PyPOTS

1.4 verified Mon Apr 27 auth: no python

A Python toolbox for machine learning on partially-observed time series. Current version 1.4, released with bug fixes. Active development with frequent releases.

pip install pypots

Common errors

error ValueError: Expected data to be 3D, got 2D ↓

cause Input data is 2D instead of 3D (n_samples, n_timesteps, n_features).

fix

Reshape data to 3D: data = data.reshape(n_samples, -1, n_features) or add a dimension.

error ModuleNotFoundError: No module named 'pypots.data' ↓

cause Trying to import from the old 'pypots.data' subpackage which was removed in v1.0.

fix

Use 'from pypots.utils.random import set_random_seed' or other correct imports from actual subpackages.

error AttributeError: 'NoneType' object has no attribute 'endswith' ↓

cause Known bug in TimeLLM model when loading pretrained LLM, fixed in v1.2.

fix

Upgrade to pypots>=1.2.

Warnings

gotcha PyPOTS requires data in 3D numpy arrays (samples, timesteps, features). Passing 2D or 4D arrays will cause errors. ↓

fix Ensure input shape is (n_samples, n_timesteps, n_features).

deprecated The old import 'from pypots.data import ...' is deprecated. Use 'from pypots.utils.random import ...' instead. ↓

fix Use correct import paths as shown in official docs.

breaking In v1.0, the API changed significantly. Many model classes and utilities moved to new subpackages. ↓

fix Refer to the migration guide at https://pypots.readthedocs.io/en/latest/release.html

Imports

pypots
```
import pypots
```
The main package; no common wrong pattern.

Quickstart

Impute missing values in 3D time series data using SAITS.

import pypots
from pypots.utils.random import set_random_seed
import numpy as np

# Generate random data with missing values
X = np.random.randn(100, 10, 5)
X[X < -1] = np.nan  # introduce missing values

# Impute using a simple model
from pypots.imputation import SAITS

model = SAITS(
    n_steps=10,
    n_features=5,
    n_layers=2,
    d_model=32,
    d_ffn=64,
    n_heads=4,
    d_k=16,
    d_v=16,
    epochs=10,
    batch_size=32,
    loss_fn='mse',
    optimizer='adam',
    lr=1e-3,
    verbose=True
)
model.fit(X)
imputed_X = model.impute(X)
print(imputed_X.shape)