Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?


Failed to load latest commit information.
Latest commit message
Commit time
March 23, 2023 15:50
March 23, 2023 14:30
September 6, 2022 19:58
October 28, 2022 13:13
September 9, 2022 09:00
March 17, 2022 08:45
September 6, 2022 18:01
October 6, 2022 14:23
January 24, 2023 15:35
January 24, 2023 15:35;base64,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


A library of immutable and grow-only Pandas-like DataFrames with a more explicit and consistent interface. StaticFrame is suitable for applications in data science, data engineering, finance, scientific computing, and related fields where reducing opportunities for error by prohibiting in-place mutation is critical.

While many interfaces are similar to Pandas, StaticFrame deviates from Pandas in many ways: all data is immutable, and all indices are unique; the full range of NumPy data types is preserved, and date-time indices use discrete NumPy units; hierarchical indices are seamlessly integrated; and uniform approaches to element, row, and column iteration and function application are provided. Core StaticFrame depends only on NumPy and two C-extension packages (maintained by the StaticFrame team): Pandas is not a dependency.

A wide variety of table formats are supported, including input from and output to CSV, TSV, JSON, MessagePack, Excel XLSX, SQLite, HDF5, NumPy, Pandas, Arrow, and Parquet; additionally, output to xarray, VisiData, HTML, RST, Markdown, and LaTeX is supported, as well as HTML representations in Jupyter notebooks. Full serialization is also available via custom NPZ and NPY encodings, the latter supporting memory mapping.

StaticFrame features a family of multi-table containers: the Bus is a lazily-loaded container of tables, the Batch is a deferred processor of tables, the Yarn is virtual concatenation of many Buses, and the Quilt is a virtual concatenation of all tables within a single Bus or Yarn. All permit operating on large collections of tables with minimal memory overhead, as well as writing to and reading from zipped bundles of pickles, NPZ, Parquet, or delimited files, as well as XLSX workbooks, SQLite, and HDF5.




API Search:

Jupyter Notebook Tutorial: Launch Binder


Install StaticFrame via PIP:

pip install static-frame

Or, install StaticFrame via conda:

conda install -c conda-forge static-frame

To install full support of input and output routines via PIP:

pip install static-frame [extras]


Core StaticFrame requires the following:

  • Python>=3.7
  • NumPy>=1.18.5
  • automap==0.6.2
  • arraykit==0.3.0

For extended input and output, the following packages are required:

  • pandas>=0.24.2
  • xlsxwriter>=1.1.2
  • openpyxl>=3.0.9
  • xarray>=0.13.0
  • tables>=3.6.1
  • pyarrow>=0.17.0
  • visidata>=2.4

Quick-Start Guide

To get startred quickly, let's download the classic iris (flower) characteristics data set and build a simple naive Bayes classifier that can predict species from iris petal characteristics.

While StaticFrame's API has over 7,500 endpoints, much will be familiar to users of Pandas or other DataFrame libraries. Rather than offering fewer interfaces with greater configurability, StaticFrame favors more numerous interfaces with more narrow parameters and functionality. This design leads to more maintainable code. (Read more about differences between Pandas and StaticFrame here.)

We can download the data set from the UCI Machine Learning Repository and create a Frame. StaticFrame exposes all constructors on the class: here, we will use the Frame.from_csv() constructor. To download a file from the internet and provide it to a constructor, we can use StaticFrame's WWW.from_file() interface.

>>> import static_frame as sf
>>> data = sf.Frame.from_csv(sf.WWW.from_file(''), columns_depth=0)

Each record (or row) in this dataset describes observations of an iris flower, including its sepal and petal characteristics, as well as its species (of which there are three). To display just the first few rows, we can use the head() method. Notice that StaticFrame's default display makes it very clear what type of Frame, Index, and NumPy datatypes are present.

>>> data.head()
<Index> 0         1         2         3         4           <int64>
0       5.1       3.5       1.4       0.2       Iris-setosa
1       4.9       3.0       1.4       0.2       Iris-setosa
2       4.7       3.2       1.3       0.2       Iris-setosa
3       4.6       3.1       1.5       0.2       Iris-setosa
4       5.0       3.6       1.4       0.2       Iris-setosa
<int64> <float64> <float64> <float64> <float64> <<U15>

As the columns are unlabelled, let's next add column labels. StaticFrame supports reindexing (conforming existing axis labels to new labels, potentially changing the size and ordering) and relabeling (simply applying new labels without regard to existing labels). As we can ignore the default column labels (auto-incremented integers), the relabel() method is used to provide new labels.

Note that while relabel() creates a new Frame, underlying NumPy data is not copied. As all NumPy data is immutable in StaticFrame, we can reuse it in our new container, making such operations very efficient. (Read more about no-copy operations here.)

>>> data = data.relabel(columns=('sepal_l', 'sepal_w', 'petal_l', 'petal_w', 'species'))
>>> data.head()
<Index> sepal_l   sepal_w   petal_l   petal_w   species     <<U7>
0       5.1       3.5       1.4       0.2       Iris-setosa
1       4.9       3.0       1.4       0.2       Iris-setosa
2       4.7       3.2       1.3       0.2       Iris-setosa
3       4.6       3.1       1.5       0.2       Iris-setosa
4       5.0       3.6       1.4       0.2       Iris-setosa
<int64> <float64> <float64> <float64> <float64> <<U15>

For this example, eighty percent of the data will be used to train the classifier; the remaining twenty percent will be used to test the classifier. As all records are labelled with the known species, we can conclude by measuring the effectiveness of the classifier on the test data.

To divide the data into two groups, we create a Series of contiguous integers and then extract a random selection of 80% of the values into a new Series, here named sel_train. This will be used to select our traning data. As the sample() method, given a count, randomly samples that many values, your results will be different unless use the same seed argument.

>>> sel = sf.Series(np.arange(len(data)))
>>> sel_train = sel.sample(round(len(data) * .8), seed=42)
>>> sel_train.head()
0        0
2        2
3        3
4        4
5        5
<int64>  <int64>

We will create another Series to select the test data. The drop[] interface can be used to create a new Series that excludes the training selections, leaving just the testing selections. As with many interfaces in StaticFrame (such as astype and assign), brackets can be used to do loc[] style selections.

>>> sel_test = sel.drop[sel_train]
>>> sel_test.head()
1        1
14       14
20       20
21       21
37       37
<int64>  <int64>

To select a subset of the data for training, the sel_train Series can be passed to loc[] to select just those rows.

>>> data_train = data.loc[sel_train]
>>> data_train.head()
<Index> sepal_l   sepal_w   petal_l   petal_w   species     <<U7>
0       5.1       3.5       1.4       0.2       Iris-setosa
2       4.7       3.2       1.3       0.2       Iris-setosa
3       4.6       3.1       1.5       0.2       Iris-setosa
4       5.0       3.6       1.4       0.2       Iris-setosa
5       5.4       3.9       1.7       0.4       Iris-setosa
<int64> <float64> <float64> <float64> <float64> <<U15>

With our data divided into two randomly-selected, non-overlapping groups, we can proceed to implement the naive Bayes classifier. We will compute the posterior of the test data by multiplying the prior and the likelihood. With the posterior, we can determine which species the classifier has calculated is most likely. (More on naive Bayes classifiers can be found here.)

The prior is calculated as the percentage of samples of each species in the training data. This is the "normalized" count per species. To get a Series of counts per species, we can select the species column, iterate over groups based on species name, and count the size of each group.

In StaticFrame, this can be done by calling Series.iter_group_items() to get an iterator of pairs of group label, group (where the group is a Series). This iterator (or any similar iterator) can be given to a Batch, a chaining processor of Frame or Series, to perform operations on each group. (For more on the Batch and other higher-order containers in StaticFrame, see here.)

Once the Batch is created, selections, method calls, and operator expressions can be chained as if they were being called on a single container. Processing happens to every contained container, and a container is returned, only when a finalizer method, such as to_series(), is called.

>>> counts = sf.Batch(data_train['species'].iter_group_items()).count().to_series()
>>> counts
Iris-setosa     43
Iris-versicolor 39
Iris-virginica  38
<<U15>          <int64>

As with NumPy, StaticFrame containers can be used in expressions with binary operators. The prior can be derived by dividing counts by the size of the training data. This returns a Series of the percentage of records per species.

>>> prior = counts / len(data_train)
>>> prior
Iris-setosa     0.35833333333333334
Iris-versicolor 0.325
Iris-virginica  0.31666666666666665
<<U15>          <float64>

Having calculated the prior, we can calculate likelihood next. To calculate likelihood, we will call a probability distribution function (imported from SciPy) with the test data, once for each species, given the characteristics (mean and standard deviation) observed in the test data for that species.

The Batch can again be used to calculate the mean and standard deviation, per species, from the training data. With the Frame of training data, we call iter_group_items() to group by species and, passing that iterator to Batch, call mean() (assigned to mu) or std() (assigned to sigma). Note that iter_group_items() has an optional drop parameter to remove the column used for grouping from subsequent operations.

>>> mu = sf.Batch(data_train[['sepal_l', 'sepal_w', 'species']].iter_group_items('species', drop=True)).mean().to_frame()
>>> mu
<Index>         sepal_l            sepal_w            <<U7>
Iris-setosa     4.986046511627907  3.434883720930233
Iris-versicolor 5.920512820512819  2.771794871794872
Iris-virginica  6.6078947368421055 2.9763157894736842
<<U15>          <float64>          <float64>
>>> sigma = sf.Batch(data_train[['sepal_l', 'sepal_w', 'species']].iter_group_items('species', drop=True)).std(ddof=1).to_frame()
>>> sigma
<Index>         sepal_l            sepal_w             <<U7>
Iris-setosa     0.3419700595003668 0.3477024733400345
Iris-versicolor 0.508444214804487  0.33082728674826684
Iris-virginica  0.6055516042229233 0.3513942965328924
<<U15>          <float64>          <float64>

For a unified display of these characteristics, we can build a hierarchical index on each Frame with relabel_level_add() (adding the "mu" or "sigma" labels), then vertically concatenate the tables. As StaticFrame always requires unique labels in indices, adding an additional label is required before concatenation. The built-in round function can be used for more tidy display.

>>> stats = sf.Frame.from_concat((mu.relabel_level_add('mu'), sigma.relabel_level_add('sigma')))
>>> round(stats, 2)
<Index>                          sepal_l   sepal_w   <<U7>
mu               Iris-setosa     4.99      3.43
mu               Iris-versicolor 5.92      2.77
mu               Iris-virginica  6.61      2.98
sigma            Iris-setosa     0.34      0.35
sigma            Iris-versicolor 0.51      0.33
sigma            Iris-virginica  0.61      0.35
<<U5>            <<U15>          <float64> <float64>

We can now move on to processing the test data with the characteristics derived from the training data. To do that, we will extract our previously selected test records with sel_test into a new Frame, to which we can add our posterior predictions and final species classifications.

It is common to process data in table by adding columns from left to right. StaticFrame permits this limited form of mutability with the grow-only FrameGO. While underlying NumPy arrays are still always immutable, columns can be added to a FrameGO with bracket-style assignments. A FrameGO can be created from a Frame with the to_frame_go() method. As mentioned elsewhere, underlying immutable NumPy arrays are not copied: this is an efficient, no-copy operation.

Passing two arguments to loc[], we can select rows with the values from sel_test, and we can select columns with a list of labels for the sepal length and sepal width.

>>> data_test = data.loc[sel_test.values, ['sepal_l', 'sepal_w']].to_frame_go()
>>> data_test.head()
<IndexGO> sepal_l   sepal_w   <<U7>
1         4.9       3.0
14        5.8       4.0
20        5.4       3.4
21        5.1       3.7
37        4.9       3.1
<int64>   <float64> <float64>

StaticFrame interfaces make extensive use of iterators and generators. As used below, the Frame.from_fields() constructor will create a Frame from any iterable (or generator) of column arrays.

The likelihood_of_species() function (defined below), for each index label in mu (which provides each unique iris species), calculates a probability density function for the test data, given the mu (mean) and sigma (standard deviation) for the species. An array of the sum of the log is yielded.

>>> from scipy.stats import norm
>>> def likelihood_of_species():
...     for label in mu.index:
...             pdf = norm.pdf(data_test.values, mu.loc[label], sigma.loc[label])
...             yield np.log(pdf).sum(axis=1)

While the generator function above is easy to read, it is hard to copy and paste. If you are following along, using the one-line generator expression, below, will be easier. The two are equivalent:

>>> likelihood_of_species = (np.log(norm.pdf(data_test.values, mu.loc[label], sigma.loc[label])).sum(axis=1) for label in mu.index)

With this generator expression defined, we call the from_fields constructor to produce the likelihood table, providing column labels from mu.index and index labels from data_test.index. For each test record row we now have a likelihood per species.

>>> likelihood = sf.Frame.from_fields(likelihood_of_species, columns=mu.index, index=data_test.index)
>>> round(likelihood.head(), 2)
<Index> Iris-setosa Iris-versicolor Iris-virginica <<U15>
1       -0.52       -2.31           -4.27
14      -3.86       -6.97           -5.42
20      -0.45       -2.38           -3.01
21      -0.05       -5.29           -5.51
37      -0.2        -2.56           -4.33
<int64> <float64>   <float64>       <float64>

We can calculate the posterior by multiplying likelihood by prior. Whenever performing binary operations on Frame and Series, indices will be aligned and, if necessary, reindexed before processing.

>>> posterior = likelihood * prior
>>> round(posterior.head(), 2)
<Index> Iris-setosa Iris-versicolor Iris-virginica <<U15>
1       -0.19       -0.75           -1.35
14      -1.38       -2.27           -1.72
20      -0.16       -0.77           -0.95
21      -0.02       -1.72           -1.75
37      -0.07       -0.83           -1.37
<int64> <float64>   <float64>       <float64>

We can now add columns to our data_test FrameGO. To determine our best prediction of species for each row of the test data, the column label (the species) of the maximum a posteriori estimate is selected with loc_max().

>>> data_test['predict'] = posterior.loc_max(axis=1)
>>> data_test.head()
<IndexGO> sepal_l   sepal_w   predict     <<U7>
1         4.9       3.0       Iris-setosa
14        5.8       4.0       Iris-setosa
20        5.4       3.4       Iris-setosa
21        5.1       3.7       Iris-setosa
37        4.9       3.1       Iris-setosa
<int64>   <float64> <float64> <<U15>

We can add two additional columns to evaluate the effectivess of the classifier. First, we can add an "observed" column by adding the original "species" column from the original data Frame. In assigning a Series to a Frame, only values found in the intersection of the indices will be added as a column.

>>> data_test['observed'] = data['species']
>>> data_test.head()
<IndexGO> sepal_l   sepal_w   predict     observed    <<U8>
1         4.9       3.0       Iris-setosa Iris-setosa
14        5.8       4.0       Iris-setosa Iris-setosa
20        5.4       3.4       Iris-setosa Iris-setosa
21        5.1       3.7       Iris-setosa Iris-setosa
37        4.9       3.1       Iris-setosa Iris-setosa
<int64>   <float64> <float64> <<U15>      <<U15>

Having populated a column of predicted and observed values, we can compare the two to get a Boolean column indicating when the classifier calculated a correct predicton.

>>> data_test['correct'] = data_test['predict'] == data_test['observed']
>>> data_test.tail()
<IndexGO> sepal_l   sepal_w   predict         observed       correct <<U8>
129       7.2       3.0       Iris-virginica  Iris-virginica True
130       7.4       2.8       Iris-virginica  Iris-virginica True
140       6.7       3.1       Iris-virginica  Iris-virginica True
144       6.7       3.3       Iris-virginica  Iris-virginica True
149       5.9       3.0       Iris-versicolor Iris-virginica False
<int64>   <float64> <float64> <<U15>          <<U15>         <bool>

To find the percentage of correct classifications among the test data, we can sum the correct Boolean column and divide that by the size of the test data.

>>> data_test["correct"].sum() / len(data_test)

This simple naive Bayes classifier can predict iris species correctly about 73% of the time.

For further introduction to StaticFrame, including links to articles, videos, and documentation, see here.