PaddlePaddle动态图实现VGG(眼底筛查为例)

本案例参考课程:百度架构师手把手教深度学习的内容。 主要目的为练习vgg动态图的PaddlePaddle实现。

本案例已经在AISTUDIO共享,链接为:

https://aistudio.baidu.com/aistudio/projectdetail/244766

数据集iChallenge-PM:

数据集图片 iChallenge-PM中既有病理性近视患者的眼底图片,也有非病理性近视患者的图片,命名规则如下:

病理性近视(PM):文件名以P开头

非病理性近视(non-PM):

高度近似(high myopia):文件名以H开头

正常眼睛(normal):文件名以N开头

我们将病理性患者的图片作为正样本,标签为1; 非病理性患者的图片作为负样本,标签为0。从数据集中选取两张图片,通过LeNet提取特征,构建分类器,对正负样本进行分类,并将图片显示出来。

算法:

VGG VGG是当前最流行的CNN模型之一,2014年由Simonyan和Zisserman提出,其命名来源于论文作者所在的实验室Visual Geometry Group。AlexNet模型通过构造多层网络,取得了较好的效果,但是并没有给出深度神经网络设计的方向。VGG通过使用一系列大小为3x3的小尺寸卷积核和pooling层构造深度卷积神经网络,并取得了较好的效果。VGG模型因为结构简单、应用性极强而广受研究者欢迎,尤其是它的网络结构设计方法,为构建深度神经网络提供了方向。

图3 是VGG-16的网络结构示意图,有13层卷积和3层全连接层。VGG网络的设计严格使用3×33\times 33×3的卷积层和池化层来提取特征,并在网络的最后面使用三层全连接层,将最后一层全连接层的输出作为分类的预测。 在VGG中每层卷积将使用ReLU作为激活函数,在全连接层之后添加dropout来抑制过拟合。使用小的卷积核能够有效地减少参数的个数,使得训练和测试变得更加有效。比如使用两层3×33\times 33×3卷积层,可以得到感受野为5的特征图,而比使用5×55 \times 55×5的卷积层需要更少的参数。由于卷积核比较小,可以堆叠更多的卷积层,加深网络的深度,这对于图像分类任务来说是有利的。VGG模型的成功证明了增加网络的深度,可以更好的学习图像中的特征模式。

关键代码:

import os

import numpy as np

import matplotlib.pyplot as plt

%matplotlib inline

from PIL import Image

DATADIR = '/home/aistudio/work/palm/PALM-Training400/PALM-Training400'

# 文件名以N开头的是正常眼底图片,以P开头的是病变眼底图片

file1 = 'N0012.jpg'

file2 = 'P0095.jpg'

# 读取图片

img1 = Image.open(os.path.join(DATADIR, file1))

img1 = np.array(img1)

img2 = Image.open(os.path.join(DATADIR, file2))

img2 = np.array(img2)

# 画出读取的图片

plt.figure(figsize=(16, 8))

f = plt.subplot(121)

f.set_title('Normal', fontsize=20)

plt.imshow(img1)

f = plt.subplot(122)

f.set_title('PM', fontsize=20)

plt.imshow(img2)

plt.show()

In[3]

# 查看图片形状

img1.shape, img2.shape

((2056, 2124, 3), (2056, 2124, 3))

In[5]

#定义数据读取器

import cv2

import random

import numpy as np

# 对读入的图像数据进行预处理

def transform_img(img):

    # 将图片尺寸缩放道 224x224

    img = cv2.resize(img, (224, 224))

    # 读入的图像数据格式是[H, W, C]

    # 使用转置操作将其变成[C, H, W]

    img = np.transpose(img, (2,0,1))

    img = img.astype('float32')

    # 将数据范围调整到[-1.0, 1.0]之间

    img = img / 255.

    img = img * 2.0 - 1.0

    return img

# 定义训练集数据读取器

def data_loader(datadir, batch_size=10, mode = 'train'):

    # 将datadir目录下的文件列出来,每条文件都要读入

    filenames = os.listdir(datadir)

    def reader():

        if mode == 'train':

            # 训练时随机打乱数据顺序

            random.shuffle(filenames)

        batch_imgs = []

        batch_labels = []

        for name in filenames:

            filepath = os.path.join(datadir, name)

            img = cv2.imread(filepath)

            img = transform_img(img)

            if name[0] == 'H' or name[0] == 'N':

                # H开头的文件名表示高度近似,N开头的文件名表示正常视力

                # 高度近视和正常视力的样本,都不是病理性的,属于负样本,标签为0

                label = 0

            elif name[0] == 'P':

                # P开头的是病理性近视,属于正样本,标签为1

                label = 1

            else:

                raise('Not excepted file name')

            # 每读取一个样本的数据,就将其放入数据列表中

            batch_imgs.append(img)

            batch_labels.append(label)

            if len(batch_imgs) == batch_size:

                # 当数据列表的长度等于batch_size的时候,

                # 把这些数据当作一个mini-batch,并作为数据生成器的一个输出

                imgs_array = np.array(batch_imgs).astype('float32')

                labels_array = np.array(batch_labels).astype('float32').reshape(-1, 1)

                yield imgs_array, labels_array

                batch_imgs = []

                batch_labels = []

        if len(batch_imgs) > 0:

            # 剩余样本数目不足一个batch_size的数据,一起打包成一个mini-batch

            imgs_array = np.array(batch_imgs).astype('float32')

            labels_array = np.array(batch_labels).astype('float32').reshape(-1, 1)

            yield imgs_array, labels_array

    return reader

# 查看数据形状

DATADIR = '/home/aistudio/work/palm/PALM-Training400/PALM-Training400'

train_loader = data_loader(DATADIR,

                          batch_size=10, mode='train')

data_reader = train_loader()

data = next(data_reader)

data[0].shape, data[1].shape

((10, 3, 224, 224), (10, 1))

In[6]

!pip install xlrd

import pandas as pd

df=pd.read_excel('/home/aistudio/work/palm/PALM-Validation-GT/PM_Label_and_Fovea_Location.xlsx')

df.to_csv('/home/aistudio/work/palm/PALM-Validation-GT/labels.csv',index=False)

#训练和评估代码

import os

import random

import paddle

import paddle.fluid as fluid

import numpy as np

DATADIR = '/home/aistudio/work/palm/PALM-Training400/PALM-Training400'

DATADIR2 = '/home/aistudio/work/palm/PALM-Validation400'

CSVFILE = '/home/aistudio/work/palm/PALM-Validation-GT/labels.csv'

# 定义训练过程

def train(model):

    with fluid.dygraph.guard():

        print('start training ... ')

        model.train()

        epoch_num = 5

        # 定义优化器

        opt = fluid.optimizer.Momentum(learning_rate=0.001, momentum=0.9)

        # 定义数据读取器,训练数据读取器和验证数据读取器

        train_loader = data_loader(DATADIR, batch_size=10, mode='train')

        valid_loader = valid_data_loader(DATADIR2, CSVFILE)

        for epoch in range(epoch_num):

            for batch_id, data in enumerate(train_loader()):

                x_data, y_data = data

                img = fluid.dygraph.to_variable(x_data)

                label = fluid.dygraph.to_variable(y_data)

                # 运行模型前向计算,得到预测值

                logits = model(img)

                # 进行loss计算

                loss = fluid.layers.sigmoid_cross_entropy_with_logits(logits, label)

                avg_loss = fluid.layers.mean(loss)

                if batch_id % 10 == 0:

                    print("epoch: {}, batch_id: {}, loss is: {}".format(epoch, batch_id, avg_loss.numpy()))

                # 反向传播,更新权重,清除梯度

                avg_loss.backward()

                opt.minimize(avg_loss)

                model.clear_gradients()

            model.eval()

            accuracies = []

            losses = []

            for batch_id, data in enumerate(valid_loader()):

                x_data, y_data = data

                img = fluid.dygraph.to_variable(x_data)

                label = fluid.dygraph.to_variable(y_data)

                # 运行模型前向计算,得到预测值

                logits = model(img)

                # 二分类,sigmoid计算后的结果以0.5为阈值分两个类别

                # 计算sigmoid后的预测概率,进行loss计算

                pred = fluid.layers.sigmoid(logits)

                loss = fluid.layers.sigmoid_cross_entropy_with_logits(logits, label)

                # 计算预测概率小于0.5的类别

                pred2 = pred * (-1.0) + 1.0

                # 得到两个类别的预测概率,并沿第一个维度级联

                pred = fluid.layers.concat([pred2, pred], axis=1)

                acc = fluid.layers.accuracy(pred, fluid.layers.cast(label, dtype='int64'))

                accuracies.append(acc.numpy())

                losses.append(loss.numpy())

            print("[validation] accuracy/loss: {}/{}".format(np.mean(accuracies), np.mean(losses)))

            model.train()

        # save params of model

        fluid.save_dygraph(model.state_dict(), 'mnist')

        # save optimizer state

        fluid.save_dygraph(opt.state_dict(), 'mnist')

# 定义评估过程

def evaluation(model, params_file_path):

    with fluid.dygraph.guard():

        print('start evaluation .......')

        #加载模型参数

        model_state_dict, _ = fluid.load_dygraph(params_file_path)

        model.load_dict(model_state_dict)

        model.eval()

        eval_loader = load_data('eval')

        acc_set = []

        avg_loss_set = []

        for batch_id, data in enumerate(eval_loader()):

            x_data, y_data = data

            img = fluid.dygraph.to_variable(x_data)

            label = fluid.dygraph.to_variable(y_data)

            # 计算预测和精度

            prediction, acc = model(img, label)

            # 计算损失函数值

            loss = fluid.layers.cross_entropy(input=prediction, label=label)

            avg_loss = fluid.layers.mean(loss)

            acc_set.append(float(acc.numpy()))

            avg_loss_set.append(float(avg_loss.numpy()))

        # 求平均精度

        acc_val_mean = np.array(acc_set).mean()

        avg_loss_val_mean = np.array(avg_loss_set).mean()

        print('loss={}, acc={}'.format(avg_loss_val_mean, acc_val_mean))

In[8]

# -*- coding:utf-8 -*-

# VGG模型代码

import numpy as np

import paddle

import paddle.fluid as fluid

from paddle.fluid.layer_helper import LayerHelper

from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, FC

from paddle.fluid.dygraph.base import to_variable

# 定义vgg块,包含多层卷积和1层2x2的最大池化层

class vgg_block(fluid.dygraph.Layer):

    def __init__(self, name_scope, num_convs, num_channels):

        """

        num_convs, 卷积层的数目

        num_channels, 卷积层的输出通道数,在同一个Incepition块内,卷积层输出通道数是一样的

        """

        super(vgg_block, self).__init__(name_scope)

        self.conv_list = []

        for i in range(num_convs):

            conv_layer = self.add_sublayer('conv_' + str(i), Conv2D(self.full_name(),

                                        num_filters=num_channels, filter_size=3, padding=1, act='relu'))

            self.conv_list.append(conv_layer)

        self.pool = Pool2D(self.full_name(), pool_stride=2, pool_size = 2, pool_type='max')

    def forward(self, x):

        for item in self.conv_list:

            x = item(x)

        return self.pool(x)

class VGG(fluid.dygraph.Layer):

    def __init__(self, name_scope, conv_arch=((2, 64),

                                (2, 128), (3, 256), (3, 512), (3, 512))):

        super(VGG, self).__init__(name_scope)

        self.vgg_blocks=[]

        iter_id = 0

        # 添加vgg_block

        # 这里一共5个vgg_block,每个block里面的卷积层数目和输出通道数由conv_arch指定

        for (num_convs, num_channels) in conv_arch:

            block = self.add_sublayer('block_' + str(iter_id),

                    vgg_block(self.full_name(), num_convs, num_channels))

            self.vgg_blocks.append(block)

            iter_id += 1

        self.fc1 = FC(self.full_name(),

                      size=4096,

                      act='relu')

        self.drop1_ratio = 0.5

        self.fc2= FC(self.full_name(),

                      size=4096,

                      act='relu')

        self.drop2_ratio = 0.5

        self.fc3 = FC(self.full_name(),

                      size=1,

                      )

    def forward(self, x):

        for item in self.vgg_blocks:

            x = item(x)

        x = fluid.layers.dropout(self.fc1(x), self.drop1_ratio)

        x = fluid.layers.dropout(self.fc2(x), self.drop2_ratio)

        x = self.fc3(x)

        return x

with fluid.dygraph.guard():

    model = VGG("VGG")

train(model)

start training ...

epoch: 0, batch_id: 0, loss is: [0.7242754]

epoch: 0, batch_id: 10, loss is: [0.6634571]

epoch: 0, batch_id: 20, loss is: [0.7898234]

epoch: 0, batch_id: 30, loss is: [0.60537547]

[validation] accuracy/loss: 0.9424999952316284/0.35623037815093994

epoch: 1, batch_id: 0, loss is: [0.31599292]

epoch: 1, batch_id: 10, loss is: [0.1198744]

epoch: 1, batch_id: 20, loss is: [0.46862125]

epoch: 1, batch_id: 30, loss is: [0.2300901]

[validation] accuracy/loss: 0.92249995470047/0.2342415601015091

epoch: 2, batch_id: 0, loss is: [0.22039299]

epoch: 2, batch_id: 10, loss is: [0.65977865]

epoch: 2, batch_id: 20, loss is: [0.37409317]

epoch: 2, batch_id: 30, loss is: [0.1841044]

[validation] accuracy/loss: 0.9325000643730164/0.22097690403461456

epoch: 3, batch_id: 0, loss is: [0.4992897]

epoch: 3, batch_id: 10, loss is: [0.31177607]

epoch: 3, batch_id: 20, loss is: [0.1721839]

epoch: 3, batch_id: 30, loss is: [0.38319916]

[validation] accuracy/loss: 0.9199999570846558/0.20679759979248047

epoch: 4, batch_id: 0, loss is: [0.20610766]

epoch: 4, batch_id: 10, loss is: [0.06688808]

epoch: 4, batch_id: 20, loss is: [0.3352648]

epoch: 4, batch_id: 30, loss is: [0.28062168]

[validation] accuracy/loss: 0.9149999618530273/0.21788272261619568

with fluid.dygraph.guard():

    model = VGG("VGG")

train(model)

你可能感兴趣的:(PaddlePaddle动态图实现VGG(眼底筛查为例))