??? ????? ?? GPU ?? ????

Reading Time: 5 minutes

??? ???? ???? ?? ???? ?? ? ????, ????? ?? ??? ??? ???? ? ?? ???? ?????. GPU ???? ?? ?????? ???? ??? ????, ??? ?? ??? ???? ????.

RAPIDS? ???? ?? ???? GPU? ???? ??? ???? ?????? ?????? ??? NVIDIA?? ??? ?? ?? ??????? ????? ??????. ??? ??? ?? cuDF, ?? ??? ?? cuML, ??? ??? ?? cuGraph? ?? ??? ???? RAPIDS? ?? Python ?????? ???? ??? ? ?? ??? ??????? ? ??? ????? ??? ? ??? ??????.

? ?????? ?? ??? ??? ???????? ?? CPU ??? ???? ??????? GPU ?? ?????? ???? ?? ?? ????? ???.

???? ?? ???? ????? RAPIDS ????

RAPIDS? ???? ??? ????? ? ?? ????? ????. ?? ??? ?? ??? ???? ???? ?? RAPIDS ?? ???? ??? ?? ????, pip ??, Docker ??? ?? Conda? ?? ??? ?? ?????? ??? ? ?? ?? ??? ????. ?? ???? ???? RAPIDS? ????? RAPIDS ???? ?? ???? ????? ????. ???? ?? ?? ????? ?? ?? CUDA ??? ???? RAPIDS ??? ???? ???? ????? ???.

pandas? ?? cuDF ? GPU ??

RAPIDS? ??? ???? GPU ?? ?????? ?? ??? ?? ?????? ??? ? ?? ??? ????? ????. ? ? cuDF? ?? pandas ?? ??????? GPU? ???? ??? ??? ???? ??? ? ?? ??? ???, ?? ??? ?? ??? ????.

????? ??? ???? ? ??? ??? GPU?? ???? ?? pandas? ???? ?? cuDF ??? ????? ????, load_ext cudf.pandas? RAPIDS ?? ????? ????, ??? ??? ??? ????? ?? ????? ?? cuDF DataFrame? ??? ???? pandas? ??? ??? ??? ??? ? ????.

?? pandas? ?????, cuDF pandas? .csv, .json, .pickle, .paraquet ? ??? ?? ??? ????? GPU ?? ??? ??? ???? ????.

??? ??? cudf.pandas ???? ????? ? ?? .csv ??? ???? ??? ?????:

%load_ext cudf.pandas

import pandas as pd 
import cupy as cp 
   
train = pd.read_csv('./Titanic/train.csv') 
test = pd.read_csv('./Titanic/test.csv') 
concat = pd.concat([train, test], axis = 0) 

cudf.pandas ???? ???? ?? ???? ??? ??? ???, ???, ??? ?? ??? pandas ??? GPU?? ??? ? ????. cuDF ???? pandas API? ???? CPU?? GPU? ???? ???? ??? ??? ?? ?? ??? ??????.

target_rows = 1_000_000
repeats = -(-target_rows // len(train))  # Ceiling division
train_df = pd.concat([train] * repeats, ignore_index=True).head(target_rows)
print(train_df.shape)  # (1000000, 2)
  
repeats = -(-target_rows // len(test))  # Ceiling division
test_df = pd.concat([test] * repeats, ignore_index=True).head(target_rows)
print(test_df.shape)  # (1000000, 2)
  
combine = [train_df, test_df]
  
  
(1000000, 12)
(1000000, 11)

filtered_df = train_df[(train_df['Age'] > 30) & (train_df['Fare'] > 50)] 
grouped_df = train_df.groupby('Embarked')[['Fare', 'Age']].mean() 
additional_info = pd.DataFrame({ 
    'PassengerId': [1, 2, 3], 
    'VIP_Status': ['No', 'Yes', 'No'] 
  }) 
merged_df = train_df.merge(additional_info, on='PassengerId', 
how='left')

??? ??: CPU ? GPU ??? ??? ?? ??

??? ???? ???? ?? ???? ?? ?? ?? ??? ??? ???? ??? ????? ???? ??? ?????. ???? ??? CPU? GPU ????? ???? ??? ????, ????? ??? ????, ?????? ???? ????? ?? ??? ???? ?? ??? ????.

cudf.pandas.profile? ?? ?? ????? ??? ?? ??? ?? ??? ??? ???? ???? ??? ???. ?? ?? ??? ? ??? ????? CPU?? ??? ??? GPU?? ???? ??? ?????:

%%cudf.pandas.profile
train_df[['Pclass', 'Survived']].groupby(['Pclass'], 
as_index=False).mean().sort_values(by='Survived', ascending=False)

        Pclass    Survived
0         1        0.629592
1         2        0.472810
2         3        0.242378
  
                         Total time elapsed: 5.131 seconds
                         5 GPU function calls in 5.020 seconds
                         0 CPU function calls in 0.000 seconds
  
                                       Stats
  
+------------------------+------------+-------------+------------+------------+-------------+------------+
| Function           | GPU ncalls  | GPU cumtime | GPU percall | CPU ncalls | CPU cumtime | CPU percall |
+------------------------+------------+-------------+------------+------------+-------------+------------+
| DataFrame.__getitem__ | 1          | 5.000       | 5.000      | 0          | 0.000       | 0.000      |
| DataFrame.groupby     | 1          | 0.000       | 0.000      | 0          | 0.000       | 0.000      |
| GroupBy.mean          | 1          | 0.007       | 0.007      | 0          | 0.000       | 0.000      |
| DataFrame.sort_values | 1          | 0.002       | 0.002      | 0          | 0.000       | 0.000      |
| DataFrame.__repr__    | 1          | 0.011       | 0.011      | 0          | 0.000       | 0.000      |
+------------------------+------------+-------------+------------+------------+-------------+------------+

??? ???? ???? ?? cuDF ??, ???? ?? ??? ?? ?? ??? ?? ??? ??? ?? ??? CPU ???? ????? ??? ??? ???? ? ??? ???. ??? ??? ?? ??? ? ??? ?? ? ???? ??? ??? ???? ?? ?????. ? ??(Loader)? ?? ??? ????? GPU ??? ?? cudf.pandas ????? ?????? ?????.

?? %%time ? %%timeit? ?? Python ?? ???(Magic commands)? ???? ?? ?? ??? ????? ????? pandas (CPU)? pandas? cuDF ???(GPU) ?? ???? ?? ??? ? ????. ??? ??? GPU ??? ?? ?? ? ?? ??? ??? ?? ????? ?????. ?? %%time? ??? ????? CPU? GPU ?? ?? ?? ??? ???? ???? ?? ??? ?? ??? ? ?? ??? ??? ?????.

%%time 
   
print("Before", train_df.shape, test_df.shape, combine[0].shape, combine[1].shape) 
   
train_df = train_df.drop(['Ticket', 'Cabin'], axis=1) 
test_df = test_df.drop(['Ticket', 'Cabin'], axis=1) 
combine = [train_df, test_df] 
   
print("After", train_df.shape, test_df.shape, combine[0].shape, combine[1].shape) 

CPU output:
Before (999702, 12) (999856, 11) (999702, 12) (999856, 11)
After  (999702, 10) (999856, 9)  (999702, 10) (999856, 9)
  
CPU times: user 56.6 ms, sys: 8.08 ms, total: 64.7 ms
  
Wall time: 63.3 ms

GPU output:
Before (999702, 12) (999856, 11) (999702, 12) (999856, 11)
After  (999702, 10) (999856, 9)  (999702, 10) (999856, 9)
  
CPU times: user 6.65 ms, sys: 0 ns, total: 6.65 ms
  
Wall time: 5.46 ms

%%time ??? ?? ??? 10? ??? CPU?? 63.3???(ms)?? ? ??? GPU?? 5.46ms? ???????. ?? ??? ??? ???? cuDF pandas? ?? GPU ??? ???? ??????, ?? ???? ???? ???? ???? ?? ?? ??? ???? %%timeit? ???? ? ?? ????? ?? ? ????.

%%timeit  
   
for dataset in combine: 
    dataset['Title'] = dataset.Name.str.extract(' ([A-Za-z]+)\\.', expand=False) 
   
pd.crosstab(train_df['Title'], train_df['Sex']) 

CPU output:
1.11 s ± 7.49 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

GPU output:
89.6 ms ± 959 μs per loop (mean ± std. dev. of 7 runs, 10 loops each)

%%timeit ??? GPU ??? ?? 10?? ?? ??? ?? ??, CPU?? ??? 1.11??? ?? ??? GPU?? ??? 89.6ms? ???????, ?? ???? ??? ??? ?? cuDF pandas? ???? ?????.

GPU ??? ??

?? ?? ??? ???? ??? ?? ???? GPU? ????? ???? ??? ???? ?? ?????. ??? ?? ???(Type command)? ???? ??? CPU?? ???? ??? GPU?? ???? ??? ???? NumPy ??? CuPy ??? ??? ? ????.

type(guess_ages)

<module 'pandas' (ModuleAccelerator(fast=cudf, slow=pandas))>

??? np.array?? ???? CPU?? ???? ?? ????. ??? cupy.ndarray?? ???? GPU?? ???? ?? ????. ? ?? ??? ?? ?????? ??? ??? GPU ???? ???? ??? ??? ? ????.

??, print ??? ???? ??? GPU? ???? ??? ???? cuDF ??????? ???? ??? ??? ? ????. ???? ?? ??(cuDF) ?? ?? ??(pandas)? ?? ??? ??? ???? ????. ? ??? ??? ?? GPU? ??? ?? ???? ?? ????? ??? ?? ??? ? ????.

print(pd)

<module 'pandas' (ModuleAccelerator(fast=cudf, slow=pandas))>

?????, df.info? ?? ??? ???? cuDF ??????? ??? ???? ??? GPU ???? ?????? ??? ? ????. ?? ?? ??? GPU?? ????? ??? CPU? ?????? ??? ? ????.

train_df.info()

<class 'cudf.core.dataframe.DataFrame'>
  
RangeIndex: 1000000 entries, 0 to 999999
Data columns (total 9 columns):
 #   Column   Non-Null Count   Dtype   
---  ------   --------------   ----- 
 0   Survived 1000000 non-null  int64   
 1   Pclass   1000000 non-null  int64   
 2   Sex      1000000 non-null  int64   
 3   Age      1000000 non-null  float64 
 4   SibSp    1000000 non-null  int64   
 5   Parch    1000000 non-null  int64   
 6   Fare     1000000 non-null  float64 
 7   Embarked 997755 non-null   object
 8   Title    1000000 non-null  int64   
dtypes: float64(2), int64(6), object(1)
memory usage: 65.9+ MB

RAPIDS? cuDF pandas? ?? ??? ?? ?? CPU ?? ??? ??????? GPU ?? ??? ???? ???? ??? ?? ??? ??????, %%time, %%timeit, %%cudf.pandas.profile? ?? ????? ??? ?? ??? ???? ??? ???? ???? ???? ? ????. ?? type, print(pd), df.info? ?? ??? ??? ?? GPU ???? ??? ? ?? ?????? GPU ???? ????? ???? ??? ??? ? ????.

???? ??? ??? ??? ??? ?? ???? ?? ???? Jupyter ???? ?????.

GPU ?? ??? ???? ?? ??? ????? ??? ????? ???? 10?: RAPIDS cuDF? CuPy ????? ?? ??? Google Colab?? ??? ?? 50??? ????? RAPIDS cuDF ???? ?????.