从dataframe列获取以前的值

Question

我对Python还不熟悉。现在，我有一个数据帧，看起来像这样。你知道吗

DocumentId                    offset        feature
0                              0              2000             
0                              7              2000
0                              16             0
0                              27             0
0                              36             0 
0                              40             0
0                              46             0
0                              57             0
0                              63             0
0                              78             0
0                              88             0
0                              91             0   
0                              103           2200
1                               109           0
1                               113           2200
1                               126           2200  
1                               131           2200 
1                                137           2200
1                                 142            0
1                                 152            0
1                                 157           200 
1                                159           200
1                                 161           200
1                                   167            0
1                                 170           200

现在，在这个数据框中，我有一个名为feature的列。它有一些值，其中也有一个值0，其他值是positive values。你知道吗

现在，每当正值变为零值，我们就要取它的前三个值。所以，我们假设它总是从正值开始。你知道吗

所以从这个数据集中

我会先得到结果的

 offset1     feature   offset2   feature   offset3   feature   
       -           -         0        2000       7       2000  

   for the `(-)` as it does not have the third value.

现在，它继续检查值，每当值从正到零变为正时，我就要取前三个值。你知道吗

所以，在这里

after 2200 which is at offset 103 value becomes the 0 which is at 109

现在，我必须取109 offset的前三个值，这样数据就变成

offset1     feature    offset2    feature   offset3   feature   
   -           -          0        2000       7       2000  
   103        2200        91         0         88         0   

在这之后，从正到0的值变为偏移量142，因此，我必须取前三个。你知道吗

offset1     feature    offset2    feature   offset3   feature   
       -           -          0        2000       7       2000  
       103        2200        91         0         88         0 
       137        2200        131      2200       126      2200
       161        200         159       200       157      200



offset4       feature      offset5  feature      offset6   feature
     103           2200         109       0           113       2200
      113          2200          126      2200        131       2200
      157          200           159      200         161        200  

所以，如果我们在数据帧中看到一个正数，如果我们向下，那么在

the place `103` the 0 becomes a positive value which is the 2200

现在，我取偏移量的下两个值.which are 109 and 113 and also the current number value which will be 103

因此，结果将是

offset4       feature      offset5  feature      offset6   feature
     103           2200         109       0           113       2200

Now, after this If I go down, then immediately at 113 the previous value was 0 that gets changed to 2200

所以，我将取113后面的两个值，也就是126，131，那么下一个值就变成了

offset4      feature4      offset5   feature5    offset5    feature5
113          2200          126       2200         131       2200

Now again down then at 157 it gets changed to the 200 as so it will be same。你知道吗

对于下一个值，我取下两个值和它本身，所以它变成3个值。在这里，每当值以这种方式从0 to a positive更改时，我就取下一个值。你知道吗

我有什么办法能达到这个结果吗？谢谢。你知道吗

我试过的是

zero_indexes = list(input_with[input_csv['RFC_PREDICTEDFEATURE'] == 0].index)
df2 = pd.DataFrame()
for each_zero_index in zero_indexes:
    value = input_with['feature'].loc[each_zero_index - 1: each_zero_index].values[0]
    if value != 0 :
        df1 = input_with.loc[each_zero_index - 3: each_zero_index]
        df2 = df2.append(df1) 

Answer 1

使用strides：

def rolling_window(a, window):
    shape = a.shape[:-1] + (a.shape[-1] - window + 1, window)
    strides = a.strides + (a.strides[-1],)
    return np.lib.stride_tricks.as_strided(a, shape=shape, strides=strides)

n = 3
x = np.concatenate([[np.nan] * (n - 1), df['feature'].values])
#change order by indexing [:, ::-1]
arr1 = rolling_window(x, n)[:, ::-1]

---

print (arr1)
[[2000.   nan   nan]
 [2000. 2000.   nan]
 [   0. 2000. 2000.]
 [   0.    0. 2000.]
 [   0.    0.    0.]
 [   0.    0.    0.]
 [   0.    0.    0.]
 [   0.    0.    0.]
 [   0.    0.    0.]
 [   0.    0.    0.]
 [   0.    0.    0.]
 [   0.    0.    0.]
 [2200.    0.    0.]
 [   0. 2200.    0.]
 [2200.    0. 2200.]
 [2200. 2200.    0.]
 [2200. 2200. 2200.]
 [2200. 2200. 2200.]
 [   0. 2200. 2200.]
 [   0.    0. 2200.]
 [ 200.    0.    0.]
 [ 200.  200.    0.]
 [ 200.  200.  200.]
 [   0.  200.  200.]
 [ 200.    0.  200.]]

---

x = np.concatenate([[np.nan] * (n - 1), df['offset'].values])
arr2 = rolling_window(x, n)[:, ::-1]

---

print (arr2)
[[  0.  nan  nan]
 [  7.   0.  nan]
 [ 16.   7.   0.]
 [ 27.  16.   7.]
 [ 36.  27.  16.]
 [ 40.  36.  27.]
 [ 46.  40.  36.]
 [ 57.  46.  40.]
 [ 63.  57.  46.]
 [ 78.  63.  57.]
 [ 88.  78.  63.]
 [ 91.  88.  78.]
 [103.  91.  88.]
 [109. 103.  91.]
 [113. 109. 103.]
 [126. 113. 109.]
 [131. 126. 113.]
 [137. 131. 126.]
 [142. 137. 131.]
 [152. 142. 137.]
 [157. 152. 142.]
 [159. 157. 152.]
 [161. 159. 157.]
 [167. 161. 159.]
 [170. 167. 161.]]

---

然后在arr1的第一个'column'中为正值和下一个0值创建掩码：

m =  np.append(arr1[1:, 0] == 0, False) & (arr1[:, 0] != 0)
arr1 = arr1[m] 
arr2 = arr2[m] 

#change order in first row
arr1[:1] = arr1[:1, ::-1]
arr2[:1] = arr2[:1, ::-1]


#create DataFrames and join together    
df1 = pd.DataFrame(arr1).add_prefix('feature_')
df2 = pd.DataFrame(arr2).add_prefix('offset_')
print (df1)
   feature_0  feature_1  feature_2
0        NaN     2000.0     2000.0
1     2200.0        0.0        0.0
2     2200.0     2200.0     2200.0
3      200.0      200.0      200.0

print (df2)
   offset_0  offset_1  offset_2
0       NaN       0.0       7.0
1     103.0      91.0      88.0
2     137.0     131.0     126.0
3     161.0     159.0     157.0

---

#https://stackoverflow.com/a/45122187/2901002
#change order of columns
a = np.arange(n).astype(str)
cols = [item for x in a for item in ('offset_' + x, 'feature_' + x)]
df = pd.concat([df1, df2], axis=1)[cols]
print (df)
   offset_0  feature_0  offset_1  feature_1  offset_2  feature_2
0       NaN        NaN       0.0     2000.0       7.0     2000.0
1     103.0     2200.0      91.0        0.0      88.0        0.0
2     137.0     2200.0     131.0     2200.0     126.0     2200.0
3     161.0      200.0     159.0      200.0     157.0      200.0

编辑：

def rolling_window(a, window):
    shape = a.shape[:-1] + (a.shape[-1] - window + 1, window)
    strides = a.strides + (a.strides[-1],)
    return np.lib.stride_tricks.as_strided(a, shape=shape, strides=strides)

n = 3
x = np.concatenate([[np.nan] * (n - 1), df['feature'].values])
arr1 = rolling_window(x, n)[:, ::-1]
#print (arr1)

x = np.concatenate([[np.nan] * (n - 1), df['offset'].values])
arr2 = rolling_window(x, n)[:, ::-1]
#print (arr2)

x1 = np.concatenate([arr1[:, 0], [np.nan] * (n - 1)])
arr11 = rolling_window(x1, n)
#print (arr11)

x2 = np.concatenate([arr2[:, 0], [np.nan] * (n - 1)])
arr22 = rolling_window(x2, n)
#print (arr22)

---

m1 =  np.append(False, arr11[:-1, 0] == 0) & (arr11[:, 0] != 0)

arr11 = arr11[m1] 
arr22 = arr22[m1] 

df11 = pd.DataFrame(arr11).rename(columns = lambda x: x + 4).add_prefix('feature_')
df22 = pd.DataFrame(arr22).rename(columns = lambda x: x + 4).add_prefix('offset_')
print (df11)
   feature_4  feature_5  feature_6
0     2200.0        0.0     2200.0
1     2200.0     2200.0     2200.0
2      200.0      200.0      200.0
3      200.0        NaN        NaN

print (df22)
   offset_4  offset_5  offset_6
0     103.0     109.0     113.0
1     113.0     126.0     131.0
2     157.0     159.0     161.0
3     170.0       NaN       NaN

---

a = np.arange(4, n + 4).astype(str)
cols = [item for x in a for item in ('offset_' + x, 'feature_' + x)]
df1 = pd.concat([df11, df22], axis=1)[cols]
print (df1)
   offset_4  feature_4  offset_5  feature_5  offset_6  feature_6
0     103.0     2200.0     109.0        0.0     113.0     2200.0
1     113.0     2200.0     126.0     2200.0     131.0     2200.0
2     157.0      200.0     159.0      200.0     161.0      200.0
3     170.0      200.0       NaN        NaN       NaN        NaN

---

m =  np.append(arr1[1:, 0] == 0, False) & (arr1[:, 0] != 0)
arr1 = arr1[m] 
arr2 = arr2[m] 

arr1[:1] = arr1[:1, ::-1]
arr2[:1] = arr2[:1, ::-1]

---

df1 = pd.DataFrame(arr1).add_prefix('feature_')
df2 = pd.DataFrame(arr2).add_prefix('offset_')
print (df1)
   feature_0  feature_1  feature_2
0        NaN     2000.0     2000.0
1     2200.0        0.0        0.0
2     2200.0     2200.0     2200.0
3      200.0      200.0      200.0

print (df2)
   offset_0  offset_1  offset_2
0       NaN       0.0       7.0
1     103.0      91.0      88.0
2     137.0     131.0     126.0
3     161.0     159.0     157.0

---

a = np.arange(n).astype(str)
cols = [item for x in a for item in ('offset_' + x, 'feature_' + x)]
df = pd.concat([df1, df2], axis=1)[cols]
print (df)
   offset_0  feature_0  offset_1  feature_1  offset_2  feature_2
0       NaN        NaN       0.0     2000.0       7.0     2000.0
1     103.0     2200.0      91.0        0.0      88.0        0.0
2     137.0     2200.0     131.0     2200.0     126.0     2200.0
3     161.0      200.0     159.0      200.0     157.0      200.0