【paddlepaddle速成】paddlepaddle图像分类从模型自定义到测试
这是给大家准备的paddlepaddle与visualdl速成例子
言有三
毕业于中国科学院,计算机视觉方向从业者,有三工作室等创始人
作者 | 言有三(微信号Longlongtogo)
编辑 | 言有三
这一次我们讲讲paddlepadle这个百度开源的机器学习框架,一个图像分类任务从训练到测试出结果的全流程。
将涉及到paddlepaddle和visualdl,git如下:https://github.com/PaddlePaddle
相关的代码、数据都在我们 Git 上,希望大家 Follow 一下这个 Git 项目,后面会持续更新不同框架下的任务。
https://github.com/longpeng2008/LongPeng_ML_Course
01
paddlepaddle是什么
正所谓google有tensorflow,facebook有pytorch,amazon有mxnet,作为国内机器学习的先驱,百度也有PaddlePaddle,其中Paddle即Parallel Distributed Deep Learning(并行分布式深度学习),整体使用起来与tensorflow非常类似。
sudo pip install paddlepaddle
安装就是一条命令,话不多说上代码。
02
paddlepaddle训练
训练包括三部分,数据的定义,网络的定义,以及可视化和模型的存储。
2.1 数据定义
定义一个图像分类任务的dataset如下:
from multiprocessing import cpu_count
import paddle.v2 as paddle
class Dataset:
def __init__(self,cropsize,resizesize):
self.cropsize = cropsize
self.resizesize = resizesize
def train_mapper(self,sample):
img, label = sample
img = paddle.image.load_image(img)
img = paddle.image.simple_transform(img, self.resizesize, self.cropsize, True)
#print "train_mapper",img.shape,label
return img.flatten().astype('float32'), label
def test_mapper(self,sample):
img, label = sample
img = paddle.image.load_image(img)
img = paddle.image.simple_transform(img, self.resizesize, self.cropsize, False)
#print "test_mapper",img.shape,label
return img.flatten().astype('float32'), label
def train_reader(self,train_list, buffered_size=1024):
def reader():
with open(train_list, 'r') as f:
lines = [line.strip() for line in f.readlines()]
print "len of train dataset=",len(lines)
for line in lines:
img_path, lab = line.strip().split(' ')
yield img_path, int(lab)
return paddle.reader.xmap_readers(self.train_mapper, reader,
cpu_count(), buffered_size)
def test_reader(self,test_list, buffered_size=1024):
def reader():
with open(test_list, 'r') as f:
lines = [line.strip() for line in f.readlines()]
print "len of val dataset=",len(lines)
for line in lines:
img_path, lab = line.strip().split(' ')
yield img_path, int(lab)
return paddle.reader.xmap_readers(self.test_mapper, reader,
cpu_count(), buffered_size)
从上面代码可以看出:
(1) 使用了paddle.image.load_image进行图片的读取,
paddle.image.simple_transform进行了简单的图像变换,这里只有图像crop操作,更多的使用可以参考API。
(2) 使用了paddle.reader.xmap_readers进行数据的映射。
2.2 网络定义
# coding=utf-8
import paddle.fluid as fluid
def simplenet(input):
# 定义卷积块
conv1 = fluid.layers.conv2d(input=input, num_filters=12,stride=2,padding=1,filter_size=3,act="relu")
bn1 = fluid.layers.batch_norm(input=conv1)
conv2 = fluid.layers.conv2d(input=bn1, num_filters=12,stride=2,padding=1,filter_size=3,act="relu")
bn2 = fluid.layers.batch_norm(input=conv2)
conv3 = fluid.layers.conv2d(input=bn2, num_filters=12,stride=2,padding=1,filter_size=3,act="relu")
bn3 = fluid.layers.batch_norm(input=conv3)
fc1 = fluid.layers.fc(input=bn3, size=128, act=None)
return fc1,conv1
与之前的caffe,pytorch,tensorflow框架一样,定义了一个3层卷积与2层全连接的网络。为了能够更好的进行可视化,我们使用了PaddlePaddle Fluid,Fluid的设计也是用来让用户像Pytorch和Tensorflow Eager Execution一样可以执行动态计算而不需要创建图。
2.3可视化
paddlepaddle有与之配套使用的可视化框架,即visualdl。
visualdl是百度数据可视化实验室发布的深度学习可视化平台,它的定位与tensorboard很像,可视化内容包含了向量,参数直方图分布,模型结构,图像等功能,以后我们会详细给大家讲述,这次直接在代码中展示如何使用。
安装使用pip install --upgrade visualdl,使用下面的命令可以查看官方例子:
vdl_create_scratch_log
visualDL --logdir ./scratch_log --port 8080
http://127.0.0.1:8080
下面是loss和直方图的查看
在咱们项目中,具体使用方法如下
# 首先定义相关变量
# 创建VisualDL,并指定log存储路径
logdir = "./logs"
logwriter = LogWriter(logdir, sync_cycle=10)
# 创建loss的趋势图
with logwriter.mode("train") as writer:
loss_scalar = writer.scalar("loss")
# 创建acc的趋势图
with logwriter.mode("train") as writer:
acc_scalar = writer.scalar("acc")
# 定义输出频率
num_samples = 4
# 创建卷积层和输出图像的图形化展示
with logwriter.mode("train") as writer:
conv_image = writer.image("conv_image", num_samples, 1)
input_image = writer.image("input_image", num_samples, 1)
# 创建可视化的训练模型结构
with logwriter.mode("train") as writer:
param1_histgram = writer.histogram("param1", 100)
然后在训练过程中进行记录,这是完整的训练代码,红色部分就是记录结果。
# coding=utf-8
import numpy as np
import os
import paddle.fluid as fluid
import paddle.fluid.framework as framework
import paddle.v2 as paddle
from paddle.fluid.initializer import NormalInitializer
from paddle.fluid.param_attr import ParamAttr
from visualdl import LogWriter
from dataset import Dataset
from net_fluid import simplenet
# 创建VisualDL,并指定当前该项目的VisualDL的路径
logdir = "./logs"
logwriter = LogWriter(logdir, sync_cycle=10)
# 创建loss的趋势图
with logwriter.mode("train") as writer:
loss_scalar = writer.scalar("loss")
# 创建acc的趋势图
with logwriter.mode("train") as writer:
acc_scalar = writer.scalar("acc")
# 定义输出频率
num_samples = 4
# 创建卷积层和输出图像的图形化展示
with logwriter.mode("train") as writer:
conv_image = writer.image("conv_image", num_samples, 1)
input_image = writer.image("input_image", num_samples, 1)
# 创建可视化的训练模型结构
with logwriter.mode("train") as writer:
param1_histgram = writer.histogram("param1", 100)
def train(use_cuda, learning_rate, num_passes, BATCH_SIZE=128):
class_dim = 2
image_shape = [3, 48, 48]
image = fluid.layers.data(name='image', shape=image_shape, dtype='float32')
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
net, conv1 = simplenet(image)
# 获取全连接输出
predict = fluid.layers.fc(
input=net,
size=class_dim,
act='softmax',
param_attr=ParamAttr(name="param1", initializer=NormalInitializer()))
# 获取损失
cost = fluid.layers.cross_entropy(input=predict, label=label)
avg_cost = fluid.layers.mean(x=cost)
# 计算batch,从而来求平均的准确率
batch_size = fluid.layers.create_tensor(dtype='int64')
print "batchsize=",batch_size
batch_acc = fluid.layers.accuracy(input=predict, label=label, total=batch_size)
# 定义优化方法
optimizer = fluid.optimizer.Momentum(
learning_rate=learning_rate,
momentum=0.9,
regularization=fluid.regularizer.L2Decay(5 * 1e-5))
opts = optimizer.minimize(avg_cost)
# 是否使用GPU
place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
# 创建调试器
exe = fluid.Executor(place)
# 初始化调试器
exe.run(fluid.default_startup_program())
# 保存结果
model_save_dir = "./models"
# 获取训练数据
resizesize = 60
cropsize = 48
mydata = Dataset(cropsize=cropsize,resizesize=resizesize)
mydatareader = mydata.train_reader(train_list='./all_shuffle_train.txt')
train_reader = paddle.batch(reader=paddle.reader.shuffle(reader=mydatareader,buf_size=50000),batch_size=128)
# 指定数据和label的对应关系
feeder = fluid.DataFeeder(place=place, feed_list=[image, label])
step = 0
sample_num = 0
start_up_program = framework.default_startup_program()
param1_var = start_up_program.global_block().var("param1")
accuracy = fluid.average.WeightedAverage()
# 开始训练,使用循环的方式来指定训多少个Pass
for pass_id in range(num_passes):
# 从训练数据中按照一个个batch来读取数据
accuracy.reset()
for batch_id, data in enumerate(train_reader()):
loss, conv1_out, param1, acc, weight = exe.run(fluid.default_main_program(),
feed=feeder.feed(data),
fetch_list=[avg_cost, conv1, param1_var, batch_acc,batch_size])
accuracy.add(value=acc, weight=weight)
pass_acc = accuracy.eval()
# 重新启动图形化展示组件
if sample_num == 0:
input_image.start_sampling()
conv_image.start_sampling()
# 获取taken
idx1 = input_image.is_sample_taken()
idx2 = conv_image.is_sample_taken()
# 保证它们的taken是一样的
assert idx1 == idx2
idx = idx1
if idx != -1:
# 加载输入图像的数据数据
image_data = data[0][0]
input_image_data = np.transpose(
image_data.reshape(image_shape), axes=[1, 2, 0])
input_image.set_sample(idx, input_image_data.shape,
input_image_data.flatten())
# 加载卷积数据
conv_image_data = conv1_out[0][0]
conv_image.set_sample(idx, conv_image_data.shape,
conv_image_data.flatten())
# 完成输出一次
sample_num += 1
if sample_num % num_samples == 0:
input_image.finish_sampling()
conv_image.finish_sampling()
sample_num = 0
# 加载趋势图的数据
loss_scalar.add_record(step, loss)
acc_scalar.add_record(step, acc)
# 添加模型结构数据
param1_histgram.add_record(step, param1.flatten())
# 输出训练日志
print("loss:" + str(loss) + " acc:" + str(acc) + " pass_acc:" + str(pass_acc))
step += 1
model_path = os.path.join(model_save_dir,str(pass_id))
if not os.path.exists(model_save_dir):
os.mkdir(model_save_dir)
fluid.io.save_inference_model(model_path,['image'],[predict],exe)
if __name__ == '__main__':
# 开始训练
train(use_cuda=False, learning_rate=0.005, num_passes=300)
2.4训练结果
看看acc和loss的曲线,可见已经收敛
03
paddlepaddle测试
训练的时候使用了fluid,测试的时候也需要定义调试器,加载训练好的模型,完整的代码如下
# encoding:utf-8
import sys
import numpy as np
import paddle.v2 as paddle
from PIL import Image
import os
import cv2
# coding=utf-8
import numpy as np
import paddle.fluid as fluid
import paddle.fluid.framework as framework
import paddle.v2 as paddle
from paddle.fluid.initializer import NormalInitializer
from paddle.fluid.param_attr import ParamAttr
from visualdl import LogWriter
from net_fluid import simplenet
if __name__ == "__main__":
# 开始预测
type_size = 2
testsize = 48
imagedir = sys.argv[1]
images = os.listdir(imagedir)
# 定义调试器
save_dirname = "./models/299"
exe = fluid.Executor(fluid.CPUPlace())
inference_scope = fluid.core.Scope()
with fluid.scope_guard(inference_scope):
# 加载模型
[inference_program,feed_target_names,fetch_targets] = fluid.io.load_inference_model(save_dirname,exe)
predicts = np.zeros((type_size,1))
for image in images:
imagepath = os.path.join(imagedir,image)
img = paddle.image.load_image(imagepath)
img = paddle.image.simple_transform(img,testsize,testsize,False)
img = img[np.newaxis,:]
#print img.shape
results = np.argsort(-exe.run(inference_program,feed={feed_target_names[0]:img},
fetch_list=fetch_targets)[0])
label = results[0][0]
predicts[label] += 1
print predicts
由于所有框架的测试流程都差不多,所以就不对每一部分进行解释了,大家可以自行去看代码。