sgn 0.5.0

A framework to help navigate buffers through a graph. The buffers must flow.

SGN Documentation

SGN is a lightweight Python library for creating and executing task graphs asynchronously for streaming data. With only builtin-dependencies, SGN is easy to install and use. This page is for the base library sgn, but there is a family of libraries that extend the functionality of SGN, including:

• sgn-ts: TimeSeries utilities for SGN
• sgn-ligo: LSC specific utilities for SGN

Installation

To install SGN, simply run:

pip install sgn

SGN has no dependencies outside of the Python standard library, so it should be easy to install on any system.

Quickstart

To get started with SGN, you can create a simple task graph that represents a simple data processing pipeline with integers. Here's an example:

import functools
from sgn import CallableTransform, CollectSink, IterSource, Pipeline


# Define a function to use in the pipeline
def scale(frame, factor: float):
    return None if frame.data is None else frame.data * factor


# Create source element
src = IterSource(
    name="src1",
    source_pad_names=["H1"],
    iters={"H1": [1, 2, 3]},
)

# Create a transform element using an arbitrary function
trn1 = CallableTransform.from_callable(
    name="t1",
    sink_pad_names=["H1"],
    callable=functools.partial(scale, factor=10),
    output_pad_name="H1",
)

# Create the sink so we can access the data after running
snk = CollectSink(
    name="snk1",
    sink_pad_names=("H1",),
)

# Create the Pipeline
p = Pipeline()

# Connect elements using pipeline.connect()
p.connect(src, trn1)  # Connects matching pad names automatically
p.connect(trn1, snk)  # H1 -> H1

# Run the pipeline
p.run()

# Check the result of the sink queue to see outputs
assert list(snk.collects["H1"]) == [10, 20, 30]

Advanced Grouping and Selection

SGN also supports grouping elements and selecting specific pads for more complex pipelines:

from sgn import Pipeline, IterSource, NullSink
from sgn.groups import group, select

# Create multiple sources
src1 = IterSource(name="src1", source_pad_names=["H1"])
src2 = IterSource(name="src2", source_pad_names=["L1", "V1"])

# Create sink
sink = NullSink(name="sink", sink_pad_names=["H1", "L1"])

# Group sources and select specific pads
sources = group(src1, select(src2, "L1"))  # Include all of src1, only L1 from src2

pipeline = Pipeline()
pipeline.connect(sources, sink)  # Automatic matching: H1->H1, L1->L1

The above example can be modified to use any data type, including json-friendly nested dictionaries, lists, and strings. The CallableTransform class can be used to create a transform element using any arbitrary function. The DequeSource and DequeSink classes are used to create source and sink elements that use collections.deque to store data.

General Concepts

Graph Construction

• Sources: Sources are the starting point of a task graph. They produce data that can be consumed by other tasks.

• Transforms: Transforms are tasks that consume data from one or more sources, process it, and produce new data.

• Sinks: Sinks are tasks that consume data from one or more sources and do something with it. This could be writing the data to a file, sending it over the network, or anything else.

Control Flow

Using these concepts, you can create complex task graphs using SGN that process and move data in a variety of ways. The SGN library provides a simple API for creating and executing task graphs, with a few key types:

• Frame: A frame is a unit of data that is passed between tasks in a task graph. Frames can contain any type of data, and can be passed between tasks in a task graph.

• Pad: A pad is a connection point between two tasks in a task graph. Pads are used to pass frames between tasks, and can be used to connect tasks in a task graph. An edge is a connection between two pads in a task graph.

• Element: An element is a task in a task graph. Elements can be sources, transforms, or sinks, and can be connected together to create a task graph.

• Pipeline: A pipeline is a collection of elements that are connected together to form a task graph. Pipelines can be executed to process data, and can be used to create complex data processing workflows.

Developer's Guide

SGN will execute a fixed graph of "pads", which are asynchronous function calls bound to classes called "elements".

Data must have an origin and a end point in all graphs. These are called sources and sinks. Elements that create data are called source elements and elements that collect data are called sink elements. Likewise, pads on elements are also called source and sink pads. Data passed between pads are stored in a Frame.

    /       ----------------------      <
   /       |   Source Element 1   |      \
  /        |                      |       \
 /          ---[source pad 'a']---         \
|                     |                     \
|                     | data flow            | The event loop runs this graph over and
\                     V                      | over pulling data through the pads
 \          ---[sink pad 'x'] ---           /
  \        |                     |         /
   \       |   Sink Element 1    |        /
    >      |                     |       /
           ---------------------       /

The whole graph execution is orchestrated by an event loop that will execute until end of stream. Here is a simple example implementing the above graph

from sgn.base import SourceElement, SinkElement, Frame
from sgn.apps import Pipeline

class MySourceClass(SourceElement):
    def new(self, pad):
        return Frame(data="hello")

class MySinkClass(SinkElement):
    def pull(self, pad, frame):
        print (frame.data)

source = MySourceClass(source_pad_names = ("a",))
sink = MySinkClass(sink_pad_names = ("x",))

pipeline = Pipeline()

pipeline.insert(source, sink, link_map = {sink.snks["x"]: source.srcs["a"]})

pipeline.run()

If you run this, it will run forever and you will see

hello
hello
hello
hello
hello
hello
hello
hello
hello
hello
...

You would need to send SIG INT or SIG kill to stop the program. Lets add a feature to end the stream after 10 Frames.

from sgn.base import SourceElement, SinkElement, Frame
from sgn.apps import Pipeline

class MySourceClass(SourceElement):
    def __post_init__(self):
        super().__post_init__()
        self.cnt = 0
    def new(self, pad):
        self.cnt += 1
        return Frame(data="hello", EOS=self.cnt > 10)

class MySinkClass(SinkElement):
    def pull(self, pad, frame):
        if frame.EOS:
            self.mark_eos(pad)
        print (frame.data)

source = MySourceClass(source_pad_names = ("a",))
sink = MySinkClass(sink_pad_names = ("x",))

pipeline = Pipeline()

pipeline.insert(source, sink, link_map = {sink.snks["x"]: source.srcs["a"]})

pipeline.run()

Now you would see the word "hello" printed 11 times. The 11th time the Frame is marked as EOS, which means end of stream. The sink class checks the data it has gotten and marks the pad as EOS. When all sink element sink pads are at EOS the pipeline stops running (in this case there is just one sink element with one sink pad).

What if we want more than one pad? It is possible to have many source and sink pads on an element. SGN provides basic bookkeeping utilities for you, but generally what the "correct" behavior is is up to you. Lets try a more complicated example with multiple pads:

 ---------------------------------------------
|                                             |
|              Source Element 1               |
|                                             |
 --- [source pad 'a'] --- [source pad 'b'] ---
           |                 |
           | data flow       |
           V                 V
 --- [sink pad 'x'  ] --- [sink pad 'y'  ] ---
|                                             |
|               Sink Element 1                |
|                                             |
----------------------------------------------

from dataclasses import dataclass
from sgn.base import SourceElement, SinkElement, Frame
from sgn.apps import Pipeline

@dataclass
class MySourceClass(SourceElement):
    # Of the form {"pad name": <data to put on the pad}
    pad_str_map: dict=None
    def __post_init__(self):
        # We will just use pad_str_map to define the source pad names too
        self.source_pad_names = tuple(self.pad_str_map)
        super().__post_init__()
        # save a pad map also hashed by pad not the string
        # NOTE: this must be done after super() post init so that the source pads exist
        self.pad_map = {self.srcs[p]: d for p,d in self.pad_str_map.items()}
        self.cnt = 0
    def new(self, pad):
        self.cnt += 1
        return Frame(data=self.pad_map[pad], EOS=self.cnt > 10)

class MySinkClass(SinkElement):
    def pull(self, pad, frame):
        if frame.EOS:
            self.mark_eos(pad)
        print (frame.data)

source = MySourceClass(pad_str_map = {"a": "Hello!", "b":"How are you?"})
sink = MySinkClass(sink_pad_names = ("x","y"))

pipeline = Pipeline()

pipeline.insert(source, sink, link_map = {sink.snks["x"]: source.srcs["a"], sink.snks["y"]: source.srcs["b"],})

pipeline.run()

Running this produces the following output:

e1-056827:~ crh184$ ./sgn-readme
Hello!
How are you?
Hello!
How are you?
Hello!
How are you?
Hello!
How are you?
Hello!
How are you?
Hello!
How are you?

Note that the total number of outputs is 12. We had the counter in the new() method which is a pad dependent method. It will be called once for each pad during each loop iteration. What if we wanted 10 loop iterations before sending EOS? There is a convenient "internal" pad inside of every element that is guaranteed to be called before any source pads and after any sink pads. Let's modify the code to use that:

from dataclasses import dataclass
from sgn.base import SourceElement, SinkElement, Frame
from sgn.apps import Pipeline

@dataclass
class MySourceClass(SourceElement):
    # Of the form {"pad name": <data to put on the pad}
    pad_str_map: dict=None
    def __post_init__(self):
        # We will just use pad_str_map to define the source pad names too
        self.source_pad_names = tuple(self.pad_str_map)
        super().__post_init__()
        # save a pad map also hashed by pad not the string
        # NOTE: this must be done after super() post init so that the source pads exist
        self.pad_map = {self.srcs[p]: d for p,d in self.pad_str_map.items()}
        self.cnt = 0
    def internal(self):
        self.cnt += 1
    def new(self, pad):
        return Frame(data=self.pad_map[pad], EOS=self.cnt > 10)

class MySinkClass(SinkElement):
    def pull(self, pad, frame):
        if frame.EOS:
            self.mark_eos(pad)
        print (frame.data)

source = MySourceClass(pad_str_map = {"a": "Hello!", "b":"How are you?"})
sink = MySinkClass(sink_pad_names = ("x","y"))

pipeline = Pipeline()

pipeline.insert(source, sink, link_map = {sink.snks["x"]: source.srcs["a"], sink.snks["y"]: source.srcs["b"],})

pipeline.run()

Now the output has the expected number of iterations

Hello!
How are you?
Hello!
How are you?
Hello!
How are you?
Hello!
How are you?
Hello!
How are you?
Hello!
How are you?
Hello!
How are you?
Hello!
How are you?
Hello!
How are you?
Hello!
How are you?
Hello!
How are you?

We can also use the internal method to make a more useful sink output, e.g.,

from dataclasses import dataclass
from sgn.base import SourceElement, SinkElement, Frame
from sgn.apps import Pipeline

@dataclass
class MySourceClass(SourceElement):
    # Of the form {"pad name": <data to put on the pad}
    pad_str_map: dict=None
    def __post_init__(self):
        # We will just use pad_str_map to define the source pad names too
        self.source_pad_names = tuple(self.pad_str_map)
        super().__post_init__()
        # save a pad map also hashed by pad not the string
        # NOTE: this must be done after super() post init so that the source pads exist
        self.pad_map = {self.srcs[p]: d for p,d in self.pad_str_map.items()}
        self.cnt = 0
    def internal(self):
        self.cnt += 1
    def new(self, pad):
        return Frame(data=self.pad_map[pad], EOS=self.cnt > 10)

class MySinkClass(SinkElement):
    def __post_init__(self):
        super().__post_init__()
        self.combined_string = ""
    def internal(self):
        print (self.combined_string)
        self.combined_string = ""
    def pull(self, pad, frame):
        if frame.EOS:
            self.mark_eos(pad)
        self.combined_string += " %s" % frame.data

source = MySourceClass(pad_str_map = {"a": "Hello!", "b":"How are you?"})
sink = MySinkClass(sink_pad_names = ("x","y"))

pipeline = Pipeline()

pipeline.insert(source, sink, link_map = {sink.snks["x"]: source.srcs["a"], sink.snks["y"]: source.srcs["b"],})

pipeline.run()

which now produces

 Hello! How are you?
 Hello! How are you?
 Hello! How are you?
 Hello! How are you?
 Hello! How are you?
 Hello! How are you?
 Hello! How are you?
 Hello! How are you?
 Hello! How are you?
 Hello! How are you?
 Hello! How are you?

Graphs can have other elements called "transform elements." These have both source and sink pads. Also, it is possible to connect a source pad to multiple sink pads (but not the other way around). Lets try to implement this graph

 ---------------------------------------------
|                                             |
|              Source Element 1               |
|                                             |
 --- [source pad 'a'] --- [source pad 'b'] ---
           |\                |\
           | \               | \
           |  \              |  \_________________________________________
           |   \_____________|_________________________                   \
           |                 |                         \                   \
           |                 |                          V                   V
           |                 |                 --- [sink pad 'l'  ] --- [sink pad 'm'  ] ---
           |                 |                |                                             |
           |                 |                |            Transform Element 1              |
           |                 |                |                                             |
           |                 |                 ------------- [source pad 'n'] --------------
           |                 |                                   /
           |                 |                                  /
           |                 |                                 /
           |                 |                                /
           |                 |                               /
           | data flow       |                              /
           V                 V                             V
 --- [sink pad 'x'  ] --- [sink pad 'y'  ] --- [sink pad 'z'  ] ---
|                                                                  |
|               Sink Element 1                                     |
|                                                                  |
-------------------------------------------------------------------

from dataclasses import dataclass
from sgn.base import SourceElement, SinkElement, TransformElement, Frame
from sgn.apps import Pipeline

@dataclass
class MySourceClass(SourceElement):
    # Of the form {"pad name": <data to put on the pad}
    pad_str_map: dict=None
    def __post_init__(self):
        # We will just use pad_str_map to define the source pad names too
        self.source_pad_names = tuple(self.pad_str_map)
        super().__post_init__()
        # save a pad map also hashed by pad not the string
        # NOTE: this must be done after super() post init so that the source pads exist
        self.pad_map = {self.srcs[p]: d for p,d in self.pad_str_map.items()}
        self.cnt = 0
    def internal(self):
        self.cnt += 1
    def new(self, pad):
        return Frame(data=self.pad_map[pad], EOS=self.cnt > 10)

class MyTransformClass(TransformElement):
    def __post_init__(self):
        # written to assume a single source pad
        assert len(self.source_pad_names) == 1
        super().__post_init__()
        self.out_string = ""
        self.out_frame = None
        self.EOS = False
    def pull(self, pad, frame):
        self.out_string += " %s" % frame.data
        self.EOS |= frame.EOS
    def internal(self):
        # Reverse the data for fun.
        self.outframe = Frame(data=self.out_string[::-1], EOS=self.EOS)
        self.out_string = ""
    def new(self, pad):
        # This element just has one source pad
        return self.outframe


class MySinkClass(SinkElement):
    def __post_init__(self):
        super().__post_init__()
        self.combined_string = ""
    def internal(self):
        print (self.combined_string)
        self.combined_string = ""
    def pull(self, pad, frame):
        if frame.EOS:
            self.mark_eos(pad)
        self.combined_string += " %s" % frame.data

source = MySourceClass(pad_str_map = {"a": "Hello!", "b":"How are you?"})
transform = MyTransformClass(sink_pad_names = ("l","m",), source_pad_names = ("n",))
sink = MySinkClass(sink_pad_names = ("x","y","z"))

pipeline = Pipeline()

pipeline.insert(source,
               transform,
               sink,
               link_map = {sink.snks["x"]: source.srcs["a"],
                           sink.snks["y"]: source.srcs["b"],
                           transform.snks["l"]: source.srcs["a"],
                           transform.snks["m"]: source.srcs["b"],
                           sink.snks["z"]: transform.srcs["n"]
                          }
               )

pipeline.run()

which produces

 Hello! How are you? ?uoy era woH !olleH
 Hello! How are you? ?uoy era woH !olleH
 Hello! How are you? ?uoy era woH !olleH
 Hello! How are you? ?uoy era woH !olleH
 Hello! How are you? ?uoy era woH !olleH
 Hello! How are you? ?uoy era woH !olleH
 Hello! How are you? ?uoy era woH !olleH
 Hello! How are you? ?uoy era woH !olleH
 Hello! How are you? ?uoy era woH !olleH
 Hello! How are you? ?uoy era woH !olleH
 Hello! How are you? ?uoy era woH !olleH

All you need to know about pads and names

Pads are hashable and they also have string names (though that name is not used as the hash). When developing you might get a bit turned around about how to access and reference pads by name. Here are a few rules:

• Elements have a notion of a short pad name. These are verbatim what get passed to source_pad_names and sink_pad_names.
• The Element base classes will initialize pads with long pad names of the form <element name>:["snk" | "source"]:<short name>.
• These long names are almost never needed for anything programmatically but they can be handy to print out because they carry extra information encoded in the name.
• Usually you will use helper attributes to reference pads by their short names or to look up a pad's short name.

Below is a bit of interactive python code that should be all you need to sort this out.

>>> from sgn.base import SourceElement
>>> e = SourceElement(name="example", source_pad_names=("alice","bob"))
>>> # Here are some relevant ways to access pad information
>>> # All of the "short" names -- these will be the strings provided by source_pad_names in the initialization
>>> print (e.source_pad_names)
('alice', 'bob')
>>> # A dictionary mapping the short name to a given pad object, e.g.,
>>> p = e.srcs["alice"]
>>> print (type(p))
<class 'sgn.base.SourcePad'>
>>> # The pad's long name
>>> print (p.name)
example:src:alice
>>> # A reverse dictionary mapping a pad to a short name
>>> print (e.rsrcs[p])
alice

Some useful API docs from this guide

Below are some API docs for concepts that came up in this guide

• SourceElement
• TransformElement
• SinkElement
• Frame
• Pipeline