API Cheat Sheet

Import conventions:

import pixeltable as pxt
import pixeltable.functions as pxtf

Operations summary

Task	Code
Create a (mutable) table	t = pxt.create_table('table_name', {'col_1': pxt.String, 'col_2': pxt.Int, ...})
Create a view	t = pxt.create_view('view_name', base_tbl.where(base_tbl.col > 10))
Create a view with iterator	t = pxt.create_view('view_name', base_tbl, iterator=FrameIterator.create(video=base_tbl.col, fps=0))
Create a snapshot	t = pxt.create_snapshot('snapshot_name', base_tbl)

The following functions apply to tables, views, and snapshots.

Task	Code
Use an existing table	t = pxt.get_table('video_data')
Rename a table	pxt.move('video_data', 'vd')
Move a table	pxt.move('video_data', 'experiments.video_data')
List tables	pxt.list_tables()
Delete a table	pxt.drop_table('video_data')
Delete a table and all its views	pxt.drop_table('video_data', force=True)

Directories

Task	Code
Create a directory	pxt.create_dir('experiments')
Rename or move a directory	pxt.move('experiments', 'project_x.experiments')
Delete a directory	pxt.drop_dir('experiments')
Delete a directory and all its contents	pxt.drop_dir('experiments', force=True)
List directories	pxt.list_dirs('project_x')

Frame extraction for video data

Create a table with video data and view for the frames:

import pixeltable as pxt
from pixeltable.iterators import FrameIterator
t = pxt.create_table('tbl_name', {'video': pxt.Video})
f = pxt.create_view('frame_view_name', t, iterator=FrameIterator.create(video=t.video, fps=0))

fps=0 extracts frames at the original frame rate.

Pixeltable types

Pixeltable type	Corresponding Python type
pxt.String	str
pxt.Int	int
pxt.Float	float
pxt.Bool	bool
pxt.Timestamp	datetime.datetime
pxt.Json	list or dict
pxt.Array	numpy.ndarray
pxt.Image	PIL.Image.Image
pxt.Video	str (the file path or URL)
pxt.Audio	str (the file path or URL)
pxt.Document	str (the file path or URL)

Table operations summary

Action	Code
Print table schema	t.describe()
Query a table	t.select(t.col2, t.col3 + 5).show()
Query a table with a filter	t.where(t.col1 == 'green').select(t.col2).show()
Insert a single row into a table	t.insert(col1='green', ...)
Insert multiple rows into a table	t.insert([{'col1': 'green', ...}, {'col1': 'red', ...}, ...])
Add a column	t.add_column(new_col_name=pxt.Int)
Add a column (alternate form)	t[new_col_name] = pxt.Int
Rename a column	t.rename_column('col_name', 'new_col_name')
Drop a column	t.drop_column('col_name')
Undo the last update operation (add/rename/drop column or insert)	t.revert()

Querying a table

Action	Code
Look at 10 rows	t.collect(10)
Look at the oldest 10 rows	t.head(10)
Look at the most recently added 10 rows	t.tail(10)
Look at all rows	t.collect()
Iterate over all rows as dictionaries	for row in t.collect(): ...
Look at row for frame 15	t.where(t.pos == 15).show()
Look at rows before index 15	t.where(t.pos < 15).show()
Look at rows before index 15 with RGB frames	t.where((t.pos < 15) & (t.frame.mode == 'RGB')).collect()

Pixeltable supports the standard comparison operators (>=, >, ==, <=, <). == None is the equivalent of isna()/isnull() in Pandas.

Boolean operators are the same as in Pandas: & for and, | for or, ~ for not. They also require parentheses, for example: (t.pos < 15) & (t.frame.mode == 'RGB') or ~(t.frame.mode == 'RGB').

Selecting and transforming columns

Action	Code
Only retrieve the frame index and frame	t.select(t.frame_idx, t.frame).collect()
Look at frames rotated 90 degrees	t.select(t.frame.rotate(90)).collect()
Overlay frame with itself rotated 90 degrees	t.select(pxt.functions.pil.image.blend(t.frame, t.frame.rotate(90))).collect()

Computed columns

The values in a computed column are automatically filled when data is added:

t.add_column(c_added=t.frame.rotate(30))

Alternatively:

t['c_added'] = t.frame.rotate(30)

Computed columns and media columns (video, image, audio) have attributes errortype and errormsg, which contain the exception type and string in rows where the computation expression or media type validation results in an exception (the column value itself will be None).

Example:

t.where(t.c_added.errortype != None).select(t.c_added.errortype, t.c_added.errormsg).show()

returns the exception type and message for rows with an exception.

Inserting data into a table

t.insert([{'video': '/path/to/video1.mp4'}, {'video': '/path/to/video2.mp4'}])

Each row is a dictionary mapping column names to column values (do not provide values for computed columns).

Attributes and methods on image data

Images are represented as PIL.Image.Image instances in memory and support a lot of the attributes and methods documented here.

Available attributes are: mode, height, width.

Available methods are: convert, crop, effect_spread, entropy, filter, getbands, getbbox, getchannel, getcolors, getextrema, getpalette, getpixel, getprojection, histogram, point, quantize, reduce, remap_palette, resize, rotate, transform, transpose.

Methods can be chained, for example: t.frame.resize((224, 224)).rotate(90).convert('L')