天池：零样本目标识别
天池：零样本目标识别
转载： https://blog.csdn.net/nima1994/article/details/82420637
2018之江杯全球人工智能大赛-零样本图像目标识别
简单数据分析
jupyter：github地址
数据预处理
将label_list和class_wordembeddings合并，处理后结果如标签\t特征。
sample_processing.py
import pandas as pd
dir = "d:/ZSL_ImageGame/DatasetA_train_20180813/"
def words_embed():
"""
处理label_list和class_wordembeddings,文本特征
:return: data/word_embeddings.csv
"""
train_labels = pd.read_csv(dir + "label_list.txt", sep='\t', header=None)
print(train_labels.head())
train_words = pd.read_csv(dir +"class_wordembeddings.txt", sep=' ', header=None)
print(train_words.head())
res = pd.merge(train_labels, train_words, left_on=1, right_on=0)
res = res.drop([1, '1_x', '0_y'], axis=1)
print(res.head())
res.to_csv('output/word_embeddings.csv', index=None, header=None, sep='\t')
def traindata():
"""
连接数据
:return:
"""
train = pd.read_csv(dir + 'train.txt', header=None, sep='\t')
train_attr = pd.read_csv(dir + "attributes_per_class.txt", sep='\t', header=None)
train_words=pd.read_csv('output/word_embeddings.csv',sep='\t',header=None)
print(train.head())
print(train_attr.head())
print(train_words.head())
# res.to_csv('output/traindata.csv',index=None,header=None,sep='\t')
if __name__ == '__main__':
# words_embed()
traindata()- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
- 19
- 20
- 21
- 22
- 23
- 24
- 25
- 26
- 27
- 28
- 29
- 30
- 31
- 32
- 33
- 34
- 35
- 36
- 37
- 38
- 39
选取训练集和验证集
有关表明，submit.txt存放着验证数据。如下代码，其标签ZJL160打错成了ZJL178，当然178还打着178。submit.txt含有标签160~200，所有样本都可在train.txt找到。
| False Label | True Label |
|---|---|
| 178 | 160 |
| 178 | 178 |
import numpy as np
import pandas as pd
src="d:/ZSL_ImageGame/DatasetA_train_20180813/"
valid=pd.read_csv(src+"submit.txt",header=None,sep="\t")
train=pd.read_csv(src+'train.txt',header=None,sep='\t')
print(valid.head())
print(valid.shape)
"""
0 1
0 9f1d3113f1fcb573596ca99ecb712364.jpeg ZJL178
1 9f73904f7a72fa7285b80f2ae8286066.jpeg ZJL178
2 619bf8d90e1fa19a7f2966bd38b27ccd.jpeg ZJL178
3 6773ca5a1a615fc0d67f836e0772ff46.jpeg ZJL178
4 e1badc8feb1e4d4a6e44eb382d13bc24.jpeg ZJL178
(8291, 2)
"""
print(train.head())
print(train.shape)
"""
0 1
0 a6394b0f513290f4651cc46792e5ac86.jpeg ZJL1
1 2fb89ef2ace869d3eb3bdd3afe184e1c.jpeg ZJL1
2 eda9f3bef2bd8da038f6acbc8355fc25.jpeg ZJL1
3 7d93ef45972154aae150b4f9980a79c0.jpeg ZJL1
4 fb901b4f9a8e396c1d0155bccc5e5671.jpeg ZJL1
(38221, 2)
"""
print(sorted([int(str(i)[3:]) for i in set(train.iloc[:,1].values)]))
"""
[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 18, 19, 21, 22, 23, 24, 25, 26, 28, 29, 30, 31, 32, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 135, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 149, 150, 151, 152, 153, 154, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200]
"""
mer=pd.merge(valid,train,left_on=0,right_on=0)
print(mer.head())
print(mer.shape)
"""
0 1_x 1_y
0 9f1d3113f1fcb573596ca99ecb712364.jpeg ZJL178 ZJL160
1 9f73904f7a72fa7285b80f2ae8286066.jpeg ZJL178 ZJL160
2 619bf8d90e1fa19a7f2966bd38b27ccd.jpeg ZJL178 ZJL160
3 6773ca5a1a615fc0d67f836e0772ff46.jpeg ZJL178 ZJL160
4 e1badc8feb1e4d4a6e44eb382d13bc24.jpeg ZJL178 ZJL160
(8291, 3)
"""
# ## 标签ZJL160的打错成了ZJL178
from sklearn.metrics import classification_report
from sklearn.metrics import confusion_matrix
import seaborn as sns
import matplotlib.pyplot as plt
from sklearn.metrics import cohen_kappa_score
true_label=mer.iloc[:,1].values
res=mer.iloc[:,2].values
print()
print(classification_report(true_label,res))
print("kappa: ",cohen_kappa_score(true_label,res))
mat = confusion_matrix(true_label,res)
sns.heatmap(mat,annot=True,square=True,fmt="d")
plt.show()- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
- 19
- 20
- 21
- 22
- 23
- 24
- 25
- 26
- 27
- 28
- 29
- 30
- 31
- 32
- 33
- 34
- 35
- 36
- 37
- 38
- 39
- 40
- 41
- 42
- 43
- 44
- 45
- 46
- 47
- 48
- 49
- 50
- 51
- 52
- 53
- 54
- 55
- 56
- 57
- 58
- 59
- 60
- 61
- 62
- 63
- 64
此处可以删除错打成ZJL178的标签或者将错打的改成168，如下为删除代码：
mer=mer[~((mer.iloc[:,1]=="ZJL178") &(mer.iloc[:,2]=="ZJL160"))]- 1
这样，submit.txt就从8291变成了8094，然后在从训练集删除验证集的部分。通过submit.txt和train.txt merge上看，submit的样本全部来自train.txt（其中的160~200）。干脆简单的，直接从train.txt中选取ZJLXX~ZJLXX作为验证，其他作为训练。
import numpy as np
import pandas as pd
src="d:/ZSL_ImageGame/DatasetA_train_20180813/"
train=pd.read_csv(src+'train.txt',header=None,sep='\t')
# print(train.head())
print("train.txt数量",train.shape)
s= ["ZJL"+str(i) for i in range(196,201)]
# print(s)
zsl_validate=train[train.iloc[:,1].isin(s)].sample(frac=1,random_state=2018)
# print(validate.head())
print("零样本测试集",zsl_validate.shape)
traindata=train[~train.iloc[:,1].isin(s)].sample(frac=1,random_state=2018)
# print("",traindata.shape)
train_img=traindata.iloc[:-1000,:]
print("图片训练集",train_img.shape)
validate_img=traindata.iloc[-1000:,:]
print("图片验证集",validate_img.shape)
zsl_validate.to_csv("../data/zsl_validate.csv", sep="\t", header=None, index=None)
train_img.to_csv("../data/train_img.csv", sep="\t", header=None, index=None)
validate_img.to_csv("../data/validate_img.csv", sep="\t", header=None, index=None)- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
- 19
- 20
- 21
- 22
- 23
- 24
- 25
- 26
- 27
- 28
train.txt数量 (38221, 2)
零样本测试集 (1016, 2)
图片训练集 (36205, 2)
图片验证集 (1000, 2)
实验部分
整体思路如下，KNN使用1近邻： 
实验中我使用train.txt中的2000条数据作为验证集，其他作为训练集来训练CNN模型，分别使用VGG16,Xception,自定义模型微调和重训练；发现验证集的精度0.2多，训练集能达到0.9以上，据说训练样本打的标记不够好。也就是说，图片特征这一步就出现问题了 \wulian \mudenggoudai ， 
映射使用的是神经网络（据说使用岭回归较好），线上精度较差。 
此处仅贴出批量数据训练的flow（个人认为有轻微参考价值）。
import keras
import pandas as pd
import numpy as np
import cv2
from keras_preprocessing.image import ImageDataGenerator
train_attr = pd.read_csv("d:/ZSL_ImageGame/DatasetA_train_20180813/attributes_per_class.txt", sep='\t', header=None)
train_words = pd.read_csv(r"data/word_embeddings.csv", sep='\t', header=None)
labelkeys = sorted(set(train_attr.iloc[:, 0].values.tolist()))
datagen = ImageDataGenerator(rescale=1./255,rotation_range=20,width_shift_range=0.2,height_shift_range=0.2,
shear_range=0.2,zoom_range=0.5,horizontal_flip=True,fill_mode='nearest')
class DataPiple:
def __init__(self,target,imgsize=64,impro=False):
"""
:param target:
:param impro: 是否数据增强
"""
self.target = pd.read_csv(target, header=None, sep='\t').sample(frac=1,random_state=2018)
self.fea_size=len(self.target)
self.impro=impro
self.imgsize=imgsize
def readOne(self,pos):
t=self.target.iloc[pos,:]
im = cv2.imread("d:/ZSL_ImageGame/DatasetA_train_20180813/train/" + t[0])
im = cv2.resize(im, dsize=(self.imgsize, self.imgsize))
attr=train_attr[train_attr[0]==t[1]].values[0,1:]
word=train_words[train_words[0]==t[1]].values[0,1:]
label=np.zeros(shape=(len(labelkeys)),dtype=np.uint8)
label[labelkeys.index(t[1])]=1
return im,attr,word,label
def readFeather(self,pos,size):
ims=[]
attrs=[]
words=[]
labels=[]
for i in range(pos,min(pos+size,self.fea_size)):
im,attr,word,label=self.readOne(i)
ims.append(im)
attrs.append(attr)
words.append(word)
labels.append(label)
ims=np.array(ims)
# 做数据增强
if self.impro == True:
ims=datagen.flow(ims,batch_size=len(ims),shuffle=False).__next__()
else:
ims=ims/255.0
ims = np.array(ims)
attrs=np.array(attrs)
words=np.array(words)
labels=np.array(labels)
return ims,attrs,words,labels
def create_inputs(self,size=64):
while True:
for i in range(0,self.fea_size,size):
ims, attrs, words,labels=self.readFeather(i,size)
# print(ims.shape)
# print(attrs.shape)
# print(words.shape)
yield ims, labels
if __name__ == '__main__':
# print(labelkeys)
dp=DataPiple(target=r"D:\ZSL_ImageGame\DatasetA_train_20180813\train.txt",impro=True)
s=dp.create_inputs(64)
r,p=s.__next__()
print(len(r),r.shape,p.shape)
import matplotlib.pyplot as plt
plt.imshow(r[2])
plt.show()
# print(labelkeys)- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
- 19
- 20
- 21
- 22
- 23
- 24
- 25
- 26
- 27
- 28
- 29
- 30
- 31
- 32
- 33
- 34
- 35
- 36
- 37
- 38
- 39
- 40
- 41
- 42
- 43
- 44
- 45
- 46
- 47
- 48
- 49
- 50
- 51
- 52
- 53
- 54
- 55
- 56
- 57
- 58
- 59
- 60
- 61
- 62
- 63
- 64
- 65
- 66
- 67
- 68
- 69
- 70
- 71
- 72
- 73
- 74
- 75
- 76
- 77
- 78
- 79
- 80
- 81
- 82
- 83
- 84
- 85
- 86
- 87
- 88
- 89
其他代码放在Github上。第一次参加此种比赛，无论设备（训练真的慢，……，训练差不多我就终止程序了）还是水平，发现真是cai奥，还是学好基础再来闲逛。此记，用以缅怀。