def test_cache_compose_cache(self):
pt.cache.CACHE_DIR = self.test_dir
import pandas as pd
queries = pd.DataFrame([["q1", "chemical"]], columns=["qid", "query"])
br1 = pt.BatchRetrieve(pt.get_dataset("vaswani").get_index(), wmodel="TF_IDF")
br2 = pt.BatchRetrieve(pt.get_dataset("vaswani").get_index(), wmodel="BM25")
cache = ~ (~br1 >> br2)
self.assertEqual(0, len(cache.chest._keys))
cache(queries)
cache(queries)
self.assertEqual(0.5, cache.stats())
#lets see if another cache of the same object would see the same cache entries.
cache2 = ~(~br1 >> br2)
cache2(queries)
self.assertEqual(1, cache2.stats())
# check that the cache report works
all_report = pt.cache.list_cache()
self.assertTrue(len(all_report) > 0)
report = list(all_report.values())[0]
self.assertEqual(1, report["queries"])
self.assertTrue("transformer" in report)
self.assertTrue("size" in report)
self.assertTrue("lastmodified" in report)
pt.cache.CACHE_DIR = None
def test_mrt(self):
brs = [
pt.BatchRetrieve(pt.datasets.get_dataset("vaswani").get_index(),
wmodel="DPH"),
pt.BatchRetrieve(pt.datasets.get_dataset("vaswani").get_index(),
wmodel="BM25")
]
topics = pt.datasets.get_dataset("vaswani").get_topics().head(10)
qrels = pt.datasets.get_dataset("vaswani").get_qrels()
pt.Experiment(brs,
topics,
qrels,
eval_metrics=["map", "mrt"],
baseline=0,
highlight="color")
pt.Experiment(brs,
topics,
qrels,
eval_metrics=["map", "mrt"],
highlight="color")
pt.Experiment(brs, topics, qrels, eval_metrics=["map", "mrt"])
pt.Experiment(brs,
topics,
qrels,
eval_metrics=["map", "mrt"],
baseline=0,
highlight="color")
def test_fbr_reranking(self):
if not pt.check_version("5.3"):
self.skipTest("Requires Terrier 5.3")
# this test examines the use of ScoringMatchingWithFat
JIR = pt.autoclass('org.terrier.querying.IndexRef')
indexref = JIR.of(self.here + "/fixtures/index/data.properties")
# we only want a candidate set of 2 documents
firstpass = pt.BatchRetrieve(indexref, wmodel="BM25") % 2
pipe = firstpass >> pt.FeaturesBatchRetrieve(
indexref, features=["WMODEL:DPH", "WMODEL:PL2"])
input = pd.DataFrame([["1", "Stability"]], columns=['qid', 'query'])
result = pipe.transform(input)
self.assertTrue("qid" in result.columns)
self.assertTrue("docno" in result.columns)
self.assertTrue("score" in result.columns)
self.assertTrue("features" in result.columns)
self.assertEqual(2, len(result))
self.assertEqual(result.iloc[0]["features"].size, 2)
pipe_simple = firstpass >> (pt.BatchRetrieve(indexref, wmodel="DPH")**
pt.BatchRetrieve(indexref, wmodel="PL2"))
result2 = pipe.transform(input)
import numpy as np
f1 = np.stack(result["features"].values)
f2 = np.stack(result2["features"].values)
self.assertTrue(np.array_equal(f1, f2))
def _check_index(index_small):
if has_direct1:
contents1 = TestBackground.get_contents(0, index_small)
self.assertEqual(contents1, contents_big)
inv1 = index_small.getInvertedIndex()
print(inv1.getClass().getName())
lex1 = index_small.getLexicon()
for t in contents_big:
pointer = lex1[t]
print(pointer.toString())
p = inv1.getPostings(pointer)
print(p.getClass().getName())
rtr = p.next()
self.assertEqual(0, rtr)
self.assertEqual(
p.getDocumentLength(),
index_big.getDocumentIndex().getDocumentLength(1047))
self.assertEqual(contents_big[t], p.getFrequency())
self.assertEqual(p.next(), p.EOL)
from jnius import JavaException
try:
br1 = pt.BatchRetrieve(index_small, wmodel="Tf")
brall = pt.BatchRetrieve(index_big, wmodel="Tf")
with_doc = pd.DataFrame(
[["q1", q, "1048", 1047]],
columns=["qid", "query", "docno", "docid"])
rtr1 = br1.transform(q)
except JavaException as ja:
print(ja.stacktrace)
raise ja
rtrall = brall(with_doc)
self.assertTrue(
np.array_equal(rtr1["score"].values, rtrall["score"].values))
def test_save_docs_CE(self):
index = pt.get_dataset("vaswani").get_index()
dph = pt.BatchRetrieve(index, wmodel="DPH")
pipe = dph \
>> pt.rewrite.stash_results() \
>> pt.BatchRetrieve(index, wmodel="BM25") \
>> pt.rewrite.Bo1QueryExpansion(index) \
>> pt.rewrite.reset_results() \
>> dph
rtr1 = dph.search("chemical reactions")
rtr2 = pipe.search("chemical reactions")
# Bo1 should be applied as a re-ranker, hence the
# number of docs in rtr1 and rtr2 should be equal
self.assertEqual(len(rtr1), len(rtr2))
# check columns are passed through where we expect
pipeP3 = dph \
>> pt.rewrite.stash_results() \
>> pt.BatchRetrieve(index, wmodel="BM25")
res3 = pipeP3.search("chemical reactions")
self.assertIn("stashed_results_0", res3.columns)
pipeP4 = dph \
>> pt.rewrite.stash_results() \
>> pt.BatchRetrieve(index, wmodel="BM25") \
>> pt.rewrite.Bo1QueryExpansion(index)
res4 = pipeP3.search("chemical reactions")
self.assertIn("stashed_results_0", res4.columns)
def test_qe(self):
if not pt.check_version("5.3"):
self.skipTest("Requires Terrier 5.3")
dataset = pt.datasets.get_dataset("vaswani")
indexref = dataset.get_index()
qe = pt.rewrite.QueryExpansion(indexref)
br = pt.BatchRetrieve(indexref)
queriesIn = pd.DataFrame([["1", "compact"]], columns=["qid", "query"])
res = br.transform(queriesIn)
queriesOut = qe.transform(res)
self.assertEqual(len(queriesOut), 1)
query = queriesOut.iloc[0]["query"]
self.assertTrue("compact^1.82230972" in query)
self.assertTrue("applypipeline:off " in query)
pipe = br >> qe >> br
# lets go faster, we only need 18 topics. qid 16 had a tricky case
t = dataset.get_topics().head(18)
all_qe_res = pipe.transform(t)
map_pipe = pt.Utils.evaluate(all_qe_res, dataset.get_qrels(), metrics=["map"])["map"]
br_qe = pt.BatchRetrieve(indexref, controls={"qe":"on"})
map_qe = pt.Utils.evaluate(br_qe.transform(t), dataset.get_qrels(), metrics=["map"])["map"]
self.assertAlmostEqual(map_qe, map_pipe, places=4)
def _sdm(self, freq):
dataset = pt.datasets.get_dataset("vaswani")
indexer = pt.TRECCollectionIndexer(self.test_dir, blocks=True)
indexref = indexer.index(dataset.get_corpus())
if freq:
sdm = pt.rewrite.SDM(prox_model="org.terrier.matching.models.Tf")
else:
sdm = pt.rewrite.SDM()
queriesIn = pd.DataFrame([["1", "compact"], ["2", "compact memories"]], columns=["qid", "query"])
queriesOut = sdm.transform(queriesIn)
self.assertEqual(len(queriesOut), 2)
self.assertEqual(queriesOut.iloc[0]["query"], "compact")
query2 = queriesOut.iloc[1]["query"]
self.assertTrue("#1" in query2)
self.assertTrue("#uw8" in query2)
self.assertTrue("#combine" in query2)
br_normal = pt.BatchRetrieve(indexref)
pipe = sdm >> br_normal
if freq:
br_normal.controls["wmodel"] = "Tf"
resTest_pipe = pipe.transform(queriesIn)
# this BR should do the same thing as the pipe, but natively in Terrier
br_sdm = pt.BatchRetrieve(indexref,
controls = {"sd" :"on"},
properties={"querying.processes" : "terrierql:TerrierQLParser,parsecontrols:TerrierQLToControls,"\
+"parseql:TerrierQLToMatchingQueryTerms,matchopql:MatchingOpQLParser,applypipeline:ApplyTermPipeline,"\
+"sd:DependenceModelPreProcess,localmatching:LocalManager$ApplyLocalMatching,qe:QueryExpansion,"\
+"labels:org.terrier.learning.LabelDecorator,filters:LocalManager$PostFilterProcess"})
if freq:
br_sdm.controls["wmodel"] = "Tf"
br_sdm.controls["dependencemodel"] = "org.terrier.matching.models.Tf"
resTest_native = br_sdm.transform(queriesIn)
#print (resTest_pipe[resTest_pipe["qid"]=="2"])
#print (resTest_native[resTest_native["qid"]=="2"])
for index, row in resTest_pipe.iterrows():
#print(index)
#print(row["query"])
#print(row)
#print(resTest_native.iloc[index])
self.assertEqual(row['qid'], resTest_native.iloc[index]["qid"])
self.assertEqual(row['docno'], resTest_native.iloc[index]["docno"])
# TODO I cannot get this test to pass with freq=False more precisely than 1dp
#9.165638 in resTest_pipe vs 9.200683 in resTest_native
self.assertAlmostEqual(row['score'], resTest_native.iloc[index]["score"], 1)
t = dataset.get_topics().head(5)
pipe_res = pipe.transform(t)
#br_normal.saveResult(pipe_res, "/tmp/sdm.res", run_name="DPH")
self.assertAlmostEqual(
pt.Utils.evaluate(pipe_res, dataset.get_qrels(), metrics=["map"])["map"],
pt.Utils.evaluate(br_sdm.transform(t), dataset.get_qrels(), metrics=["map"])["map"],
places=4)
def test_wrong(self):
brs = [
pt.BatchRetrieve(pt.datasets.get_dataset("vaswani").get_index(), wmodel="DPH"),
pt.BatchRetrieve(pt.datasets.get_dataset("vaswani").get_index(), wmodel="BM25")
]
topics = pt.datasets.get_dataset("vaswani").get_topics().head(10)
qrels = pt.datasets.get_dataset("vaswani").get_qrels()
with self.assertRaises(TypeError):
pt.Experiment(brs, topics, qrels, eval_metrics=["map"], filter_qrels=True)
def test_baseline(self):
dataset = pt.get_dataset("vaswani")
df = pt.Experiment(
[pt.BatchRetrieve(dataset.get_index(), wmodel="BM25"), pt.BatchRetrieve(dataset.get_index(), wmodel="DPH")],
dataset.get_topics().head(10),
dataset.get_qrels(),
eval_metrics=["map", "ndcg"],
baseline=0)
self.assertTrue("map +" in df.columns)
self.assertTrue("map -" in df.columns)
self.assertTrue("map p-value" in df.columns)
def test_parallel_joblib_ops(self):
dataset = pt.get_dataset("vaswani")
topics = dataset.get_topics().head(3)
dph = pt.BatchRetrieve(dataset.get_index())
tf = pt.BatchRetrieve(dataset.get_index(), wmodel="Tf")
for pipe in [
dph, dph % 10, dph >> tf, dph + tf,
pt.apply.query(lambda row: row["query"] + " chemical") >> dph
]:
res1 = pipe(topics)
res2 = pipe.parallel(3)(topics)
self.assertEqual(len(res1), len(res2))
def test_mrt(self):
index = pt.datasets.get_dataset("vaswani").get_index()
brs = [
pt.BatchRetrieve(index, wmodel="DPH"),
pt.BatchRetrieve(index, wmodel="BM25")
]
topics = pt.datasets.get_dataset("vaswani").get_topics().head(10)
qrels = pt.datasets.get_dataset("vaswani").get_qrels()
measures = ["map", "mrt"]
pt.Experiment(brs, topics, qrels, eval_metrics=measures)
self.assertTrue("mrt" in measures)
pt.Experiment(brs, topics, qrels, eval_metrics=["map", "mrt"], highlight="color")
pt.Experiment(brs, topics, qrels, eval_metrics=["map", "mrt"], baseline=0, highlight="color")
def test_baseline_and_tests(self):
dataset = pt.get_dataset("vaswani")
numt=10
res1 = pt.BatchRetrieve(dataset.get_index(), wmodel="BM25")(dataset.get_topics().head(numt))
res2 = pt.BatchRetrieve(dataset.get_index(), wmodel="DPH")(dataset.get_topics().head(numt))
# t-test
with warnings.catch_warnings():
warnings.simplefilter("ignore")
df = pt.Experiment(
[res1, res2],
dataset.get_topics().head(numt),
dataset.get_qrels(),
eval_metrics=["map", "ndcg"],
baseline=0)
self.assertTrue("map +" in df.columns)
self.assertTrue("map -" in df.columns)
self.assertTrue("map p-value" in df.columns)
# wilcoxon signed-rank test
df = pt.Experiment(
[res1, res2],
dataset.get_topics().head(numt),
dataset.get_qrels(),
eval_metrics=["map", "ndcg"],
test='wilcoxon',
baseline=0)
self.assertTrue("map +" in df.columns)
self.assertTrue("map -" in df.columns)
self.assertTrue("map p-value" in df.columns)
# user-specified TOST
# TOST will omit warnings here, due to low numbers of topics
import statsmodels.stats.weightstats
fn = lambda X,Y: (0, statsmodels.stats.weightstats.ttost_ind(X, Y, -0.01, 0.01)[0])
#This filter doesnt work
with warnings.catch_warnings(record=True) as w:
warnings.filterwarnings("always")
df = pt.Experiment(
[res1, res2],
dataset.get_topics().head(numt),
dataset.get_qrels(),
eval_metrics=["map", "ndcg"],
test=fn,
baseline=0)
print(w)
self.assertTrue("map +" in df.columns)
self.assertTrue("map -" in df.columns)
self.assertTrue("map p-value" in df.columns)
def test_fbr_reranking2(self):
if not pt.check_version("5.4"):
self.skipTest("Requires Terrier 5.4")
# this test examines the use of ScoringMatchingWithFat, using a particular case known to with Terrier 5.3
JIR = pt.Class('org.terrier.querying.IndexRef')
indexref = JIR.of(self.here + "/fixtures/index/data.properties")
# we only want a candidate set of 3 documents
firstpass = pt.BatchRetrieve(indexref, wmodel="BM25") % 3
pipe1 = firstpass >> pt.FeaturesBatchRetrieve(indexref,
features=["WMODEL:PL2"])
pipe2 = firstpass >> pt.BatchRetrieve(indexref, wmodel="PL2")
input = pd.DataFrame([["1", "Stability"]], columns=['qid', 'query'])
result0 = firstpass.transform(input)
result1 = pipe1.transform(input)
result2 = pipe2.transform(input)
result1["feature0"] = result1.apply(lambda row: row["features"][0],
axis=1)
#BM25 score
result0_map = {row.docno: row.score for row in result0.itertuples()}
result1S_map = {row.docno: row.score for row in result1.itertuples()}
#PL2 score
result1F_map = {
row.docno: row.feature0
for row in result1.itertuples()
}
result2_map = {row.docno: row.score for row in result2.itertuples()}
print(result1F_map)
print(result2_map)
# check features scores
# NB: places can go no less than 4, as two documents have similar PL2 scores
for rank, row in enumerate(result0.itertuples()):
docno = row.docno
# check that score is unchanged
self.assertAlmostEqual(
result1S_map[docno],
result0_map[docno],
msg="input score mismatch at rank %d for docno %s" %
(rank, docno),
places=4)
# check that feature score is correct
self.assertAlmostEqual(
result1F_map[docno],
result2_map[docno],
msg="feature score mismatch at rank %d for docno %s" %
(rank, docno),
places=4)
def test_num_results(self):
JIR = pt.autoclass('org.terrier.querying.IndexRef')
indexref = JIR.of(self.here + "/fixtures/index/data.properties")
retr = pt.BatchRetrieve(indexref, num_results=10)
input = pd.DataFrame([["1", "Stability"]], columns=['qid', 'query'])
result = retr.transform(input)
self.assertEqual(len(result), 10)
if not pt.check_version("5.5"):
return
retr = pt.BatchRetrieve(indexref, num_results=1001)
result = retr.search("results")
self.assertEqual(len(result), 1001)
def test_compile_to_fbr(self):
indexref = pt.IndexRef.of(self.here + "/fixtures/index/data.properties")
# we only want a candidate set of 2 documents
firstpass = pt.BatchRetrieve(indexref, wmodel="BM25")
pipe_f_fbr = firstpass >> pt.FeaturesBatchRetrieve(indexref, features=["WMODEL:DPH", "WMODEL:PL2"])
pipe_fbr = pt.FeaturesBatchRetrieve(indexref, wmodel="BM25", features=["WMODEL:DPH", "WMODEL:PL2"])
pipe_raw = firstpass >> ( pt.BatchRetrieve(indexref, wmodel="DPH") ** pt.BatchRetrieve(indexref, wmodel="PL2") )
input = pd.DataFrame([["1", "Stability"]], columns=['qid', 'query'])
res1 = (pipe_f_fbr %2)(input)
res2 = (pipe_fbr % 2)(input)
res3 = (pipe_raw % 2)(input)
compiled = (pipe_raw % 2).compile()
print(repr(compiled))
res4 = compiled(input)
def test_baseline_corrected(self):
dataset = pt.get_dataset("vaswani")
for corr in ['hs', 'bonferroni', 'holm-sidak']:
df = pt.Experiment(
[pt.BatchRetrieve(dataset.get_index(), wmodel="BM25"), pt.BatchRetrieve(dataset.get_index(), wmodel="DPH")],
dataset.get_topics().head(10),
dataset.get_qrels(),
eval_metrics=["map", "ndcg"],
baseline=0, correction='hs')
self.assertTrue("map +" in df.columns)
self.assertTrue("map -" in df.columns)
self.assertTrue("map p-value" in df.columns)
self.assertTrue("map p-value corrected" in df.columns)
self.assertTrue("map reject" in df.columns)
def test_various_metrics(self):
topics = pt.datasets.get_dataset("vaswani").get_topics().head(10)
res = [
pt.BatchRetrieve(pt.datasets.get_dataset("vaswani").get_index(),
wmodel="DPH")(topics),
pt.BatchRetrieve(pt.datasets.get_dataset("vaswani").get_index(),
wmodel="BM25")(topics)
]
qrels = pt.datasets.get_dataset("vaswani").get_qrels()
# what we ask for -> what we should get as a metric
family2measure = {
'ndcg_cut_5': 'ndcg_cut_5',
'P': "P@5",
'P_5': "P_5",
"iprec_at_recall": "[email protected]",
"official": "AP",
"set": "SetP",
"recall": "R@5",
"recall_1000": "recall_1000"
}
# what we ask for -> what we should NOT get
family2black = {
'ndcg_cut_5': 'ndcg_cut_10',
'P_5': "P_100",
"recall_1000": "recall_5"
}
for m in family2measure:
df1 = pt.Experiment(res, topics, qrels, eval_metrics=[m])
df2 = pt.Experiment(res,
topics,
qrels,
eval_metrics=[m],
baseline=0)
df3 = pt.Experiment(res,
topics,
qrels,
eval_metrics=[m],
perquery=True)
self.assertIn(family2measure[m], df1.columns)
self.assertIn(family2measure[m], df2.columns)
self.assertTrue(len(df3[df3["measure"] == family2measure[m]]) > 0)
# check that we dont get back measures that we did NOT ask for
if m in family2black:
self.assertNotIn(family2black[m], df1.columns)
self.assertNotIn(family2black[m], df2.columns)
self.assertTrue(
len(df3[df3["measure"] == family2black[m]]) == 0)
def test_save(self):
index = pt.datasets.get_dataset("vaswani").get_index()
brs = [
pt.BatchRetrieve(index, wmodel="DPH"),
pt.BatchRetrieve(index, wmodel="BM25")
]
topics = pt.datasets.get_dataset("vaswani").get_topics().head(10)
qrels = pt.datasets.get_dataset("vaswani").get_qrels()
df1 = pt.Experiment(brs, topics, qrels, eval_metrics=["map", "mrt"], save_dir=self.test_dir)
# check save_dir files are there
self.assertTrue(os.path.exists(os.path.join(self.test_dir, "BR(DPH).res.gz")))
self.assertTrue(os.path.exists(os.path.join(self.test_dir, "BR(BM25).res.gz")))
df2 = pt.Experiment(brs, topics, qrels, eval_metrics=["map", "mrt"], save_dir=self.test_dir)
# a successful experiment using save_dir should be faster
self.assertTrue(df2.iloc[0]["mrt"] < df1.iloc[0]["mrt"])
def test_num_python_wmodel(self):
indexref = self.here + "/fixtures/index/data.properties"
Tf = lambda keyFreq, posting, entryStats, collStats: posting.getFrequency(
)
retr = pt.BatchRetrieve(indexref, wmodel=Tf)
input = pd.DataFrame([["1", "Stability"]], columns=['qid', 'query'])
result = retr.transform(input)
def test_two_term_query_correct_qid_docid_score(self):
JIR = pt.autoclass('org.terrier.querying.IndexRef')
indexref = JIR.of(self.here + "/fixtures/index/data.properties")
retr = pt.BatchRetrieve(indexref)
input = pd.DataFrame([["1", "Stability"], ["2", "Generator"]],
columns=['qid', 'query'])
result = retr.transform(input)
exp_result = parse_res_file(
os.path.dirname(os.path.realpath(__file__)) +
"/fixtures/two_queries_result")
for index, row in result.iterrows():
self.assertEqual(row['qid'], exp_result[index][0])
self.assertEqual(row['docno'], exp_result[index][1])
self.assertAlmostEqual(row['score'], exp_result[index][2])
input = pd.DataFrame([[1, "Stability"], [2, "Generator"]],
columns=['qid', 'query'])
result = retr.transform(input)
exp_result = parse_res_file(
os.path.dirname(os.path.realpath(__file__)) +
"/fixtures/two_queries_result")
for index, row in result.iterrows():
self.assertEqual(str(row['qid']), exp_result[index][0])
self.assertEqual(row['docno'], exp_result[index][1])
self.assertAlmostEqual(row['score'], exp_result[index][2])
def rank_publications(self, query, page, rpp):
itemlist = []
if query is not None:
if self.idx is None:
try:
self.idx = pt.IndexFactory.of('./index/data.properties')
except Exception as e:
print('No index available: ', e)
if self.idx is not None:
topics = pd.DataFrame.from_dict({'qid': [0], 'query': [query]})
retr = pt.BatchRetrieve(self.idx, controls={"wmodel": "TF_IDF"})
retr.setControl("wmodel", "TF_IDF")
retr.setControls({"wmodel": "TF_IDF"})
res = retr.transform(topics)
itemlist = list(res['docno'][page * rpp:(page + 1) * rpp])
return {
'page': page,
'rpp': rpp,
'query': query,
'itemlist': itemlist,
'num_found': len(itemlist)
}
def test_vaswani(self):
import pyterrier as pt
dataset = pt.datasets.get_dataset("vaswani")
self.assertIsNotNone(dataset)
topics = dataset.get_topics()
self.assertIsNotNone(topics)
self.assertEqual(len(topics), 93)
qrels = dataset.get_qrels()
self.assertIsNotNone(qrels)
self.assertEqual(len(qrels), 2083)
indexref = dataset.get_index()
self.assertIsNotNone(indexref)
with pt.IndexFactory.of(indexref) as index:
self.assertIsNotNone(index)
self.assertEqual(
index.getCollectionStatistics().getNumberOfDocuments(), 11429)
# do it once again, to ensure it works locally
dataset = pt.datasets.get_dataset("vaswani")
topics = dataset.get_topics()
self.assertIsNotNone(topics)
self.assertEqual(len(topics), 93)
# test the newer get_topicsqrels
pt.Experiment([pt.BatchRetrieve(dataset.get_index())],
*dataset.get_topicsqrels(), ["map"])
def __init__(self,
candidates,
num_candidates_samples,
path_index,
sample_data,
set_rm3=False,
seed=42):
random.seed(seed)
self.candidates = candidates
self.candidates_df = pd.DataFrame(
self.candidates,
columns=["candidate"
]).reset_index().rename(columns={'index': 'docno'})
self.candidates_df['docno'] = self.candidates_df['docno'].astype(str)
self.num_candidates_samples = num_candidates_samples
self.path_index = path_index
if set_rm3:
self.name = "BM25RM3NS_pyterrier"
else:
self.name = "BM25NS_pyterrier"
self.sample_data = sample_data
self._create_index()
self.bm25_pipeline = pt.BatchRetrieve(
self.indexref, wmodel="BM25") % self.num_candidates_samples
def test_batching(self):
vaswani = pt.datasets.get_dataset("vaswani")
br = pt.BatchRetrieve(vaswani.get_index())
rtr1 = pt.Experiment([br],
vaswani.get_topics().head(10),
vaswani.get_qrels(),
["map", "ndcg", "num_q", "mrt"])
rtr2 = pt.Experiment([br],
vaswani.get_topics().head(10),
vaswani.get_qrels(),
["map", "ndcg", "num_q", "mrt"],
batch_size=2)
self.assertTrue("mrt" in rtr1.columns)
self.assertTrue("mrt" in rtr2.columns)
rtr1.drop(columns=["mrt"], inplace=True)
rtr2.drop(columns=["mrt"], inplace=True)
pd.testing.assert_frame_equal(rtr1, rtr2)
rtr1 = pt.Experiment([br],
vaswani.get_topics().head(10),
vaswani.get_qrels(), ["map", "ndcg", "num_q"],
perquery=True)
rtr2 = pt.Experiment([br],
vaswani.get_topics().head(10),
vaswani.get_qrels(), ["map", "ndcg", "num_q"],
batch_size=2,
perquery=True)
pd.testing.assert_frame_equal(rtr1, rtr2)
def test_bad_measure(self):
vaswani = pt.datasets.get_dataset("vaswani")
br = pt.BatchRetrieve(vaswani.get_index())
with self.assertRaises(KeyError):
pt.Experiment([br],
vaswani.get_topics().head(10), vaswani.get_qrels(),
[map])
def test_parallel_joblib_experiment(self):
self.skip_windows()
dataset = pt.get_dataset("vaswani")
br = pt.BatchRetrieve(dataset.get_index())
df = pt.Experiment([br, br.parallel(3)], dataset.get_topics(),
dataset.get_qrels(), ["map", "mrt"])
self.assertEqual(df.iloc[0]["map"], df.iloc[1]["map"])
def test_perquery(self):
vaswani = pt.datasets.get_dataset("vaswani")
br = pt.BatchRetrieve(vaswani.get_index())
rtr = pt.Experiment([br], vaswani.get_topics().head(10), vaswani.get_qrels(), ["map", "ndcg"], perquery=True)
print(rtr)
rtr = pt.Experiment([br], vaswani.get_topics().head(10), vaswani.get_qrels(), ["map", "ndcg"], perquery=True, dataframe=False)
print(rtr)
def test_qe(self):
if not pt.check_version("5.3"):
self.skipTest("Requires Terrier 5.3")
dataset = pt.datasets.get_dataset("vaswani")
indexref = dataset.get_index()
index = pt.IndexFactory.of(indexref)
# given their defaults, there three expressions are identical, all use Bo1
qe1 = pt.rewrite.QueryExpansion(index)
qe2 = pt.rewrite.DFRQueryExpansion(index)
qe3 = pt.rewrite.Bo1QueryExpansion(index)
# lets go faster, we only need 18 topics. qid 16 had a tricky case
t = dataset.get_topics().head(18)
qrels = dataset.get_qrels()
for qe in [qe1, qe2, qe3]:
br = pt.BatchRetrieve(index)
queriesIn = pd.DataFrame([["1", "compact"]],
columns=["qid", "query"])
res = br.transform(queriesIn)
queriesOut = qe.transform(res)
self.assertEqual(len(queriesOut), 1)
self.assertTrue("query_0" in queriesOut.columns)
self.assertEqual(queriesOut.iloc[0]["query_0"], "compact")
query = queriesOut.iloc[0]["query"]
self.assertTrue("compact^1.82230972" in query)
self.assertTrue("applypipeline:off " in query)
pipe = br >> qe >> br
# check the pipe doesnt cause an error
str(pipe)
all_qe_res = pipe.transform(t)
map_pipe = pt.Utils.evaluate(all_qe_res, qrels,
metrics=["map"])["map"]
br_qe = pt.BatchRetrieve(indexref, controls={"qe": "on"})
map_qe = pt.Utils.evaluate(br_qe.transform(t),
qrels,
metrics=["map"])["map"]
self.assertAlmostEqual(map_qe, map_pipe, places=4)
def test_save_trec_generator(self):
br = pt.BatchRetrieve(pt.get_dataset("vaswani").get_index(),
wmodel="TF_IDF")
filepath = os.path.join(self.test_dir, "test.res")
pt.io.write_results(br.transform_gen(
pt.get_dataset("vaswani").get_topics().head()),
filepath,
format="trec")
def test_parallel_joblib_experiment_br_callback(self):
self.skip_windows()
dataset = pt.get_dataset("vaswani")
Tf = lambda keyFreq, posting, entryStats, collStats: posting.getFrequency(
)
br = pt.BatchRetrieve(dataset.get_index(), wmodel=Tf)
df = pt.Experiment([br, br.parallel(3)], dataset.get_topics(),
dataset.get_qrels(), ["map", "mrt"])
self.assertEqual(df.iloc[0]["map"], df.iloc[1]["map"])
