PyTerrier Transformers

PyTerrier’s retrieval architecture is based on three concepts:

• dataframes with pre-defined types (each with a minimum set of known attributes), as detailed in the data model.

• the transformation of those dataframes by standard information retrieval operations, defined as transformers.

• the compsition of transformers, supported by the operators defined on transformers.

In essence, a PyTerrier transformer is a class with a transform() method, which takes as input a dataframe, and changes it, before returning it.

Input	Output	Cardinality	Example	Concrete Transformer Example
Q	Q	1 to 1	Query rewriting	pt.rewrite.SDM()
Q	Q x D	1 to N	Retrieval	pt.BatchRetrieve()
Q x D	Q	N to 1	Query expansion	pt.rewrite.RM3()
Q x D	Q x D	1 to 1	Re-ranking	pt.apply.doc_score()
Q x D	Q x Df	1 to 1	Feature scoring	pt.FeaturesBatchRetrieve()

Optimisation

Some operators applied to transformer can be optimised by the underlying search engine - for instance, cutting a ranking earlier. So while the following two pipelines are semantically equivalent, the latter might be more efficient:

pipe1 = BatchRetrieve(index, "BM25") % 10
pipe2 = pipe1.compile()

Fitting

When fit() is called on a pipeline, all estimators (transformers that also have a fit() method, as specified by Estimator) within the pipeline are fitted, in turn. This allows one (or more) stages of learning to be integrated into a retrieval pipeline. See Learning to Rank for examples.

When calling fit on a composed pipeline (i.e. one created using the >> operator), this will will call fit() on any estimators within that pipeline.

Transformer base classes

Transformer

This class is the base class for all transformers.

class pyterrier.Transformer

name = 'Transformer'

Base class for all transformers. Implements the various operators >> + * | & as well as search() for executing a single query and compile() for rewriting complex pipelines into more simples ones.

static identity()

Instantiates a transformer that returns exactly its input.

This can be useful for adding the candidate ranking score as a feature in for learning-to-rank:

bm25 = pt.BatchRetrieve(index, wmodel="BM25")
two_feat_pipe = bm25 >> pt.Transformer.identify() ** pt.BatchRetrieve(index, wmodel="PL2")

This will return a pipeline that produces a score column (BM25), but also has a features column containing BM25 and PL2 scores.

Return type:

Transformer

static from_df(input, uniform=False)

Instantiates a transformer from an input dataframe. Some rows from the input dataframe are returned in response to a query on the transform() method. Depending on the value uniform, the dataframe passed as an argument to transform() can affect this selection.

If uniform is True, input will be returned in its entirety each time. If uniform is False, rows from input that match the qid values from the argument dataframe.

Return type:

Transformer

transform(topics_or_res)

Abstract method for all transformations. Typically takes as input a Pandas DataFrame, and also returns one.

Return type:

DataFrame

transform_iter(input)

Method that proesses an iter-dict by instantiating it as a dataframe and calling transform(). Returns the DataFrame returned by transform(). This can be a handier version of transform() that avoids constructing a dataframe by hand. Alo used in the implementation of index() on a composed pipeline.

Return type:

DataFrame

transform_gen(input, batch_size=1, output_topics=False)

Method for executing a transformer pipeline on smaller batches of queries. The input dataframe is grouped into batches of batch_size queries, and a generator returned, such that transform() is only executed for a smaller batch at a time.

Return type:

Iterator[DataFrame]

Parameters:

• input (DataFrame) – a dataframe to process

• batch_size (int) – how many input instances to execute in each batch. Defaults to 1.

search(query, qid='1', sort=True)

Method for executing a transformer (pipeline) for a single query. Returns a dataframe with the results for the specified query. This is a utility method, and most uses are expected to use the transform() method passing a dataframe.

Return type:

DataFrame

Parameters:

• query (str) – String form of the query to run

• qid (str) – the query id to associate to this request. defaults to 1.

• sort (bool) – ensures the results are sorted by descending rank (defaults to True)

Example:

bm25 = pt.BatchRetrieve(index, wmodel="BM25")
res = bm25.search("example query")

# is equivalent to
queryDf = pd.DataFrame([["1", "example query"]], columns=["qid", "query"])
res = bm25.transform(queryDf)

compile()

Rewrites this pipeline by applying of the Matchpy rules in rewrite_rules. Pipeline optimisation is discussed in the ICTIR 2020 paper on PyTerrier.

Return type:

Transformer

parallel(N, backend='joblib')

Returns a parallelised version of this transformer. The underlying transformer must be “picklable”.

Return type:

Transformer

Parameters:

• N (int) – how many processes/machines to parallelise this transformer over.

• backend (str) – which multiprocessing backend to use. Only two backends are supported, ‘joblib’ and ‘ray’. Defaults to ‘joblib’.

get_parameter(name)

Gets the current value of a particular key of the transformer’s configuration state. By default, this examines the attributes of the transformer object, using hasattr() and setattr().

set_parameter(name, value)

Adjusts this transformer’s configuration state, by setting the value for specific parameter. By default, this examines the attributes of the transformer object, using hasattr() and setattr().

Default Method

You can invoke a transformer’s transfor method simply by calling the default method. If t is a transformer:

df_in = pt.new.queries(['test query'], qid=['q1'])
df_out = t.transform(df_in)
df_out = t(df_in)

The default method can also detect iterable dictionaries, and pass those directly to transform_iter() (which typically calls transform()). So the following expression is equivalent to the examples in the previous code block:

df_out = t([{'qid' : 'q1', 'query' : 'test query'}])

This can be more succinct than creating new dataframes for testing transformer implementations.

Operator Support

By extending Transformer, all transformer implementations gain the necessary “dunder” methods (e.g. __rshift__()) to support the transformer operators (>>, + etc). NB: This class used to be called pyterrier.transformer.TransformerBase

Estimator

This base class exposes a fit() method that can be used for transformers that can be trained.

class pyterrier.Estimator

This is a base class for things that can be fitted.

fit(topics_or_res_tr, qrels_tr, topics_or_res_va, qrels_va)

Method for training the transformer.

Parameters:

• topics_or_res_tr (DataFrame) – training topics (usually with documents)

• qrels_tr (DataFrame) – training qrels

• topics_or_res_va (DataFrame) – validation topics (usually with documents)

• qrels_va (DataFrame) – validation qrels

The ComposedPipeline implements fit(), which applies the interimediate transformers on the specified training (and validation) topics, and places the output into the fit() method of the final transformer.

Indexer

This base class exposes a index() method that can be used for transformers that create an index.

class pyterrier.Indexer

index(iter, **kwargs)

Takes an iterable of dictionaries (“iterdict”), and consumes them. There is no return; This method is typically used to implement indexers.

The ComposedPipeline also implements index(), which applies the interimediate transformers on the specified documents to be indexed, and places the output into the index() method of the final transformer.

Internal transformers

A significant number of transformers are defined in pyterrier.ops to implement operators etc. Its is not expected to use these directly but they are documented for completeness.

Symbol	Name	Implementing transformer
>>	compose/then	ComposedPipeline
|	set-union	SetUnionTransformer
&	set-intersection	SetIntersectionTransformer
+	linear	CombSumTransformer
+	scalar-product	ScalarProductTransformer
%	rank-cutoff	RankCutoffTransformer
**	feature-union	FeatureUnionPipeline
^	concatenate	ConcatenateTransformer
~	cache	ChestCacheTransformer

Indexing Pipelines

Transformers can be chained to create indexing pipelines. The last element in the chain is assumed to be an indexer like IterDictIndexer - it should implement an index() method like pt.Indexer. For instance:

docs = [ {"docno" : "1", "text" : "a" } ]
indexer = pt.text.sliding() >> pt.IterDictIndexer()
indexer.index(docs)

This is implemented by several methods:

• The last stage of the pipeline should have an index() method that accepts an iterable of dictionaries

• ComposedPipeline has a special index() method that breaks the input iterable into chunks (the size of chunks can be altered by a batch_size kwarg) and passes those through the intermediate pipeline stages (i.e. all but the last).

• In the intermediate pipeline stages, the transform_iter() method is called - by default this instantiates a DataFrame on batch_size records, which is passed to transform().

• These are passed to index() of the last pipeline stage.

Writing your own transformer

The first step to writing your own transformer for your own code is to consider the type of change being applied. Several common transformations are supported through the functions in the pyterrier.apply - Custom Transformers module. See the pyterrier.apply - Custom Transformers documentation.

However, if your transformer has state, such as an expensive model to be loaded at startup time, you may want to extend pt.Transformer directly.

Here are some hints for writing Transformers:

• Except for an indexer, you should implement a transform() method.

• If your approach ranks results, use pt.model.add_ranks() to add the rank column. (pt.apply.doc_score will call add_ranks automatically).

• If your approach can be trained, your transformer should extend Estimator, and implement the fit() method.

• If your approach is an indexer, your transformer should extend Indexer and implement index() method.

Mocking Transformers from DataFrames

You can make a Transformer object from dataframes. For instance, a unifom transformer will always return the input dataframe any time transform() is called:

df = pt.new.ranked_documents([[1,2]])
uniformT = pt.Transformer.from_df(df, uniform=True)
# uniformT.transform() always returns df, regardless of arguments

You can also create a Transformer object from existing results, e.g. saved on disk using pt.io.write_results() etc. The resulting “source transformer” will return all results by matching on the qid of the input:

res = pt.io.read_results("/path/to/baseline.res.gz")
baselineT = pt.Transformer.from_df(res, uniform=True)

Q1 = pt.new.queries("test query", qid="Q1")
resQ1 = baselineT.transform(Q1)