利用MxNet实现图像分类任务
这篇文章将利用MxNet以及其前端gluon 实现一个完整的图像分类任务，其中主要包括以下几个方面：
- 图像I/O
- 搭建网络
- 进行训练
- 验证算法
- 输出结果
1. 训练数据I/O
将处理好的训练数据读入，进行训练。
训练数据的格式基本按照一个子类一个子文件夹的形式保持，具体可以参考MXNet的数据I/O
1.1 程序的第一步，首先导入相关的包
#import some packages
import sys
import collections
import datetime #用于计时
import gluonbook as gb #用于导入一些功能函数
import math
import numpy as np
import mxnet as mx #mxnet
from mxnet import autograd, gluon, init, nd, image #导入自动梯度，gluon前端，图像等模块
from mxnet.gluon import data as gdata, loss as gloss, model_zoo, nn #导入模型相关模块
import os
import shutil #用于预处理复制文件
import zipfile
import matplotlib.pyplot as plt #绘图工具导入
1.2 随后定义精度计算函数、图像增广函数等辅助函数
# 图像增广和辅助函数
# 计算 Average Precision
def calculate_ap(labels, outputs):
cnt = 0
ap = 0.
for label, output in zip(labels, outputs):
for lb, op in zip(label.asnumpy().astype(np.int),
output.asnumpy()):
op_argsort = np.argsort(op)[::-1] #输出排序后的index，最大概率的值对应的index
lb_int = int(lb) #标签对应的整数
ap += 1.0 / (1+list(op_argsort).index(lb_int)) #精度计算 正确的个数
cnt += 1
return ((ap, cnt))
# 训练集图片增广
def transform_train(data, label):
im = data.astype('float32') / 255 #归并到0~1之间
#图像增强的函数组定义，并利用ImageNet的预训练均值、方差归一化输入图像
auglist = image.CreateAugmenter(data_shape=(3, 224, 224), resize=256,
rand_crop=True, rand_mirror=True,
mean = np.array([0.485, 0.456, 0.406]),
std = np.array([0.229, 0.224, 0.225]))
for aug in auglist:
im = aug(im)
im = nd.transpose(im, (2,0,1)) #改变
return (im, nd.array([label]).asscalar())
# 验证集图片增广，没有随机裁剪和翻转
def transform_val(data, label):
im = data.astype('float32') / 255
auglist = image.CreateAugmenter(data_shape=(3, 224, 224), resize=256,
mean = np.array([0.485, 0.456, 0.406]),
std = np.array([0.229, 0.224, 0.225]))
for aug in auglist:
im = aug(im)
im = nd.transpose(im, (2,0,1)) #改变格式为 channel width height
return (im, nd.array([label]).asscalar())
# 在验证集上预测并评估
def validate(net, val_data, ctx):
metric = mx.metric.Accuracy()
L = gluon.loss.SoftmaxCrossEntropyLoss()
AP = 0.
AP_cnt = 0
val_loss = 0
for i, batch in enumerate(val_data):
data = gluon.utils.split_and_load(batch[0], ctx_list=ctx,
batch_axis=0, even_split=False)
label = gluon.utils.split_and_load(batch[1], ctx_list=ctx,
batch_axis=0, even_split=False)
outputs = [net(X) for X in data]
metric.update(label, outputs)
loss = [L(yhat, y) for yhat, y in zip(outputs, label)]
val_loss += sum([l.mean().asscalar() for l in loss]) / len(loss) #平均损失
ap, cnt = calculate_ap(label, outputs)
AP += ap
AP_cnt += cnt #精度也要求平均
_, val_acc = metric.get()
return ((val_acc, AP / AP_cnt, val_loss / len(val_data)))
1.3 读取训练和验证数据
这时候可以利用gluon的内置函数来对数据进行读取了，只需要输入对应数据的文件夹即可，参考MXNet I/O
#读取数据文件
train_set = gdata.vision.ImageFolderDataset('./train_dis/',flag=1)
valid_set = gdata.vision.ImageFolderDataset('./valid_dis/',flag=1)
#check data classes
print(train_set) #check数据的长度是否正确，应为训练图像总数量
print(train_set.synsets) #also has items attributes，现实分类别是否正确，应为类别数目
print(valid_set)
print(valid_set.synsets) #also has items attributes
<mxnet.gluon.data.vision.datasets.ImageFolderDataset object at 0x7fb3d6e06710>
['0', '1', '10', '11', '12', '13', '14', '15', '16', '17', '18', '19', '2', '20', '21', '22', '23', '24', '25', '26', '27', '28', '29', '3', '30', '31', '32', '33', '34', '35', '36', '37', '38', '39', '4', '40', '41', '42', '43', '44', '45', '46', '47', '48', '49', '5', '50', '51', '52', '53', '54', '55', '56', '57', '58', '59', '6', '60', '7', '8', '9']
<mxnet.gluon.data.vision.datasets.ImageFolderDataset object at 0x7fb3d6e06668>
['0', '1', '10', '11', '12', '13', '14', '15', '16', '17', '18', '19', '2', '20', '21', '22', '23', '24', '25', '26', '27', '28', '29', '3', '30', '31', '32', '33', '34', '35', '36', '37', '38', '39', '4', '40', '41', '42', '43', '44', '45', '46', '47', '48', '49', '5', '50', '51', '52', '53', '54', '55', '56', '57', '58', '59', '6', '60', '7', '8', '9']
得到输入序列后，将图像读入迭代器中，根据显存设置批量的大小。
#data into iter and realized argumentation
batch_size = 64 #32--2821M could be 64
train_iter = gdata.DataLoader(train_set.transform(transform_train),
batch_size, shuffle=True, last_batch='keep', num_workers=4)
valid_iter = gdata.DataLoader(valid_set.transform(transform_val),
batch_size, shuffle=True, last_batch='keep', num_workers=4)
读入后check迭代器的数据，并显示目测
# check the data set in iter
print("trainiter lenght is: %d"%len(train_iter))
import matplotlib.pyplot as plt
for imgs, labels in train_iter:
print(labels) #打印label 对应类别label
print(imgs.shape) #查看batch图像的维度
break #读入一个batch
#show images
nor_parms = [[0.485, 0.456, 0.406],[0.229, 0.224, 0.225]]
#_,figs = plt.subplots(8,4,figsize=(8,4))
for i in range(8):
for j in range(4):
x = nd.transpose(imgs[i*4+j,:,:,:],(1,2,0)).asnumpy()
print(x.shape,type(x)) #查看batch中图像的维度和类型
#x[:,:,0]*nor_parms[0][0]+nor_parms[1][0]
#x[:,:,1]*nor_parms[0][1]+nor_parms[1][1]
#x[:,:,2]*nor_parms[0][2]+nor_parms[1][2]
plt.imshow(x)
plt.show()
break
trainiter lenght is: 512 #总共有512个batch，每个batch有64个训练数据
Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers).
[35. 0. 31. 19. 38. 33. 35. 33. 19. 25. 16. 26. 36. 52. 18. 16. 27. 23.
19. 4. 19. 38. 38. 11. 41. 36. 22. 36. 29. 57. 26. 55. 18. 55. 55. 16.
27. 26. 55. 10. 19. 21. 23. 19. 50. 56. 31. 14. 20. 19. 8. 54. 57. 8.
52. 19. 56. 57. 17. 42. 18. 0. 23. 55.]
<NDArray 64 @cpu_shared(0)>
(64, 3, 224, 224)
(224, 224, 3) <class 'numpy.ndarray'>
2.定义模型
这里主要使用迁移学习的方式，利用预训练模型抽取图像的基本特征，而后只需要训练最后的输出层来进行分类。
#define the net work by pre-train
def get_net(ctx):
resnet = model_zoo.vision.resnet50_v2(pretrained=True) #ctx 使用resnet_50作为基本网络抽取特征
resnet.output_new = nn.HybridSequential(prefix='') #output is the origin 得到特征，新定义一个输出
#add two fcn for finetune
resnet.output_new.add(nn.Dense(256,activation = 'relu')) #在模型基础上，定义最后两个全连接层
resnet.output_new.add(nn.Dense(61))
#initialize
resnet.output_new.initialize(init.Xavier(),ctx=ctx) #for fintune
resnet.collect_params().reset_ctx(ctx) #for whole net
return resnet
定义损失函数，这里主要使用分类的softmax交叉熵来作为损失。
#for loss
loss = gloss.SoftmaxCrossEntropyLoss() #分类损失交叉熵
def get_loss(data,net,ctx):
l=0.0 #loss
for X,y in data:
y = y.as_in_context(ctx)
#计算预训练模型输出的特征
out_features = net.features(X.as_in_context(ctx))
outputs = net.output_new(out_features) #final output
l += loss(outputs,y).mean().asscalar() #loss for the process
return l/len(data)
2.1定义训练过程
完成了以上的准备工作，读入了数据、定义好了网络和损失，我们可以开始进行训练了，训练函数定义如下，输入为网络模型、数据、训练epochs、学习率、衰减等：
#def trainning process, trainer, epochscircles, lossback, valide
def train(net,train_iter,valid_iter,num_epochs, lr, wd, ctx, lr_period, lr_decay):
trainer = gluon.Trainer(net.output_new.collect_params(), 'sgd',
{'learning_rate':lr, 'momentum':0.9, 'wd': wd})
plot_loss = [] #plot loss
tic = datetime.datetime.now()
print('Traing is begining, please waiting......')
for epoch in range(num_epochs):
train_l = 0.0 #存储训练loss
counter = 0 #训练batch周期计数器
#if epoch >0 and epoch %lr_period==0: #every period step update lr
trainer.set_learning_rate(trainer.learning_rate*lr_decay) #every steps updata lr
#print("There are %d data could train network"%len(train_iter))
for X,y in train_iter: #X~32(batch)*1024(iter)= 32768
#output for process reminding
counter +=1
if counter % 256 ==0:
print('processd %d images'%(counter*batch_size)) #一定批量就显示处理过程
#output finished
y = y.astype('float32').as_in_context(ctx)
#feature
out_features = net.features(X.as_in_context(ctx)) #预训练直接前传得到特征，未来这一步可以一次性做
#partly training fineturning
with autograd.record():
#features to output, just use features as input
outputs = net.output_new(out_features) #这里只bp最后两层，只训练最后新定义的部分
l = loss(outputs, y)
l.backward()
#for next batch
trainer.step(batch_size)
train_l += l.mean().asscalar()
#log time into
toc = datetime.datetime.now()
h, remainder = divmod((toc - tic).seconds, 3600)
m, s = divmod(remainder, 60)
time_s = "time %02d:%02d:%02d" % (h, m, s)
#validata
if valid_iter is not None: #验证数据，验证训练效果
valid_loss = get_loss(valid_iter, net, ctx)
epoch_s = ("epoch %d, train loss is %f, valid loss is %f :D "
%(epoch+1, train_l/len(train_iter),valid_loss))
else:
epoch_s = ("epoch %d, train loss is %f :D"
%(epoch+1, train_l/len(train_iter)))
tic = toc
print(epoch_s + time_s + ', lr ' + str(trainer.learning_rate))
#plot loss
plot_loss.append(train_l/len(train_iter))
plt.plot(plot_loss) #将损失优化结果保存到图里
plt.savefig("./training_loss.png")
2.2 开始训练
ctx = gb.try_gpu();num_epochs = 1000;lr = 0.01;wd = 1e-4;lr_period = 10;lr_decay = 0.99;
net = get_net(ctx) #将网络和数据定义到gpu上
train(net,train_iter,valid_iter,num_epochs, lr, wd, ctx, lr_period, lr_decay) #训练
net.output_new.collect_params().save('./output_new_2_1000.params') #训练结束后保存参数
#net.output_new.save_params('./output_new_50.params')
Traing is begining, please waiting......
processd xxxxx images
processd xxxxx images
epoch 1, train loss is 1.234988, valid loss is 0.776764 :Dtime 00:04:10, lr 0.0099
3.测试
在训练完成得到模型后，我们需要对数据进行测试。同样需要读入数据，并利用网络进行分类。
#prepaer data
test_set = gdata.vision.ImageFolderDataset('./test_dis/',flag=1)
print("There are %d test imgs"%len(test_set))
There are xxxx test imgs
定义图像读入函数
def plot_image(img_path):
with open(img_path, 'rb') as f:
img = image.imdecode(f.read()) #读入输入
#plt.imshow(img.asnumpy())
return img
接下来就是测试过程了：
#predict process
preds = []
count_p=0
for img_path,label in test_set.items: #将加载列表中每一张测试图进行分类
img = plot_image(img_path)
data, _ = transform_val(img, 0)
data = data.expand_dims(axis=0)
#plt.imshow(img.asnumpy())
#plt.show()
#print(img_path)
#break
# 计算预训练模型输出层的输入，即特征。
output_features = net.features(data.as_in_context(mx.gpu()))
# 将特征作为我们定义的输出网络的输入，计算输出。
output = nd.softmax(net.output_new(output_features))
preds.extend(output.asnumpy())
count_p +=1
#print(count_p)
if count_p%100==0:
print("processed %d imgs"%count_p)
processed 100 imgs
可以根据需要将生成的预测结果preds保存为json文件：
# use the tese_set name and predict results
with open('submission.json', 'w') as f:
f.write("[")
for i in range(len(preds)):
if i==len(preds)-1:
f.write("{"+""image_id": "+"""+test_set.items[i][0].split('/')[-1]+"""+','+""xxxx_class":"+str(preds[i].argmax())+'}')
else:
f.write("{"+""image_id": "+"""+test_set.items[i][0].split('/')[-1]+"""+','+""xxxx_class":"+str(preds[i].argmax())+'}'+',')
f.write("]")
最后检查生成的数据长度，是否和测试集数据长度相同，然后就大功告成啦~~~~
#check format
import json
user_result_list = json.load(open('./submission.json', encoding='utf-8'))
len(user_result_list)