從包含每個標簽的數據框中提取行

Question

在下面顯示的示例dataframe ，我有 5 個標簽（ class_name ）。 總共有 31 個文件（31 行）可以使用...

我正在嘗試提取 80%（可以是可變的）行（= 24 行（整數））。 但是，我想確保從每個class_name中提取至少 1 行

對於我的嘗試，我只能手動執行此操作。 這種方法變得乏味，因為class_name遠不止 10 個。您能幫我提取正確的 % 行以從每個標簽（ class_name ）中至少包含 1 個條目。

這是我的嘗試：

import math
import pandas as pd

base_path = 'G:/PandasFileSeperation'
original_df = pd.read_csv(f'{base_path}/Book2.csv')

original_df = original_df.astype(str)
length = original_df.class_name.count()
length

# Get number of groups
dfg = original_df.groupby('class_name')
numgroups = dfg.ngroups
numgroups

# Get the sizes of each group
group_size =  original_df.groupby('class_name').size()

# Get length of original dataframe
Total_dataset_size = len(original_df)

# Get number of Training samples
TrainPercent = 0.80

Train_size = int(Total_dataset_size * TrainPercent)
Train_size

# How can I change this to automatically change the label size length to give at least 1 row from each class?
Label_0_size = 5
Label_1_size = 3
Label_2_size = 7
Label_3_size = 7
Label_4_size = 2

# Training Dataset
label_percent = { 'pigs' : Label_0_size, 'goats' : Label_1_size, 'chickens' : Label_2_size, 'hens' : Label_3_size, 'sheep' : Label_4_size}

flag = True
for label, num_rows in label_percent.items():
  
  row_num = num_rows
  
  if label == 'pigs':
    row_num0 = Label_0_size
    df = original_df[original_df['class_name'] == label]
    df = df.head(row_num0)

  elif label == 'goats':
    row_num2 = Label_1_size
    df = original_df[original_df['class_name'] == label]
    df = df.head(row_num1)
    
    
  elif label == 'chickens':
    row_num2 = Label_2_size
    df = original_df[original_df['class_name'] == label]
    df = df.head(row_num2)
    
  elif label == 'hens':
    row_num3 = Label_3_size
    df = original_df[original_df['class_name'] == label]
    df = df.head(row_num3)
    
  else:
    row_num4 = Label_4_size
    df = original_df[original_df['class_name'] == label]
    df = df.head(row_num4)

  if flag == True:
    Train_df = df
    flag = False
  else:
    Train_df = pd.concat([Train_df, df])

Train_df.to_csv(f'{base_path}/Train_df.csv', encoding='utf-8')

數據框：

切片文件名	fsID	開始	結尾	顯着性	折疊	班級號	班級名稱	original_class
1-1000020520400.wav					1	1	豬
1-100004024000001.wav					1	1	豬
1-10000406050001.wav					1	1	豬
1-1000050120400.wav					1	1	豬
1-1000050320400.wav					1	1	豬
1-1000050520400.wav					1	2	山羊
1-10000601400001000.wav					1	2	山羊
1-1000060340000.wav					1	2	山羊
1-100006070500.wav					1	3	雞
1-100007020800.wav					1	3	雞
1-100007024000001.wav					1	3	雞
1-1000070320400.wav					1	3	雞
1-100007050800.wav					1	3	雞
1-100007064000001.wav					1	3	雞
1-100010620400.wav					1	3	雞
1-100040620400.wav					1	3	雞
1-10006020500.wav					1	3	雞
1-10006030500.wav					1	3	雞
1-100060520400.wav					1	4	母雞
1-10007020500.wav					1	4	母雞
2-100070420400.wav					1	4	母雞
2-100070540000.wav					1	4	母雞
2-1313131313004.wav					1	4	母雞
2-1313131313043.wav					1	4	母雞
2-1313131313044.wav					1	5	羊
2-150002020500.wav					1	5	羊
2-150002060800.wav					1	5	羊
2-150004022040001.wav					1	5	羊
2-15000406050001.wav					1	5	羊
2-150006014000001.wav					1	5	羊
2-150006024000001.wav					1	5	羊

作為開始，我們可以做一個forloop並從每個類 = 1 開始，然后遞增每個類，在每次迭代后檢查總和，直到總類 = Train_size ？

Answer 1

我認為這應該有所幫助。 對數據幀進行兩次采樣：一次使用groupby為每個名稱采樣一行，然后從數據幀的其余部分隨機采樣以完成 80% 的訓練集。

TrainPercent = 0.8
# sample one row for each class_name (5 rows)
one_each = df.groupby('class_name').sample(n=1)
# from the rest of the rows, sample int(0.8*len(df))-len(one_each) number of rows (19 rows)
rest = df.loc[~df.index.isin(one_each.index)].sample(n=int(TrainPercent*len(df))-len(one_each))
# concatenate the two
res = pd.concat([one_each, rest])