背景：模型错误有标签True，判别负；标签负，判别正。因此需要将模型的True_negative与模型的False_positive的错误预测写为list，并且将相应的错误文件名找出。
目录
一、根据预测结果找出badcase
1.1 写出模型预测结果
1.2 根据结果算出指标
输出多标签分类模型每class指标OP,OR,OF1,CP,CR,CF1
1.3 加入错误样本矩阵
上面，有一个对应numpy矩阵，大小为n_imgs*n_class
我们也可以继续构建这样一个矩阵，是错误样本则对应位置为1，非错误样本对应位置为0
def evaluation_and_badcase(scores_, targets_):
print('evaluation start...')
n, n_class = scores_.shape
print('img_numbers=',n,'n_class=',n_class)
Nc, Np, Ng = np.zeros(n_class), np.zeros(n_class), np.zeros(n_class)
ture_Positive,true_negative,false_positive=np.zeros((n,n_class)),np.zeros((n,n_class)),np.zeros((n,n_class))
cls_P,cls_R,cls_F1=np.zeros(n_class),np.zeros(n_class),np.zeros(n_class)
for k in range(n_class):
scores = scores_[:, k] # all img scores on class_k
targets = targets_[:, k] # all img labels on class_k
targets[targets == -1] = 0 # set img labels from -1 to 0
Ng[k] = np.sum(targets == 1) # ture: all ture labels sum number
Np[k] = np.sum(scores >= 0) # positive: all predict positive sum number
Nc[k] = np.sum(targets * (scores >= 0)) # true_positive: true_positive sum number
ture_Positive[:,k]=np.where(targets * (scores >= 0)==1 ,1,0)
true_negative[:,k]=np.where(targets * (scores < 0)==1 ,1,0)
false_positive[:,k]=np.where((targets==0)*(scores >= 0)==1, 1,0)
cls_P[k]=Nc[k]/Np[k]
cls_R[k]=Nc[k]/Ng[k]
cls_F1[k]=(2 * cls_P[k] * cls_R[k]) / (cls_P[k] + cls_R[k])
Np[Np == 0] = 1
print('np.sum(Nc),true_positive=',np.sum(Nc))
print('np.sum(Np),positive=',np.sum(Np))
print('np.sum(Ng),ture=',np.sum(Ng))
# for all labels num_imgs*n_classes
OP = np.sum(Nc) / np.sum(Np) # precision: true_positive/positive
OR = np.sum(Nc) / np.sum(Ng) # recall: true_positive/true
OF1 = (2 * OP * OR) / (OP + OR) # F1_score: harmonic mean of precision and recall
# average by class
CP = np.sum(Nc / Np) / n_class # precision: true_positive/positive
CR = np.sum(Nc / Ng) / n_class # recall: true_positive/true
CF1 = (2 * CP * CR) / (CP + CR) # F1_score: harmonic mean of precision and recall
return OP, OR, OF1, CP, CR, CF1,cls_P,cls_R,cls_F1,ture_Positive,true_negative,false_positive有三个这样的矩阵，ture_Positive,true_negative,false_positive，其中，true，false表示标签的正负，positive，negative表示网络预测结果的正负。
二、每类的badcase
2.1 根据标签创建list
关于np.where https://www.cnblogs.com/massquantity/p/8908859.html
遍历np.array
https://blog.csdn.net/github_39611196/article/details/79109843
[img_num,class_num]=ture_Positive.shape
for class_idx in range(class_num):
ture_negative_dict[class_idx]=[]
true_positive_dict[class_idx]=[]
false_positive_dict[class_idx]=[]
for img_idx in range(img_num):
for class_idx in range(class_num):
if true_negative[img_idx,class_idx]==1:
true_negative_num=true_negative_num+1
ture_negative_dict[class_idx].append(img_idx)
if ture_Positive[img_idx,class_idx]==1:
true_positive_num=true_positive_num+1
true_positive_dict[class_idx].append(img_idx)
if false_positive[img_idx,class_idx]==1:
false_positive_num=false_positive_num+1
false_positive_dict[class_idx].append(img_idx)2.2 将相应的指引矩阵转为dict
需要将每类的badcase存入一个dict，dict的key为类名，keyValue为图片编号的list
[img_num,class_num]=ture_Positive.shape
for class_idx in range(class_num):
ture_negative_dict[class_idx]=[]
true_positive_dict[class_idx]=[]
false_positive_dict[class_idx]=[]
for img_idx in range(img_num):
for class_idx in range(class_num):
if true_negative[img_idx,class_idx]==1:
true_negative_num=true_negative_num+1
ture_negative_dict[class_idx].append(img_idx)
if ture_Positive[img_idx,class_idx]==1:
true_positive_num=true_positive_num+1
true_positive_dict[class_idx].append(img_idx)
if false_positive[img_idx,class_idx]==1:
false_positive_num=false_positive_num+1
false_positive_dict[class_idx].append(img_idx)
print('true_positive_num',true_positive_num)
print('true_negative_num=',true_negative_num)
print('false_positive_num',false_positive_num)三、文件名写入
3.1 coco的标注文件
查找源码，看到coco的文件在下面这个文件中：
data/data/coco/data/val_anno.jsoncoco_img_list
3.2 coco读取
分别为根据文件位置读取文件，将每个文件作为item
def get_anno(self):
list_path = os.path.join(self.root, 'data', '{}_anno.json'.format(self.phase))
self.img_list = json.load(open(list_path, 'r'))
self.cat2idx = json.load(open(os.path.join(self.root, 'data', 'category.json'), 'r'))
def __len__(self):
return len(self.img_list)
def __getitem__(self, index):
item = self.img_list[index]
return self.get(item)item中的file_name即为我们需要的文件名
def get(self, item):
filename = item['file_name']
labels = sorted(item['labels'])
# fixme==========================
if self.is_grouping:
new_labels = [self.idxmap[l] for l in labels]
labels = sorted(new_labels)
# fixme===============================
img = Image.open(os.path.join(self.root, 'data', '{}2014'.format(self.phase), filename)).convert('RGB')
if self.transform is not None:
img = self.transform(img)
target = np.zeros(self.num_classes, np.float32) - 1
target[labels] = 1
return (img, filename, self.inp), target
3.3 读取后的格式
{'file_name': 'COCO_val2014_000000501919.jpg', 'labels': [8, 49, 68]},
{'file_name': 'COCO_val2014_000000436385.jpg', 'labels': [0, 49]}, {'file_name': 'COCO_val2014_000000043170.jpg', 'labels': [16, 49, 18, 68]}, {'file_name':
'COCO_val2014_000000436387.jpg', 'labels': [49, 18, 31]}, {'file_name': 'COCO_val2014_000000305317.jpg', 'labels': [40, 49]}, {'file_name':
'COCO_val2014_000000174247.jpg', 'labels': [42]}, {'file_name': 'COCO_val2014_000000436392.jpg', 'labels': [57, 75, 53]}, {'file_name':
'COCO_val2014_000000239785.jpg', 'labels': [49, 59]}, {'file_name': 'COCO_val2014_000000575372.jpg', 'labels': [72]}, {'file_name':
'COCO_val2014_000000055280.jpg', 'labels': [25]},四、文件名写入dict
读出文件名，遍历相应的array，写入dict
coco_list_path = os.path.join('data/data/coco/data/val_anno.json')
coco_img_list = json.load(open(coco_list_path, 'r'))
# print(coco_img_list)
ture_negative_name_dict,true_positive_name_dict,false_positive_name_dict={},{},{}
for class_idx in range(class_num):
ture_negative_name_dict[class_idx]=[]
true_positive_name_dict[class_idx]=[]
false_positive_name_dict[class_idx]=[]
for class_idx in range(len(false_positive_dict)):
for object_idx in range(len(false_positive_dict[class_idx])):
file_idx=false_positive_dict[class_idx][object_idx]
item=coco_img_list[file_idx]
false_positive_name_dict[class_idx].append(item['file_name'])
for class_idx in range(len(ture_negative_dict)):
for object_idx in range(len(ture_negative_dict[class_idx])):
file_idx=ture_negative_dict[class_idx][object_idx]
item=coco_img_list[file_idx]
ture_negative_name_dict[class_idx].append(item['file_name'])
with open('checkpoint/coco/resnet101_on_coco/ture_negative_name_dict.pkl','wb') as f:
print('writing checkpoint/coco/resnet101_on_coco/ture_negative_name_dict.pkl')
pickle.dump(ture_negative_name_dict, f)
with open('checkpoint/coco/resnet101_on_coco/false_positive_name_dict.pkl','wb') as f:
print('writing checkpoint/coco/resnet101_on_coco/false_positive_name_dict.pkl')
pickle.dump(false_positive_name_dict, f)最终文件格式，dict格式，前面为class的编号，后面为每个file_name
class_num 4 file_names ['COCO_val2014_000000262274.jpg', 'COCO_val2014_000000461248.jpg', 'COCO_val2014_000000397483.jpg', 'COCO_val2014_000000140020.jpg', 'COCO_val2014_000000399297.jpg', 'COCO_val2014_000000202981.jpg', 'COCO_val2014_000000072873.jpg', 'COCO_val2014_000000074499.jpg', 'COCO_val2014_000000536110.jpg', 'COCO_val2014_000000339705.jpg', 'COCO_val2014_000000144058.jpg',
'COCO_val2014_000000538005.jpg', 'COCO_val2014_000000538394.jpg', 'COCO_val2014_000000408757.jpg', 'COCO_val2014_000000505890.jpg', 'COCO_val2014_000000348475.jpg', 'COCO_val2014_000000480657.jpg', 'COCO_val2014_000000240804.jpg', 'COCO_val2014_000000436605.jpg',
'COCO_val2014_000000226244.jpg', 'COCO_val2014_000000089884.jpg', 'COCO_val2014_000000484816.jpg', 'COCO_val2014_000000306415.jpg', 'COCO_val2014_000000421307.jpg', 'COCO_val2014_000000160239.jpg', 'COCO_val2014_000000234660.jpg', 'COCO_val2014_000000359677.jpg', 'COCO_val2014_000000230247.jpg', 'COCO_val2014_000000361551.jpg', 'COCO_val2014_000000001149.jpg', 'COCO_val2014_000000082462.jpg',
'COCO_val2014_000000104327.jpg', 'COCO_val2014_000000170651.jpg', 'COCO_val2014_000000498687.jpg', 'COCO_val2014_000000500257.jpg', 'COCO_val2014_000000172935.jpg', 'COCO_val2014_000000107554.jpg',
'COCO_val2014_000000050786.jpg', 'COCO_val2014_000000456584.jpg',
'COCO_val2014_000000437859.jpg', 'COCO_val2014_000000214197.jpg', 'COCO_val2014_000000441072.jpg', 'COCO_val2014_000000444100.jpg', 'COCO_val2014_000000510122.jpg', 'COCO_val2014_000000183114.jpg', 'COCO_val2014_000000346621.jpg', 'COCO_val2014_000000382905.jpg', 'COCO_val2014_000000186472.jpg', 'COCO_val2014_000000448786.jpg',
'COCO_val2014_000000449432.jpg', 'COCO_val2014_000000390975.jpg',
'COCO_val2014_000000196415.jpg']
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
