Skip to content

Latest commit

 

History

History
475 lines (362 loc) · 17.2 KB

README.md

File metadata and controls

475 lines (362 loc) · 17.2 KB

Synthetic Population Pipeline (synpp)

Build Status

The synpp module is a tool to chain different stages of a (population) synthesis pipeline. This means that self-contained pieces of code can be run, which are dependent on the outputs of other self-contained pieces of code. Those pieces, or steps, are called stages in this module.

The following will describe the components of the pipeline and how it can be set up and configured. Scroll to the bottom to find a full example of such a pipeline which automatically downloads NYC taxi data sets, merges them together and calculates the average vehicle occupancy during a predefined period.

Installation

The synpp package releases can be installed via pip:

pip install synpp

Currently, version 1.5.1 is the active release version. Alternatively, you can clone the develop branch of this repository to use the development version. It can be installed by calling

pip install .

inside of the repository directoy.

Concepts

A typical chain of stages could, for instance, be: (C1) load raw census data, (C2) clean raw census data (dependent on C1), (H1) load raw household travel survey data, (H2) clean survey data (dependent on C2), (P1) merge census (C1) and survey (H2) data, (P2) generate a synthetic population from merged data (P1).

In synpp each stage is defined by:

  • A descriptor, which contains the stage logic.
  • Configuration options that parameterize each stage.

Defining a descriptor

A descriptor can be defined in its compact form or in its full form. Both work in the same way and can be used interchangeably in most cases.

In this readme the full form is preferred to explain each of synpp's feature as it is more expressive, but towards the end a closer look at the compact form is also provided.

A descriptor in its full form looks like:

def configure(context):
  pass

def execute(context):
  pass

def validate(context):
  pass

These functions are provided in a Python object, such as a module, a class or a class's instance. synpp expects either a String containing the path to the object, such as "pkg.subpkg.module", or the instantiated object directly.

In its compact form, the stage is defined as a function, and looks like:

@synpp.stage
def stage_to_run():
  pass

Where the @stage decorator informs synpp that it should handle this function as a stage and how it should do it.

Configuration and Parameterization

Whenever the pipeline explores a stage, configure is called first. Note that in the example above we use a Python module, but the same procedure would work analogously with a class. In configure one can tell the pipeline what the stage expects in terms of other input stages and in terms of configuration options:

def configure(context):
  # Expect an output directory
  value = context.config("output_path")

  # Expect a random seed
  value = context.config("random_seed")

  # Expect a certain stage (no return value)
  context.stage("my.pipeline.raw_data")

We could add this stage (let's call it my.pipeline.raw_data) as a dependency to another one. However, as we did not define a default value with the config method, we need to explicitly set one, like so:

def configure(context):
  context.stage("my.pipeline.raw_data", { "random_seed": 1234 })

Note that it is even possible to build recursive chains of stages using only one stage definition:

def configure(context):
  i = context.config("i")

  if i > 0:
    context.stage("this.stage", { "i": i - 1 })

Configuration options can also be defined globally in the pipeline. In case no default value is given for an option in configure and in case that no specific value is passed to the stage, a global configuration that is specific to the pipeline will be used to look up the value.

Execution

The requested configuration values and stages are afterwards available to the execute step of a stage. There those values can be used to do the "heavy work" of the stage. As the configure step already defined what kind of values to expect, we can be sure that those values and dependencies are present once execute is called.

def execute(context):
  # Load some data from another stage
  df = context.stage("my.pipeline.census.raw")

  df = df.dropna()
  df["age"] = df["age"].astype(int)

  # We could access some values if we wanted
  value = context.config("...")

  return df

Note that the execute step returns a value. This value will be pickled (see pickle package of Python) and cached on the hard drive. This means that whenever the output of this stage is requested by another stage, it doesn't need to be run again. The pipeline can simply load the cached result from hard drive.

If one has a very complex pipeline with many stages this means that changes in one stage will likely not lead to a situation where one needs to re-run the whole pipeline, but only a fraction. The synpp framework has intelligent explorations algorithms included which figure out automatically, which stages need to be re-run.

Running a pipeline

A pipeline can be started using the synpp.run method. A typical run would look like this:

config = { "random_seed": 1234 }
working_directory = "~/pipeline/cache"

synpp.run([
    { "descriptor": "my.pipeline.final_population" },
    { "descriptor": "my.pipeline.paper_analysis", "config": { "font_size": 12 } }
], config = config, working_directory = working_directory)

Here we call the stage defined by the module my.pipeline.final_population which should be available in the Python path. And we also want to run the my.pipeline.paper_analysis path with a font size parameter of 12. Note that in both cases we could also have based the bare Python module objects instead of strings.

The pipeline will now figure out how to run those stages. Probably they have dependencies and the analysis stage may even depend on the other one. Therefore, synpp explores the tree of dependencies as follows:

  • Consider the requested stages (two in this case)
  • Step by step, go through the dependencies of those stages
  • Then again, go through the dependencies of all added stages, and so on

By that the pipeline traverses the whole tree of dependencies as they are defined by the configure steps of all stages. At the same time it collects information about which configuration options and parameters are required by each stage. Note that a stage can occur twice in this dependency tree if it has different parameters.

After constructing a tree of stages, synpp devalidates some of them according to the following scheme. A stage is devalidated if ...

  • ... it is requested by the run call (and rerun_required is set to True, the default)
  • ... it is new (no meta data from a previous call is present)
  • ... the code of the stage has changed (verified with inspection)
  • ... if at least one of the requested configuration options has changed
  • ... if at least one dependency has been re-run since the last run of the stage
  • ... if list of dependencies has changed
  • ... if manual validation of the stage has failed (see below)
  • ... if any ascendant of a stage has been devalidated

This list of conditions makes sure that in almost any case of pipeline modification we end up in a consistent situation (though we cannot prove it). The only measure that may be important to enforce 'by convention' is to always run a stage after the code has been modified. Though even this can be automated.

Validation

Each stage has an additional validate step, which also receives the configuration options and the parameters. Its purpose is to return a hash value that represents the environment of the stage. To learn about the concept in general, search for "md5 hash", for instance. The idea is the following: After the execute step, the validate step is called and it will return a certain value. Next time the pipeline is resolved the validate step is called during devalidation, i.e. before the stage is actually executed. If the return value of validate now differs from what it was before, the stage will be devalidated.

This is useful to check the integrity of data that is not generated inside of the pipeline but comes from the outside, for instance:

def configure(context):
  context.config("input_path")

def validate(context):
  path = context.config("input_path")
  filesize = get_filesize(path)

  # If the file size has changed, the file must have changed,
  # hence we want to run the stage again.
  return filesize

def execute(context):
  pass # Do something with the file

Cache paths

Sometimes, results of a stage are not easily representable in Python. Even more, stages may call Java or Shell scripts which simply generate an output file. For these cases each stage has its own cache path. It can be accessed through the stage context:

def execute(context):
  # In this case we write a file to the cache path of the current stage
  with open("%s/myfile.txt" % context.path()) as f:
    f.write("my content")

  # In this case we read a file from the cache path of another stage
  with open("%s/otherfile.txt" % context.path("my.other.stage")) as f:
    value = f.read()

As the example shows, we can also access cache paths of other stages. The pipeline will make sure that you only have access to the cache path of stages that have been defined as dependencies before. Note that the pipeline cannot enforce that one stage is not corrupting the cache path of another stage. Therefore, by convention, a stage should never write to the cache path of another stage.

Aliases

Once a pipeline has been defined, the structure is relatively rigid as stages are referenced by their names. To provide more flexibility, it is possible to define aliases, for instance:

synpp.run(..., aliases = {
  "my.pipeline.final_population": "my.pipeline.final_population_replacement"
})

Whenever my.pipeline.final_population is requested, my.pipeline.final_population_replacement will be used instead. Note that this allows to define entirely virtual stages that are referenced from other stages and which are only bound to a specific execution stage when running the pipeline (see example above).

Parallel execution

The synpp package comes with some simplified ways of parallelizing code, which are built on top of the multiprocessing package. To set up a parallel routine, one can follow the following pattern:

def run_parallel(context, x):
  return x**2 + context.data("y")

def execute(context):
  data = { "y": 5 }

  with context.parallel(data) as parallel:
    result = parallel.map(run_parallel, [1, 2, 3, 4, 5])

This approach looks similar to the Pool object of multiprocessing but has some simplifications. First, the first argument of the parallel routine is a context object, which provides configuration and parameters. Furthermore, it provides data, which has been passed before in the execute function. This simplifies passing data to all parallel threads considerably to the more flexible approach in multiprocessing. Otherwise, the parallel object provides most of the functionality of Pool, like, map, async_map, imap, and unordered_imap.

Info

While running the pipeline a lot of additional information may be interesting, like how many samples of a data set have been discarded in a certain stage. However, they often would only be used at the very end of the pipeline when maybe a paper, a report or some explanatory graphics are generated. For that, the pipeline provides the set_info method:

def execute(context):
  # ...
  context.set_info("dropped_samples", number_of_dropped_samples)
  # ...

The information can later be retrieved from another stage (which has the stage in question as a dependency):

def execute(context):
  # ...
  value = context.get_info("my.other.stage", "dropped_samples")
  # ...

Note that the info functionality should only be used for light-weight information like integers, short strings, etc.

Progress

The synpp package provides functionality to show the progress of a stage similar to tqdm. However, tqdm tends to spam the console output which is especially undesired if pipelines have long runtimes and run, for instance, in Continuous Integration environments. Therefore, synpp provides its own functionality, although tqdm could still be used:

def execute(context):
  # As a
  with context.progress(label = "My progress...", total = 100) as progress:
    i = 0

    while i < 100:
      progress.update()
      i += 1

  for i in context.progress(range(100)):
    pass

Compact stage definition

As quickly introduced before, stages can also be defined in a compact form. The example offered is the simplest possible, where a stage takes no configuration parameters. Consider now the more elaborate setting:

@synpp.stage(loaded_census="my.pipeline.census.raw", sample_size="census_sample_size")
def clean_census(loaded_census, sample_size=0.1):
    ...

When synpp sees clean_census, it will under the hood convert it to a stage in its full form. Basically @synpp.stage says how the stage should be configured and the function defines the stage's logic. To put clearly, the stage above is converted by synpp to something like:

def configure(context):
  context.stage("my.pipeline.census.raw")
  context.config("census_sample_size", default=0.1)

def execute(context):
  loaded_census = context.stage("my.pipeline.census.raw")
  sample_size = context.config("census_sample_size")
  return clean_census(loaded_census, sample_size)

As you may have noticed, census_sample_size is a config option defined in the config file, and in case it isn't found, synpp will simply use the function's default. Notice also that the following wouldn't work: @synpp.stage(..., sample_size=0.2), since synpp would try to find a parameter called "0.2" in the config file that doesn't exist.

In case a parameterized stage must be passed as dependency, this can be performed in a similar way, by simply wrapping the stage in the synpp.stage() decorator. Following the previous example, we may replace the first argument with loaded_census=synpp.stage("my.pipeline.census.raw", file_path="path/to/alternative/census").

This compact way of defining stages does not support all functionality, for instance custom stage devalidation, but functionality that requires the context object are also possible via the helper method synpp.get_context().

Command-line tool

The synpp pipeline comes with a command line tool, which can be called like

python3 -m synpp [config_path]

If the config path is not given, it will assume config.yml. This file should contain everything to run a pipeline. A simple version would look like this:

# General pipeline settings
working_directory: /path/to/my/working_directory

# Requested stages
run:
  - my_first_module.my_first_stage
  - my_first_parameterized_stage:
    param1: 123
    param2: 345

# These are configuration options that are used in the pipeline
config:
  my_option: 123

It receives the working directory, a list of stages (which may be parameterized) and all configuration options. The stages listed above should be available as Python modules or classes. Furthermore, aliases can be defined as a top-level element of the file.

NYC Taxi Example

This repository contains an example of the pipline. To run it, you will need pandas as an additional Python dependency. For testing, you can clone this repository to any directory on your machine. Inside the repository directory you can find the example directory. If you did not install synpp yet, you can do this by executing

pip install .

inside of the repository directory. Afterwards, open examples/config.yml and adjust the working_directory path. This is a path that should exist on your machine and it should be empty. The best is if you simply create a new folder and add the path in config.yml.

You can now go to examples and call the pipeline code:

cd examples
python3 -m synpp

It will automatically discover config.yml (but you could path a different config file path manually as a command line argument). It will then download the NYC taxi data for January, February and March 2018 (see configuration options in config.yml). Note that this is happening in one stage for which you can find the code in nyc_taxi.download. It is parameterized by a month and a year to download the respective data set. These data sets then go into nyc_taxi.aggregate, where they are merged together. Finally, an average occupancy value is printed out in nyc_taxi.print_occupancy. So the dependency structure is as follows:

nyc_taxi.aggregate depends on multiple nyc_taxi.download(year, month)
nyc_taxi.print_occupancy depends on nyc_taxi.aggregate

After one successful run of the pipeline you can start it again. You will notice that the pipeline does not download the data again, because nothing has changed for those stages. However, if you would change the requested months in config.yml the pipeline may download the additional data sets.