-
Notifications
You must be signed in to change notification settings - Fork 1
/
run_msmarco_test.py
133 lines (104 loc) · 5.08 KB
/
run_msmarco_test.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
import polars as pl
import numpy as np
import json
import torch
from torch.utils.data import Dataset, DataLoader
import tw_bert_v2
from pathlib import Path
from collections import Counter
from nltk.corpus import stopwords
from nltk.stem import PorterStemmer
from joblib import Parallel, delayed
stemmer = PorterStemmer()
sw = list(stopwords.words('english'))
def clean_and_count(pid, passage):
s = [stemmer.stem(w) for w in passage if w not in sw]
c = Counter(s)
return (pid,c)
def clean_query(s):
s = [stemmer.stem(w) for w in s if w not in sw]
return s
def query_tf_vec(s):
c = Counter(s)
s = [c[w] for w in s]
return s
class MSMARCOData(Dataset):
def __init__(self, queries):
self.queries = queries
def __len__(self):
return len(self.queries)
def __getitem__(self, index):
qid = self.queries[index]
query = qid_map[qid]
query_tf_vec = query_tf_vecs_map[qid]
corpus = [doc_wf[d] for d in query_doc_map[qid]]
targets = query_labels_map[qid]
return query, query_tf_vec, corpus, targets
if __name__ == "__main__":
torch.set_default_device('cuda')
path = "path to msmarco"
qrels_dev = pl.read_csv(path / "collectionandqueries" / "qrels.dev.small.tsv", separator='\t', has_header=False,new_columns=["qid", "i", "pid", "label"])
top1000 = pl.read_csv(path / "top1000.dev" /"top1000.dev", separator='\t', has_header=False, new_columns=["qid", "pid", "query", "passage"])
data = top1000.join(qrels_dev[['qid', 'pid','label']], on=['qid', 'pid'], how='left').fill_null(0).sort('label', descending=True).group_by('qid').head(100)
queries = data.select(pl.col("qid")).unique().to_numpy().squeeze()
train_queries = queries[:-500]
val_queries = queries[-500:]
query_tf_vecs = data.select(pl.col("qid"), pl.col('query')).unique().with_columns([
pl.col("query").str.extract_all(r"[A-Za-z0-9']+").alias("query_terms")]).with_columns([
pl.col("query_terms").map_elements(clean_query).alias("clean_query_terms")]).with_columns([
pl.col("clean_query_terms").map_elements(query_tf_vec).alias("query_tf_vec"),
pl.col("clean_query_terms").list.join(" ").alias("clean_query")])
doc_lists = data.select(pl.col("pid"), pl.col('passage')).unique().with_columns([
pl.col("passage").str.to_lowercase().str.extract_all(r"[A-Za-z0-9']+").alias("passage_terms")]).select(pl.col("pid", "passage_terms"))
# parallelize the loop using joblib
doc_wf = dict(Parallel(n_jobs=-1)(delayed(clean_and_count)(doc[0], doc[1]) for doc in doc_lists.rows()))
query_doc_map = {row[0]: row[1] for row in data.select(pl.col("qid"), pl.col("pid")).group_by("qid").agg(pl.col("pid")).rows()}
query_labels_map = {row[0]: row[1] for row in data.select(pl.col("qid"), pl.col("label")).group_by("qid").agg(pl.col("label")).rows()}
qid_map = {row[0]: row[1] for row in query_tf_vecs.select(pl.col("qid"), pl.col("clean_query")).unique().rows()}
query_tf_vecs_map = {row[0]: row[1] for row in query_tf_vecs.select(pl.col("qid"), pl.col("query_tf_vec")).rows()}
train_dataset = MSMARCOData(train_queries)
val_dataset = MSMARCOData(val_queries)
model = tw_bert_v2.TWBERT().cuda()
optimizer = torch.optim.AdamW(model.parameters(), lr=1e-5)
criterion = tw_bert_v2.TWBERTLossFT()
accum_iter = 500
for i in range(10):
for j in range(len(train_dataset)):
query, query_tf_vec, corpus, target = train_dataset[j]
query_tf_vec = torch.tensor(query_tf_vec, dtype=torch.float32, device="cuda")
avg_doc_len = 500
if sum(target) == 0 or len(corpus) == 0:
continue
query_t, mask = tw_bert_v2.token_and_mask_query(query, tw_bert_v2.tokenizer)
term_weights = model(query_t["input_ids"], query_t["attention_mask"], mask).squeeze()[1:-1]
output = tw_bert_v2.score_vec(query, query_tf_vec, corpus, term_weights, avg_doc_len)
target = torch.tensor(target, dtype=torch.float32, device="cuda")
loss = criterion(output, target)
loss = loss / accum_iter
loss.backward()
if ((j + 1) % accum_iter == 0) or (j + 1 == len(data)):
print("Step: ", j + 1, "Train Loss: ", loss.item())
optimizer.step()
optimizer.zero_grad()
with torch.no_grad():
model.eval()
loss_ = 0
mrr = 0
for k in range(len(val_dataset)):
query, query_tf_vec, corpus, target = val_dataset[k]
query_tf_vec = torch.tensor(query_tf_vec, dtype=torch.float32, device="cuda")
avg_doc_len = 500
if sum(target) == 0 or len(corpus) == 0:
continue
query_t, mask = tw_bert_v2.token_and_mask_query(query, tw_bert_v2.tokenizer)
term_weights = model(query_t["input_ids"], query_t["attention_mask"], mask).squeeze()[1:-1]
#term_weights = torch.ones(term_weights.shape, device="cuda") # check non-optimized metrics
output = tw_bert_v2.score_vec(query, query_tf_vec, corpus, term_weights, avg_doc_len)
target = torch.tensor(target, dtype=torch.float32, device="cuda")
if term_weights.sum() != 0:
mrr += 1 / (torch.nonzero(target[output.sort(descending=True).indices])[0].item()+1)
else:
mrr += 0.
loss_ += criterion(output, target).item()
print("Val Loss: ", loss_/ len(val_dataset), "Val MRR: ", mrr/ len(val_dataset))
model.train()