基于pytorch故障诊断深度学习中的小utils（新手版自用）_pytorch故障诊断测试

1.metrics 指标函数

def recur(fn, input, *args):
    if isinstance(input, torch.Tensor) or isinstance(input, np.ndarray):  # tensor
        output = fn(input, *args)
    elif isinstance(input, list):  # list
        output = []
        for i in range(len(input)):
            output.append(recur(fn, input[i], *args))
    elif isinstance(input, tuple):  # tuple
        output = []
        for i in range(len(input)):
            output.append(recur(fn, input[i], *args))
        output = tuple(output)
    elif isinstance(input, dict):  # dict
        output = {}
        for key in input:
            output[key] = recur(fn, input[key], *args)
    elif isinstance(input, str):  # str
        output = input
    elif input is None:
        output = None
    else:
        raise ValueError('Not valid input type')
    return output
def Accuracy(output, target, topk=1):
    with torch.no_grad():
        if target.dtype != torch.int64:
            target = (target.topk(1, 1, True, True)[1]).view(-1)
        batch_size = target.size(0)
        pred_k = output.topk(topk, 1, True, True)[1]
        correct_k = pred_k.eq(target.view(-1, 1).expand_as(pred_k)).float().sum()
        acc = (correct_k * (100.0 / batch_size)).item()
    return acc
 
 
def MAccuracy(output, target, mask, topk=1):
    if torch.any(mask):
        output = output[mask]
        target = target[mask]
        acc = Accuracy(output, target, topk)
    else:
        acc = 0
    return acc
 
 
def LabelRatio(mask):
    with torch.no_grad():
        lr = mask.float().mean().item()
    return lr
 
 
class Metric(object):
    def __init__(self, metric_name):
        self.metric_name = self.make_metric_name(metric_name)
        self.metric = {'Loss': (lambda input, output: output['loss'].item()),
                       'Accuracy': (lambda input, output: recur(Accuracy, output['target'], input['target'])),
                       'PAccuracy': (lambda input, output: recur(Accuracy, output['target'], input['target'])),
                       'MAccuracy': (lambda input, output: recur(MAccuracy, output['target'], input['target'],
                                                                 output['mask'])),
                       'LabelRatio': (lambda input, output: recur(LabelRatio, output['mask']))}
 
    def make_metric_name(self, metric_name):
        return metric_name
 
    def evaluate(self, metric_names, input, output):
        evaluation = {}
        for metric_name in metric_names:
            evaluation[metric_name] = self.metric[metric_name](input, output)
        return evaluation
 
    def compare(self, val):
        if self.pivot_direction == 'down':
            compared = self.pivot > val
        elif self.pivot_direction == 'up':
            compared = self.pivot < val
        else:
            raise ValueError('Not valid pivot direction')
        return compared
 
    def update(self, val):
        self.pivot = val
        return

========================================================================

2.Python collections模块之Counter

Counter()

主要功能：可以支持方便、快速的计数，将元素数量统计，然后计数并返回一个字典，键为元素，值为元素个数。

from collections import Counter
 
list1 = ["a", "a", "a", "b", "c", "c", "f", "g", "g", "g", "f"]
dic = Counter(list1)
print(dic)
#结果:次数是从高到低的
#Counter({'a': 3, 'g': 3, 'c': 2, 'f': 2, 'b': 1})
 
print(dict(dic))
#结果:按字母顺序排序的
#{'a': 3, 'b': 1, 'c': 2, 'f': 2, 'g': 3}
 
print(dic.items()) #dic.items()获取字典的key和value
#结果:按字母顺序排序的
#dict_items([('a', 3), ('b', 1), ('c', 2), ('f', 2), ('g', 3)])
 
print(dic.keys())
#结果:
#dict_keys(['a', 'b', 'c', 'f', 'g'])
 
print(dic.values())
#结果：
#dict_values([3, 1, 2, 2, 3])
 
print(sorted(dic.items(), key=lambda s: (-s[1])))
#结果:按统计次数降序排序
#[('a', 3), ('g', 3), ('c', 2), ('f', 2), ('b', 1)]
 
for i, v in dic.items():
    if v == 1:
        print(i)
#结果:
#b

from collections import Counter
 
str1 = "aabbfkrigbgsejaae"
print(Counter(str1))
print(dict(Counter(str1)))
#结果:
#Counter({'a': 4, 'b': 3, 'g': 2, 'e': 2, 'f': 1, 'k': 1, 'r': 1, 'i': 1, 's': 1, 'j': 1})
#{'a': 4, 'b': 3, 'f': 1, 'k': 1, 'r': 1, 'i': 1, 'g': 2, 's': 1, 'e': 2, 'j': 1}
 
dic1 = {'a': 3, 'b': 4, 'c': 0, 'd': -2}
print(Counter(dic1))

python中的pop()函数

pop()

用于删除并返回列表中的一个元素（默认为最后一个元素）

>>> list1 = [1,2,4,"hello","xy","你好"]
>>> a = list1.pop()#默认弹出最后一个元素
>>> print(a，list1)
你好 [1,2,4,"hello","xy"]

>>> list2 = [1,2,4,"hello","xy","你好"]
>>> b = list2.pop(3)#弹出列表中第四个元素
>>> print(b,list2)
hello [1,2,4,"xy","你好"]

张量操作：

torch.cat() 功能：将张量按维度dim进行拼接

torch.stack() 功能：在新创建的维度dim上进行拼接

---------------------------------------------------

从list中随机抽取元素的方法

1.随机抽取一个元素

from random import choice
l = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] 
print(choice(l)) # 随机抽取一个

 2.随机抽取若干的元素（无重复）

from random import sample
l = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] 
print(sample(l, 5)) # 随机抽取5个元素

3.随机抽取若干个元素（有重复）

import numpy as np
l = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] 
idxs = np.random.randint(0, len(l), size=5) # 生成长度为5的随机数组，范围为 [0,10)，作为索引
print([l[i] for i in idxs]) # 按照索引，去l中获取到对应的值

----------------------------------------------------

打乱列表的顺序

import random
 
x = [i for i in range(10)]
print(x)
random.shuffle(x)
print(x)

---------------------------------------------------- 

set()

set() 函数创建一个无序不重复元素集，可进行关系测试，删除重复数据，还可以计算交集、差集、并集等。

----------------------------------------------------  

-------------------------------------------------- 

iid数据划分：