speechbrain.utils.data_pipeline module

A pipeline for data transformations.

Example

>>> from hyperpyyaml import load_hyperpyyaml
>>> yamlstring = '''
... pipeline: !new:speechbrain.utils.data_pipeline.DataPipeline
...     static_data_keys: [a, b]
...     dynamic_items:
...         -   func: !name:operator.add
...             takes: ["a", "b"]
...             provides: foo
...         -   func: !name:operator.sub
...             takes: ["foo", "b"]
...             provides: bar
...     output_keys: ["foo", "bar"]
... '''
>>> hparams = load_hyperpyyaml(yamlstring)
>>> hparams["pipeline"]({"a":1, "b":2})
{'foo': 3, 'bar': 1}
Author:
  • Aku Rouhe

Summary

Classes:

DataPipeline

Organises data transformations into a pipeline.

DynamicItem

Essentially represents a data transformation function.

GeneratorDynamicItem

Essentially represents a multi-step data transformation.

StaticItem

Data class that represents a static item.

Functions:

provides

Decorator which makes a DynamicItem and specifies what keys it provides.

provides_decorator

Decorator which makes a DynamicItem and specifies what keys it provides.

takes

Decorator which makes a DynamicItem and specifies its argkeys.

takes_decorator

Decorator which makes a DynamicItem and specifies its argkeys.

Reference

class speechbrain.utils.data_pipeline.StaticItem(key: str)[source]

Bases: object

Data class that represents a static item.

Static items are in-memory items so they don’t need to be computed dynamically.

key: str
class speechbrain.utils.data_pipeline.DynamicItem(takes=[], func=None, provides=[])[source]

Bases: object

Essentially represents a data transformation function.

A DynamicItem takes some arguments and computes its value dynamically when called. A straight-forward use-case is to load something from disk dynamically; take the path and provide the loaded data.

Instances of this class are often created implicitly via the @takes and @provides decorators or otherwise from specifying the taken and provided arguments and the function.

A counterpart is the GeneratorDynamicItem, which should be used for generator functions.

Parameters
  • takes (list) – The keys of the items that this needs to compute its output.

  • func (callable) – The function that is used to compute the output.

  • provides (list) – The keys that this provides.

next_takes()[source]

The next argkeys to provide to this, when called.

next_provides()[source]

The next keys that this provides, when called.

provided_in_order()[source]

Assuming that this may need to be called multiple times; which keys does it provide at that call. Returns a list, with len equal to the number of times that this may be called.

reset()[source]

Signals that this will not be called any more times on this pipeline call.

class speechbrain.utils.data_pipeline.GeneratorDynamicItem(*args, **kwargs)[source]

Bases: speechbrain.utils.data_pipeline.DynamicItem

Essentially represents a multi-step data transformation.

This is the generator function counterpart for DynamicItem (which should be used for regular functions).

A GeneratorDynamicItem first takes some arguments and then uses those in multiple steps to incrementally compute some values when called.

A typical use-case is a pipeline of transformations on data: e.g. taking in text as a string, and first a tokenized version, and then on the second call providing an integer-encoded version. This can be used even though the integer-encoder needs to be trained on the first outputs.

The main benefit is to be able to define the pipeline in a clear function, even if parts of the pipeline depend on others for their initialization.

Example

>>> lab2ind = {}
>>> def text_pipeline(text):
...     text = text.lower().strip()
...     text = "".join(c for c in text if c.isalpha() or c == " ")
...     words = text.split()
...     yield words
...     encoded = [lab2ind[word] for word in words]
...     yield encoded
>>> item = GeneratorDynamicItem(
...         func=text_pipeline,
...         takes=["text"],
...         provides=["words", "words_encoded"])
>>> # First create the integer-encoding:
>>> ind = 1
>>> for token in item("Is this it? - This is it."):
...     if token not in lab2ind:
...         lab2ind[token] = ind
...         ind += 1
>>> # Now the integers can be encoded!
>>> item()
[1, 2, 3, 2, 1, 3]
next_takes()[source]
next_provides()[source]
provided_in_order()[source]
reset()[source]
speechbrain.utils.data_pipeline.takes(*argkeys)[source]

Decorator which makes a DynamicItem and specifies its argkeys.

If the wrapped object is a generator function (has a yield statement), Creates a GeneratorDynamicItem. If the object is already a DynamicItem, just specifies the argkeys for that. Otherwise creates a new regular DynamicItem, with argkeys specified.

The args are always passed to the function at the start. Generators could support sending new arguments, but for such use cases, simply create a new dynamic item. The GeneratorDynamicItem class is meant for pipelines which take in an input and transform it in multiple ways, where the intermediate representations may be needed for e.g. fitting a BPE segmenter.

Example

>>> @takes("text")
... def tokenize(text):
...     return text.strip().lower().split()
>>> tokenize.provides = ["tokenized"]
>>> tokenize('      This Example gets tokenized')
['this', 'example', 'gets', 'tokenized']
speechbrain.utils.data_pipeline.takes_decorator(*argkeys)

Decorator which makes a DynamicItem and specifies its argkeys.

If the wrapped object is a generator function (has a yield statement), Creates a GeneratorDynamicItem. If the object is already a DynamicItem, just specifies the argkeys for that. Otherwise creates a new regular DynamicItem, with argkeys specified.

The args are always passed to the function at the start. Generators could support sending new arguments, but for such use cases, simply create a new dynamic item. The GeneratorDynamicItem class is meant for pipelines which take in an input and transform it in multiple ways, where the intermediate representations may be needed for e.g. fitting a BPE segmenter.

Example

>>> @takes("text")
... def tokenize(text):
...     return text.strip().lower().split()
>>> tokenize.provides = ["tokenized"]
>>> tokenize('      This Example gets tokenized')
['this', 'example', 'gets', 'tokenized']
speechbrain.utils.data_pipeline.provides(*output_keys)[source]

Decorator which makes a DynamicItem and specifies what keys it provides.

If the wrapped object is a generator function (has a yield statement), Creates a GeneratorDynamicItem. If the object is already a DynamicItem, just specifies the provided keys for that. Otherwise creates a new regular DynamicItem, with provided keys specified.

Note

The behavior is slightly different for generators and regular functions, if many output keys are specified, e.g. @provides(“signal”, “mfcc”). Regular functions should return a tuple with len equal to len(output_keys), while generators should yield the items one by one.

>>> @provides("signal", "feat")
... def read_feat():
...     wav = [.1,.2,-.1]
...     feat = [s**2 for s in wav]
...     return wav, feat
>>> @provides("signal", "feat")
... def read_feat():
...     wav = [.1,.2,-.1]
...     yield wav
...     feat = [s**2 for s in wav]
...     yield feat

If multiple keys are yielded at once, write e.g.,

>>> @provides("wav_read", ["left_channel", "right_channel"])
... def read_multi_channel():
...     wav = [[.1,.2,-.1],[.2,.1,-.1]]
...     yield wav
...     yield wav[0], wav[1]
speechbrain.utils.data_pipeline.provides_decorator(*output_keys)

Decorator which makes a DynamicItem and specifies what keys it provides.

If the wrapped object is a generator function (has a yield statement), Creates a GeneratorDynamicItem. If the object is already a DynamicItem, just specifies the provided keys for that. Otherwise creates a new regular DynamicItem, with provided keys specified.

Note

The behavior is slightly different for generators and regular functions, if many output keys are specified, e.g. @provides(“signal”, “mfcc”). Regular functions should return a tuple with len equal to len(output_keys), while generators should yield the items one by one.

>>> @provides("signal", "feat")
... def read_feat():
...     wav = [.1,.2,-.1]
...     feat = [s**2 for s in wav]
...     return wav, feat
>>> @provides("signal", "feat")
... def read_feat():
...     wav = [.1,.2,-.1]
...     yield wav
...     feat = [s**2 for s in wav]
...     yield feat

If multiple keys are yielded at once, write e.g.,

>>> @provides("wav_read", ["left_channel", "right_channel"])
... def read_multi_channel():
...     wav = [[.1,.2,-.1],[.2,.1,-.1]]
...     yield wav
...     yield wav[0], wav[1]
class speechbrain.utils.data_pipeline.DataPipeline(static_data_keys, dynamic_items=[], output_keys=[])[source]

Bases: object

Organises data transformations into a pipeline.

Example

>>> pipeline = DataPipeline(
...     static_data_keys=["text"],
...     dynamic_items=[
...     {"func": lambda x: x.lower(), "takes": "text", "provides": "foo"},
...     {"func": lambda x: x[::-1], "takes": "foo", "provides": "bar"},
...     ],
...     output_keys=["bar"],
... )
>>> pipeline({"text": "Test"})
{'bar': 'tset'}
add_static_keys(static_keys)[source]

Informs the pipeline about static items.

Static items are the ones provided to __call__ as data.

add_dynamic_items(dynamic_items)[source]

Add multiple dynamic items at once.

add_dynamic_item(func, takes=None, provides=None)[source]

Adds a dynamic item to the Pipeline.

Two calling conventions. For DynamicItem objects, just use: add_dynamic_item(dynamic_item) But otherwise, should use: add_dynamic_item(func, takes, provides)

Parameters
  • func (callable, DynamicItem) – If a DynamicItem is given, adds that directly. Otherwise a DynamicItem is created, and this specifies the callable to use. If a generator function is given, then create a GeneratorDynamicItem. Otherwise creates a normal DynamicItem.

  • takes (list, str) – List of keys. When func is called, each key is resolved to either an entry in the data or the output of another dynamic_item. The func is then called with these as positional arguments, in the same order as specified here. A single key can be given as a bare string.

  • provides (str, list) – For regular functions, the key or list of keys that it provides. If you give a generator function, key or list of keys that it yields, in order. Also see the provides decorator. A single key can be given as a bare string.

set_output_keys(keys)[source]

Use this to change the output keys.

Also re-evaluates execution order. So if you request different outputs, some parts of the data pipeline may be skipped.

Parameters

keys (dict, list, None) –

List of keys (str) to produce in output.

If a dict is given; it is used to map internal keys to output keys. From the output_keys dict key:value pairs the key appears outside, and value is the internal key.

compute_outputs(data)[source]
Parameters

data (dict) – Dictionary with data entries by key.

Returns

With the keys that were set.

Return type

dict

compute_specific(keys, data)[source]

Compute output of specific item, without changing output_keys.

get_selected_node_ids(selected_keys)[source]

Translates selected keys to dependency graph keys.