宠物分辨使用resnet-34迁移学习

#在执行代码时自动载入需要的模块
%reload_ext autoreload
%autoreload 2
%matplotlib inline 
# 将matplotlib的图表直接嵌入到Notebook

from fastai.vision import * # 导入包
from fastai.metrics import error_rate # 导入评价指标

path = untar_data(URLs.PETS) #给出文件路径

WindowsPath('C:/Users/Wither8848/.fastai/data/oxford-iiit-pet')

path.ls()#查看该路径下所有的文件

[WindowsPath('C:/Users/Wither8848/.fastai/data/oxford-iiit-pet/annotations'),
 WindowsPath('C:/Users/Wither8848/.fastai/data/oxford-iiit-pet/crappy'),
 WindowsPath('C:/Users/Wither8848/.fastai/data/oxford-iiit-pet/images')]

path_img = path/'images'

fnames = get_image_files(path_img)#得到图片
fnames[:5]

[WindowsPath('C:/Users/Wither8848/.fastai/data/oxford-iiit-pet/images/Abyssinian_1.jpg'),
 WindowsPath('C:/Users/Wither8848/.fastai/data/oxford-iiit-pet/images/Abyssinian_10.jpg'),
 WindowsPath('C:/Users/Wither8848/.fastai/data/oxford-iiit-pet/images/Abyssinian_100.jpg'),
 WindowsPath('C:/Users/Wither8848/.fastai/data/oxford-iiit-pet/images/Abyssinian_101.jpg'),
 WindowsPath('C:/Users/Wither8848/.fastai/data/oxford-iiit-pet/images/Abyssinian_102.jpg')]

np.random.seed(2)
pat = r'/([^/]+)_\d+.jpg$'#正则化表达式
bs=64

data = ImageDataBunch.from_name_re(path_img, fnames, pat, ds_tfms=get_transforms(), size=224, bs=bs).normalize(imagenet_stats)

data.show_batch(rows=4, figsize=(7,7))

print(data.classes)#类别名字
data.c#类别总数

['Abyssinian', 'Bengal', 'Birman', 'Bombay', 'British_Shorthair', 'Egyptian_Mau', 'Maine_Coon', 'Persian', 'Ragdoll', 'Russian_Blue', 'Siamese', 'Sphynx', 'american_bulldog', 'american_pit_bull_terrier', 'basset_hound', 'beagle', 'boxer', 'chihuahua', 'english_cocker_spaniel', 'english_setter', 'german_shorthaired', 'great_pyrenees', 'havanese', 'japanese_chin', 'keeshond', 'leonberger', 'miniature_pinscher', 'newfoundland', 'pomeranian', 'pug', 'saint_bernard', 'samoyed', 'scottish_terrier', 'shiba_inu', 'staffordshire_bull_terrier', 'wheaten_terrier', 'yorkshire_terrier']





37

'''迁移学习'''
learn = cnn_learner(data,models.resnet34,metrics = error_rate)
learn.model
learn.fit_one_cycle(4)
learn.save('stage')

'''查看错误分类的结果'''
interp = ClassificationInterpretation.from_learner(learn)
interp.plot_top_losses(9,figsize=(15,11))#查看损失最高的图片

interp.plot_confusion_matrix(figsize=(12,12), dpi=60)#查看混淆矩阵
interp.most_confused(min_val=2)#查看分错的类别

[('staffordshire_bull_terrier', 'american_pit_bull_terrier', 6),
 ('Abyssinian', 'Bengal', 5),
 ('Maine_Coon', 'Ragdoll', 5),
 ('chihuahua', 'miniature_pinscher', 5),
 ('Birman', 'Ragdoll', 4),
 ('British_Shorthair', 'Russian_Blue', 4),
 ('Ragdoll', 'Birman', 4),
 ('Egyptian_Mau', 'Bengal', 3),
 ('Russian_Blue', 'Bombay', 3),
 ('american_bulldog', 'staffordshire_bull_terrier', 3),
 ('english_cocker_spaniel', 'english_setter', 3),
 ('Abyssinian', 'Sphynx', 2),
 ('Bengal', 'Egyptian_Mau', 2),
 ('Birman', 'Siamese', 2),
 ('British_Shorthair', 'Bombay', 2),
 ('Maine_Coon', 'Bengal', 2),
 ('Maine_Coon', 'Persian', 2),
 ('Ragdoll', 'Persian', 2),
 ('Russian_Blue', 'Abyssinian', 2),
 ('Siamese', 'Birman', 2),
 ('Sphynx', 'chihuahua', 2),
 ('american_bulldog', 'boxer', 2),
 ('american_pit_bull_terrier', 'american_bulldog', 2),
 ('american_pit_bull_terrier', 'boxer', 2),
 ('american_pit_bull_terrier', 'staffordshire_bull_terrier', 2),
 ('beagle', 'basset_hound', 2),
 ('boxer', 'american_bulldog', 2),
 ('boxer', 'saint_bernard', 2),
 ('chihuahua', 'american_pit_bull_terrier', 2),
 ('german_shorthaired', 'american_pit_bull_terrier', 2),
 ('german_shorthaired', 'staffordshire_bull_terrier', 2),
 ('miniature_pinscher', 'american_pit_bull_terrier', 2),
 ('saint_bernard', 'boxer', 2),
 ('staffordshire_bull_terrier', 'american_bulldog', 2)]

'''参数优化'''
learn.unfreeze()
learn.fit_one_cycle(1)#结果并没有变好，因为学习率太大

epoch     train_loss  valid_loss  error_rate  time    
0         0.685388    0.372840    0.117727    01:17     

learn.lr_find()
learn.recorder.plot()#找出学习率

epoch     train_loss  valid_loss  error_rate  time    
0         1.473476    #na#        00:43       
LR Finder is complete, type {learner_name}.recorder.plot() to see the graph.


F:\software\programming_software\Anaconda3\lib\site-packages\fastai\sixel.py:16: UserWarning: You could see this plot with `libsixel`. See https://github.com/saitoha/libsixel
  warn("You could see this plot with `libsixel`. See https://github.com/saitoha/libsixel")

learn.unfreeze()
learn.fit_one_cycle(2, max_lr=slice(1e-6,1e-4))

epoch     train_loss  valid_loss  error_rate  time    
0         0.379295    0.312089    0.096752    01:17     
1         0.327472    0.293061    0.094723    01:16     

learn.save('stage-2')