Basins

Motivation

Basins are a powerful concept that allow you to both save time, disk space, and network usage when processing DC data. The basic idea behind basins is that you avoid duplicating feature data, by not copying all of the features from an input file to an output file, but by linking from the output file to the input file.

For instance, let’s say your pipeline is designed to compute a new feature userdef1 and you would like to open the output file in Shape-Out, visualizing this feature in combination with other features defined in the input file (e.g. deform). What you could do is to write the userdef1 feature directly to the input file or to create a copy of your input file and write userdef1 to that copy. However, this might make you feel uneasy, because you would like to avoid editing your original raw data (possible data loss) and copying an entire file seems like unnecessary overhead (redundant data, high disk usage). With basin features, you can create an empty output file, write your new feature userdef1 to it, define a basin that links to your input file, and you are done. Without ever modifying your input file. And it is even possible to create an output file that is gated (containing a subset of events from the input file).

Defining Basins

Basins may have different properties depending on the use case. Let’s dive into an example and afterwards explain what was done and how it could have been done differently.

import dclab

with (dclab.new_dataset("input.rtdc") as ds,
      dclab.RTDCWriter("output.rtdc") as hw):

    # First of all, we have to copy the metadata from the input file
    # to the output file. If we forget to do this, then dclab will
    # not be able to open the output file.
    hw.store_metadata(ds.config.as_dict(pop_filtering=True))

    # Next, we can compute and write the new feature to the output file.
    hw.store_feature("userdef1", np.random.random(len(ds)))

    # Finally, we write the basin information to the output file.
    hw.store_basin(
        basin_name="raw data",
        basin_type="file",
        basin_format="hdf5",
        basin_locs=["input.rtdc"],
    )

# You can now open the output file and verify that everything worked.
with dclab.new_dataset("output.rtdc") as ds_out:
    assert "userdef1" in ds_out, "check that the feature we wrote is there"
    assert "image" in ds_out, "check that we can access basin features"
    # You could also be more specific:
    assert "userdef1" in ds_out.features_innate
    assert "image" in ds_out.features_basin

What happened? We created an output.rtdc file that contains the metadata from the ìnput.rtdc file, a feature userdef1 filled with random data, and the basin information referencing all features from the input.rtdc file. We opened the output file and verified that it transparently gives us access to the features stored in the (basin) input file.

What could we have done differently? Take a look at the options that RTDCWriter.store_basin allows you to set. You will notice that there are three major properties that define the nature of a basin:

type of a basin: A basin can be a local file (including files on a network share), or a remote which means that it is accessible via a networking protocol (e.g. HTTP).
format of a basin: This is the subclass of RTDCBase that defines how the basin is accessed. For file-type basins, this is “hdf5” and for remote-type basins, this is “dcor”, “http”, or “s3”.
mapping of a basin: If the output file has the same number of events in the same order as the input file, then we call the mapping “same”. If the output file only contains a subset of the events from the input file, then we call the basin a mapped basin. These mapped basins allow you to minimize data redundancy for analysis pipelines that produce output files with a subset of the events from the input file. To be able to map from the input file to the output file, dclab stores the mapping information as integer indices in dedicated features enumerated basinmap0, basinmap1, etc.
features defined by a basin: You may define basins and explicitly state the features this basin provides. In combination with mapping, you could e.g. realize your own event segmentation pipeline, storing only the mask feature and extracted scalar features in you output file, while you define the image feature via the input file basin. If you combine this approach with the dcor basin format, you can distribute all of your data (raw and processed) in a very efficient and transparent manner.

Examples

Mapped basin via RTDCWriter

You can explicitly define a mapped basin via the RTDCWriter.store_basin method (see also the example after this one).

import dclab
import numpy as np

with (dclab.new_dataset("input.rtdc") as ds,
      dclab.RTDCWriter("output.rtdc") as hw):

    # metadata
    hw.store_metadata(ds.config.as_dict(pop_filtering=True))

    # take every second event from the input file
    event_mapping = np.arange(len(ds), None, 2, dtype=np.uint64)

    # write the basin
    hw.store_basin(
        basin_name="raw data",
        basin_type="file",
        basin_format="hdf5",
        basin_locs=["input.rtdc"],
        basin_map=event_mapping,
    )

# verify that this worked
with (dclab.new_dataset("input.rtdc") as ds_in,
      dclab.new_dataset("output.rtdc") as ds_out):
    assert np.allclose(ds_in["deform"][::2], ds_out["deform"])

Implicitly mapped basin via HDF5 export

It is also possible to implicitly write basin information to an exported file, achieving the same result as above (a very small output file).

import dclab
import numpy as np

with dclab.new_dataset("input.rtdc") as ds:
    # remove every second event
    ds.filter.manual[1::2] = False
    ds.apply_filter()
    # export the dataset with the mapped basin
    ds.export.hdf5(path="output.rtdc",
                   features=[],
                   filtered=True,
                   basins=True)

# verify that this worked
with (dclab.new_dataset("input.rtdc") as ds_in,
      dclab.new_dataset("output.rtdc") as ds_out):
    assert np.allclose(ds_in["deform"][::2], ds_out["deform"])

Basin internals

Storing the basin information

In the output.rtdc file, the basin is stored as a json-encoded string in an HDF5 dataset in the "/basins" group. For the HDF5 export example above, the json data looks like this:

{
  "description": "Exported with dclab 0.58.0",
  "format": "hdf5",
  "name": "Exported data",
  "type": "file",
  "features": null,
  "mapping": "basinmap0",
  "paths": [
    "/absolute/path/to/input.rtdc",
    "input.rtdc"
  ]
}

The description and name are filled automatically by dclab here. As expected, the type of the basin is file and the format of the basin is hdf5. There are a few things to notice:

The features are set to null which means None, i.e. all features from the input file are allowed.
The mapping key reads basinmap0. This is the name of the feature in which to find the mapping information from the input file to the output file. The information can be found in the HDF5 dataset /events/basinmap0 in the output file. Note that the fact that this mapping information is stored as a feature means that it is also properly gated when you define basins iteratively.
There are two paths defined, an absolute path (from the root of the file system) and a relative path (relative to the directory of the output file). This relative path makes it possible to copy-paste these two files together to other locations. You will always be able to open the output file and see the basin features defined in the input file. Internally, dclab also checks the measurement identifier of the output file against that of the input file to avoid loading basin features from the wrong file.

For the sake of completeness, let’s see how the basin information looks like when you derive the output file from a DCOR resource:

import dclab
import numpy as np

with dclab.new_dataset("fb719fb2-bd9f-817a-7d70-f4002af916f0") as ds:
    ds.filter.manual[1::2] = False
    ds.apply_filter()
    ds.export.hdf5(path="output.rtdc",
                   features=[],
                   filtered=True,
                   basins=True)

The corresponding json data:

{
  "description": "Exported with dclab 0.58.0",
  "format": "dcor",
  "name": "Exported data",
  "type": "remote",
  "features": null,
  "mapping": "basinmap0",
  "urls": [
    "https://dcor.mpl.mpg.de/api/3/action/dcserv?id=fb719fb2-bd9f-817a-7d70-f4002af916f0"
  ]
}

As you can see, paths is replaced by urls and the format and type keys changed. The rest remains the same. This also works with private DCOR resources, given that you have globally set your API token as described in the DCOR section.

Basin loading procedure

When dclab opens a dataset the defines a basin, the basin features are retrieved only when they are needed (i.e. when the user tries to access them and they are not defined as innate features). Internally, dclab instantiates an RTDCBase subclass as defined by the format key. For mapped basins, dclab additionally creates a hierarchy child from the original dataset by filling the manual filtering array with the mapping information. To see which features are defined in basins, you can check the RTDCBase.features_basin property. The basins are directly accessible via RTDCBase.basins (and the basin datasets via RTDCBase.basins[index].ds).