PyGrinder

0.7 verified Mon Apr 27 auth: no python

A Python toolkit for introducing missing values into datasets under various missingness mechanisms (MCAR, MAR, MNAR) and patterns (point, sequential, block). Current version: 0.7. Requires Python >=3.8. Released quarterly.

pip install pygrinder

Common errors

error AttributeError: module 'pygrinder' has no attribute 'calc_misssing_rate' ↓

cause Typo in function name; the correct name is calc_missing_rate (single 's' after 'miss').

fix

Use from pygrinder import calc_missing_rate

error TypeError: mcar() got multiple values for argument 'rate' ↓

cause Passing both positional and keyword argument for 'rate', or duplicate keyword.

fix

Call mcar(X, rate=0.2) with only one rate specification.

error ValueError: seq_len must be less than n_steps ↓

cause In seq_missing, seq_len equals or exceeds the number of time steps, causing an empty index list.

fix

Set seq_len < n_steps, or upgrade to pygrinder >=0.6.4 where this condition is handled.

Warnings

breaking In v0.4, all missingness-creating functions (mcar, mar_logistic, mnar_x, mnar_t) changed to return only the corrupted data (with NaN), not a tuple (X_intact, X, mask). Use fill_and_get_mask to get the mask. ↓

fix Update code: data = mcar(X, rate=0.2) instead of X_intact, data, mask = mcar(X, rate=0.2). Use fill_and_get_mask(data) to get mask.

deprecated In v0.7, mnar_num has been renamed to mnar_nonuniform. The old name mnar_num is removed. ↓

fix Replace mnar_num with mnar_nonuniform.

gotcha Functions seq_missing and block_missing require seq_len argument; setting seq_len equal to n_steps can cause an empty step_idx list bug (fixed in v0.6.4). Verify that seq_len < n_steps to avoid errors. ↓

fix Upgrade to >=0.6.4 or ensure seq_len < n_steps.

gotcha Mar_logistic had a bug in argument order (v0.6.2 and earlier) that could produce incorrect missingness. Fixed in v0.6.3. ↓

fix Upgrade to >=0.6.3.

deprecated In v0.4, the parameter 'return_masks' was available to return masks alongside corrupted data; it was removed in later versions. Use fill_and_get_mask instead. ↓

fix Switch to fill_and_get_mask to retrieve the mask.

Imports

mcar
```
from pygrinder import mcar
```
Earlier versions (pre-0.4) returned masks; now returns only corrupted data.
mar_logistic
```
from pygrinder import mar_logistic
```
Renamed from mar_logistic (always existed).
mnar_x
```
from pygrinder import mnar_x
```
Renamed from mnar_x (always existed).
mnar_t
```
from pygrinder import mnar_t
```
Renamed from mnar_t (always existed).
mnar_nonuniform
```
from pygrinder import mnar_nonuniform
```
Added in v0.7, replaces mnar_num.
seq_missing
```
from pygrinder import seq_missing
```
Added in v0.6.
block_missing
```
from pygrinder import block_missing
```
Added in v0.6.
rdo
```
from pygrinder import rdo
```
Added in v0.5.
fill_and_get_mask
```
from pygrinder import fill_and_get_mask
```
Added in v0.4.
little_mcar_test
```
from pygrinder import little_mcar_test
```
Added in v0.2.
calc_missing_rate
wrong
```
from pygrinder import calc_misssing_rate
```
correct
```
from pygrinder import calc_missing_rate
```
Typo in v0.6.1 used 'calc_misssing_rate' (extra 's'), corrected to 'calc_missing_rate' from v0.6.2 onwards.

Quickstart

Basic usage: create MCAR missingness, then fill NaNs and get mask.

import numpy as np
from pygrinder import mcar, fill_and_get_mask

X = np.random.randn(100, 10)
corrupted_X = mcar(X, rate=0.2)
X_filled, mask = fill_and_get_mask(corrupted_X, fill_value=0)
print(f"Original shape: {X.shape}, missing rate: {1 - mask.mean():.2f}")