[BUG] Large memory requirements for SimpleImputer strategy median #4794
Labels
Comments
erikrene
added
? - Needs Triage
beckernick
changed the title
|
Thanks for filing an issue. I observe the example above (~2.16GB array) spiking memory to ~18GB on my machine in 22.06. As a note, you can edit your issues if they are accidentally filed without a title or for any other reason. I've updated this issue to better reflect the behavior, which helps us evaluate. |
This was referenced Jul 13, 2022
|
This issue has been labeled |
rapids-bot bot
pushed a commit
that referenced
this issue
Aug 11, 2022
… (#4817) I have implemented a fix for [BUG] Large memory requirements for SimpleImputer strategy median #4794. I narrowed down the issue to _masked_column_median. As expected, the extra memory results from the unnecessary copy of the array (in the case where NaN is the masked value). However, in the other case (where NaN isn't the masked value) this copy is necessary. To fix this, I used in-place sorting. However, in both cases the memory usage goes from 3000 MiB (size of original array) to 13000. From my understanding, sorting should only take up an additional 3000 MiB. Is it possible to reduce memory usage further? Still, this fix still reduces the memory used by over 5000 MiB. Authors: - https://github.com/erikrene Approvers: - William Hicks (https://github.com/wphicks) URL: #4817
jakirkham
pushed a commit
to jakirkham/cuml
that referenced
this issue
Feb 27, 2023
…idsai#4794 (rapidsai#4817) I have implemented a fix for [BUG] Large memory requirements for SimpleImputer strategy median rapidsai#4794. I narrowed down the issue to _masked_column_median. As expected, the extra memory results from the unnecessary copy of the array (in the case where NaN is the masked value). However, in the other case (where NaN isn't the masked value) this copy is necessary. To fix this, I used in-place sorting. However, in both cases the memory usage goes from 3000 MiB (size of original array) to 13000. From my understanding, sorting should only take up an additional 3000 MiB. Is it possible to reduce memory usage further? Still, this fix still reduces the memory used by over 5000 MiB. Authors: - https://github.com/erikrene Approvers: - William Hicks (https://github.com/wphicks) URL: rapidsai#4817
|
This was resolved by #4817 . We now generally require less memory than the CPU scikit-learn version. Closing. %load_ext gpu_memory_profiler
%load_ext memory_profiler
from sklearn.impute import SimpleImputer as skl_SimpleImputer
from cuml.preprocessing import SimpleImputer as cu_SimpleImputer
from sklearn.datasets import make_classification
import numpy as np
import cupy as cp
import gc
NROWS = [
64e6,
128e6,
256e6,
]
NROWS = [int(x) for x in NROWS]
NULL_PCT = [
0.1,
]
for N in NROWS:
for NP in NULL_PCT:
# Create some data and randomly set some elements as null
X = np.random.normal(0, 10, size=(N, 1))
mask = np.random.choice([True, False], size=X.shape, p=[NP, 1-NP])
X[mask] = None
# Compare peak memory usage on GPU and CPU
print(f"{N:,} rows, {NP} null percent, X size: {X.nbytes/1e9} GB")
imputer = cu_SimpleImputer(strategy='median')
%gpu_memit imputer.fit(X)
imputer = skl_SimpleImputer(strategy='median')
%memit imputer.fit(X)
print()
del X
gc.collect()
64,000,000 rows, 0.1 null percent, X size: 0.512 GB
Peak GPU memory: 4225.00 MiB
peak memory: 6078.40 MiB, increment: 2136.23 MiB
128,000,000 rows, 0.1 null percent, X size: 1.024 GB
Peak GPU memory: 7219.00 MiB
peak memory: 9787.75 MiB, increment: 4272.26 MiB
256,000,000 rows, 0.1 null percent, X size: 2.048 GB
Peak GPU memory: 13203.00 MiB
peak memory: 17206.98 MiB, increment: 8544.85 MiB |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Describe the bug
Running fit_transform with SimpleImputer results in running out of memory. This only occurs when the imputation strategy is median.
Steps/Code to reproduce bug
Output:
Expected behavior
This operation should not take as much memory as it does. Using a smaller array does not result in the error. After running nvidia-smi, the memory usage increases about 5x after running the code above.
Environment details (please complete the following information):
The text was updated successfully, but these errors were encountered: