不一致,需要做通道转换,否则抛异常
if sample_image.shape[2] != 1:
img = Image.open(img_file)
img = img.convert("L")#转为灰度图
sample_image = np.array(img)
sample_image = tf.image.resize(sample_image,[128, 128])
return sample_image
#加载图片并转换为训练集和验证集,同时输出训练集、验证集样本数量
def load(img_path):
trainImageList = []
valImageList = []
path = img_path + "/images"
files = os.listdir(path)
cnt = 0
for imgFile in files:
if os.path.isdir(imgFile):
continue
file = path + "/" + imgFile
print("load image ", file)
cnt += 1
img = load_tensor_from_file(file)
img = tf.squeeze(img)
#每8张图片中抽取一个样本作为验证集
if cnt % 8 == 0:
valImageList.append(img)
else:
trainImageList.append(img)
#加载1000张样本,机器配置有限,样本过多,报00M错误
if cnt > samples_all_number:
break
trainAnnList = []
valAnnList = []
path = img_path + "/ann"
files = os.listdir(path)
cnt = 0
for imgFile in files:
if os.path.isdir(imgFile):
continue
file = path + "/" + imgFile
print("load image ", file)
img = load_ann_from_file(file)
cnt+=1
if cnt % 8 == 0:
valAnnList.append(img)
else:
trainAnnList.append(img)
#加载1000张样本,机器配置有限,样本过多,报00M错误
if cnt > samples_all_number:
break
train_num = len(trainImageList)
val_num = len(valImageList)
x = tf.convert_to_tensor(trainImageList, dtype=tf.float32)
y = tf.convert_to_tensor(trainAnnList, dtype=tf.float32)
dataset_train = tf.data.Dataset.from_tensor_slices((x, y))
x_val = tf.convert_to_tensor(valImageList, dtype=tf.float32)
y_val = tf.convert_to_tensor(valAnnList, dtype=tf.float32)
dataset_val = tf.data.Dataset.from_tensor_slices((x_val, y_val))
return dataset_train, train_num, dataset_val, val_num
train_dataset, train_num, val_dataset, val_num = load("D:/dataset/pet_images")
'''
下面的代码进行了一个简单的图像翻转扩充。然后,将图像标准化到 [0,1]。最后,如上文提到的,像素点在图像分割掩码中被标记为 {1, 2, 3} 中的一个。为了方便起见,我们将分割掩码都减 1,得到了以下的标签:{0, 1, 2}。
'''
def normalize(input_image, input_mask):
input_image = tf.cast(input_image, tf.float32) / 128.0 - 1
#mask图像数据需根据标记数据做具体转换,使用labelme工具标记的图像,需要将图像颜色转换为类别标签索引,否则loss=nan
input_mask -= 1
return input_image, input_mask
@tf.function
def load_image_train(x, y):
input_image = tf.image.resize(x, (128, 128))
input_mask = tf.image.resize(y, (128,128))
if tf.random.uniform(()) > 0.5:
input_image = tf.image.flip_left_right(input_image)
input_mask = tf.image.flip_left_right(input_mask)
input_image, input_mask = normalize(input_image, input_mask)
return input_image, input_mask
def load_image_test(x, y):
input_image = tf.image.resize(x, (128, 128))
input_mask = tf.image.resize(y, (128,128))
input_image, input_mask = normalize(input_image, input_mask)
return input_image, input_mask
TRAIN_LENGTH = train_num
#根据GPU性能调节BATCH_SIZE大小
BATCH_SIZE = 16#64
BUFFER_SIZE = 1000
STEPS_PER_EPOCH= TRAIN_LENGTH
train = train_dataset.map(load_image_train, num_parallel_calls=tf.data.experimental.AUTOTUNE)
val_dataset = val_dataset.map(load_image_test)
train_dataset = train.cache().shuffle(BUFFER_SIZE).batch(BATCH_SIZE).repeat()
train_dataset = train_dataset.prefetch(buffer_size=tf.data.experimental.AUTOTUNE)
val_dataset = val_dataset.batch(BATCH_SIZE)
def display(display_list):
global count
count=count+1
plt.figure(figsize=(15,15))
title = ['Input Image', 'True Mask', 'Predicted Mask']
for i in range(len(display_list)):
plt.subplot(1, len(display_list), i+1)
plt.title(title[i])
plt.imshow(tf.keras.preprocessing.image.array_to_img(display_list[i]))
plt.axis('off')
save_file = save_dir+"/count_%d.jpg" %count
plt.savefig(save_file)
plt.show()
np.set_printoptions(threshold=128*128)
for image, mask in train.take(1):
sample_image, sample_mask = image, mask
print(tf.reduce_min(mask), tf.reduce_max(mask), tf.reduce_mean(mask))
display([sample_image,sample_mask])
OUTPUT_CHANNELS = 3
base_model = tf.keras.applications.MobileNetV2(input_shape=[128,128,3], include_top=False)
layer_names = [
'block_1_expand_relu', #64x64
'block_3_expand_relu', #32x32
'block_6_expand_relu', #16x16
'block_13_expand_relu',#8x8
'block_16_project', #4x4
]
layers = [base_model.get_layer(name).output for name in layer_names]
#创建特征提取模型
down_stack = tf.keras.Model(inputs=base_model.input, outputs=layers)
down_stack.trainable = False
up_stack =[
pix2pix.upsample(512, 3),#4x4 -> 8x8
pix2pix.upsample(256, 3),#8x8 -> 16x16
pix2pix.upsample(128, 3),#16x16 -> 32x32
pix2pix.upsample(64, 3), #32x32 -> 64x64
]
def unet_model(output_channels):
last = tf.keras.layers.Conv2DTranspose(output_channels, 3, strides=2,padding='same', activation='softmax')
inputs = tf.keras.layers.Input(shape=[128,128,3])
x = inputs
#降频采样
skips = down_stack(x)
x = skips[-1]#取最后一次输出
skips=reversed(skips[:-1])
#升频采样
for up, skip in zip(up_stack, skips):
x = up(x)
concat = tf.keras.layers.Concatenate()
x = concat([x, skip])
x = last(x)
return tf.keras.Model(inputs=inputs, outputs=x)
model = unet_model(OUTPUT_CHANNELS)
adam = tf.keras.optimizers.Adam(lr=1e-3)
#optimizer = optimizers.Adam(lr=1e-3)
model.compile(adam, loss='sparse_categorical_crossentropy', metrics=['accuracy'])
def create_mask(pred_mask):
pred_mask = tf.argmax(pred_mask, axis=-1)
pred_mask = pred_mask[..., tf.newaxis]
return pred_mask[0]
def show_predictions(dataset=None, num=1):
if dataset:
for image, mask in dataset.take(num):
pred_mask = model.predict(image)
display([image[0], mask[0], create_mask(pred_mask)])
else:
display([sample_image, sample_mask,
create_mask(model.predict(sample_image[tf.newaxis, ...]))])
show_predictions()
class DisplayCallback(tf.keras.callbacks.Callback):
def on_epoch_end(self, epoch, logs=None):
clear_output(wait=True)
show_predictions()
print ('\nSample Prediction after epoch {}\n'.format(epoch+1))
EPOCHS = 2
VAL_SUBSPLITS = 5
VALIDATION_STEPS = val_num//BATCH_SIZE//VAL_SUBSPLITS
model_history = model.fit(train_dataset, epochs=EPOCHS,
steps_per_epoch=STEPS_PER_EPOCH,
validation_steps=VALIDATION_STEPS,
validation_data=val_dataset,
callbacks=[DisplayCallback()])
model.save("poker.h5")
loss = model_history.history['loss']
val_loss = model_history.history['val_loss']
epochs = range(EPOCHS)
plt.figure()
plt.plot(epochs, loss, 'r', label='Training loss')
plt.plot(epochs, val_loss, 'bo', label='Validation loss')
plt.title('Training and Validation Loss')
plt.xlabel('Epoch')
plt.ylabel('Loss Value')
plt.ylim([0, 1])
plt.legend()
plt.savefig(save_dir+'Training.png')
plt.show()