File size: 10,351 Bytes
2c3c6f5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239

import sys
import json
from torch.utils.data import DataLoader
from sentence_transformers import SentenceTransformer, LoggingHandler, util, models, evaluation, losses, InputExample
import logging
from datetime import datetime
import gzip
import os
import tarfile
from collections import defaultdict
from torch.utils.data import IterableDataset
import tqdm
from torch.utils.data import Dataset
import random
from shutil import copyfile

import argparse

parser = argparse.ArgumentParser()
parser.add_argument("--train_batch_size", default=64, type=int)
parser.add_argument("--max_seq_length", default=300, type=int)
parser.add_argument("--model_name", required=True)
parser.add_argument("--max_passages", default=0, type=int)
parser.add_argument("--epochs", default=10, type=int)
parser.add_argument("--pooling", default="cls")
parser.add_argument("--negs_to_use", default=None, help="From which systems should negatives be used? Multiple systems seperated by comma. None = all")
parser.add_argument("--warmup_steps", default=1000, type=int)
parser.add_argument("--lr", default=2e-5, type=float)
parser.add_argument("--name", default='')
parser.add_argument("--num_negs_per_system", default=5, type=int) 
parser.add_argument("--use_pre_trained_model", default=False, action="store_true") 
parser.add_argument("--use_all_queries", default=False, action="store_true") 
args = parser.parse_args()

print(args)

#### Just some code to print debug information to stdout
logging.basicConfig(format='%(asctime)s - %(message)s',
                    datefmt='%Y-%m-%d %H:%M:%S',
                    level=logging.INFO,
                    handlers=[LoggingHandler()])
#### /print debug information to stdout

# The  model we want to fine-tune
train_batch_size = args.train_batch_size          #Increasing the train batch size improves the model performance, but requires more GPU memory
model_name = args.model_name
max_passages = args.max_passages
max_seq_length = args.max_seq_length            #Max length for passages. Increasing it, requires more GPU memory

num_negs_per_system = args.num_negs_per_system  # We used different systems to mine hard negatives. Number of hard negatives to add from each system
num_epochs = args.epochs         # Number of epochs we want to train

# We construct the SentenceTransformer bi-encoder from scratch
if args.use_pre_trained_model:
    print("use pretrained SBERT model")
    model = SentenceTransformer(model_name)
    model.max_seq_length = max_seq_length
else:
    print("Create new SBERT model")
    word_embedding_model = models.Transformer(model_name, max_seq_length=max_seq_length)
    pooling_model = models.Pooling(word_embedding_model.get_word_embedding_dimension(), args.pooling)
    model = SentenceTransformer(modules=[word_embedding_model, pooling_model])

model_save_path = f'output/train_bi-encoder-margin_mse_en-{args.name}-{model_name.replace("/", "-")}-batch_size_{train_batch_size}-{datetime.now().strftime("%Y-%m-%d_%H-%M-%S")}'


# Write self to path
os.makedirs(model_save_path, exist_ok=True)

train_script_path = os.path.join(model_save_path, 'train_script.py')
copyfile(__file__, train_script_path)
with open(train_script_path, 'a') as fOut:
    fOut.write("\n\n# Script was called via:\n#python " + " ".join(sys.argv))


### Now we read the MS Marco dataset
data_folder = 'msmarco-data'

#### Read the corpus files, that contain all the passages. Store them in the corpus dict
corpus = {}         #dict in the format: passage_id -> passage. Stores all existent passages
collection_filepath = os.path.join(data_folder, 'collection.tsv')
if not os.path.exists(collection_filepath):
    tar_filepath = os.path.join(data_folder, 'collection.tar.gz')
    if not os.path.exists(tar_filepath):
        logging.info("Download collection.tar.gz")
        util.http_get('https://msmarco.blob.core.windows.net/msmarcoranking/collection.tar.gz', tar_filepath)

    with tarfile.open(tar_filepath, "r:gz") as tar:
        tar.extractall(path=data_folder)

logging.info("Read corpus: collection.tsv")
with open(collection_filepath, 'r', encoding='utf8') as fIn:
    for line in fIn:
        pid, passage = line.strip().split("\t")
        corpus[pid] = passage


### Read the train queries, store in queries dict
queries = {}        #dict in the format: query_id -> query. Stores all training queries
queries_filepath = os.path.join(data_folder, 'queries.train.tsv')
if not os.path.exists(queries_filepath):
    tar_filepath = os.path.join(data_folder, 'queries.tar.gz')
    if not os.path.exists(tar_filepath):
        logging.info("Download queries.tar.gz")
        util.http_get('https://msmarco.blob.core.windows.net/msmarcoranking/queries.tar.gz', tar_filepath)

    with tarfile.open(tar_filepath, "r:gz") as tar:
        tar.extractall(path=data_folder)


with open(queries_filepath, 'r', encoding='utf8') as fIn:
    for line in fIn:
        qid, query = line.strip().split("\t")
        queries[qid] = query


# Read our training file: msmarco-hard-negatives.jsonl.gz contains all queries and hard-negatives that were mined with different systems
# For each positive and mined-hard negative passage, we have a Cross-Encoder score from the cross-encoder/ms-marco-MiniLM-L-6-v2 model
# This  Cross-Encoder score allows to de-noise our hard-negatives by requiring that their CE-score is below a certain treshold
train_filepath = '/home/msmarco/data/hard-negatives/msmarco-hard-negatives-v6.jsonl.gz'

#### Create our training data
logging.info("Read train dataset")
train_queries = {}
ce_scores = {}
negs_to_use = None
with gzip.open(train_filepath, 'rt') as fIn:
    for line in tqdm.tqdm(fIn):
        if max_passages > 0 and len(train_queries) >= max_passages:
            break
            
        data = json.loads(line)
        
        if data['qid'] not in ce_scores:
            ce_scores[data['qid']] = {}
        
        # Add pos ce_scores
        for item in data['pos'] :
            ce_scores[data['qid']][item['pid']] = item['ce-score']

        #Get the positive passage ids
        pos_pids = [item['pid'] for item in data['pos']]
       
        #Get the hard negatives
        neg_pids = set()
        if negs_to_use is None:
            if args.negs_to_use is not None:    #Use specific system for negatives
                negs_to_use = args.negs_to_use.split(",")
            else:   #Use all systems
                negs_to_use = list(data['neg'].keys())
            print("Using negatives from the following systems:", negs_to_use)
            
        for system_name in negs_to_use:
            if system_name not in data['neg']:
                continue
                
            system_negs = data['neg'][system_name]
            
            negs_added = 0
            for item in system_negs:
                #Add neg ce_scores
                ce_scores[data['qid']][item['pid']] = item['ce-score']
                
                pid = item['pid']
                if pid not in neg_pids:
                    neg_pids.add(pid)
                    negs_added += 1
                    if negs_added >= num_negs_per_system:
                        break

        if args.use_all_queries or (len(pos_pids) > 0 and len(neg_pids) > 0):
            train_queries[data['qid']] = {'qid': data['qid'], 'query': queries[data['qid']], 'pos': pos_pids, 'neg': neg_pids}

logging.info("Train queries: {}".format(len(train_queries)))

# We create a custom MSMARCO dataset that returns triplets (query, positive, negative)
# on-the-fly based on the information from the mined-hard-negatives jsonl file.
class MSMARCODataset(Dataset):
    def __init__(self, queries, corpus, ce_scores):
        self.queries = queries
        self.queries_ids = list(queries.keys())
        self.corpus = corpus
        self.ce_scores = ce_scores

        for qid in self.queries:
            self.queries[qid]['pos'] = list(self.queries[qid]['pos'])
            self.queries[qid]['neg'] = list(self.queries[qid]['neg'])
            random.shuffle(self.queries[qid]['neg'])

    def __getitem__(self, item):
        query = self.queries[self.queries_ids[item]]
        query_text = query['query']
        qid = query['qid']

        if len(query['pos']) > 0:
            pos_id = query['pos'].pop(0)    #Pop positive and add at end
            pos_text = self.corpus[pos_id]
            query['pos'].append(pos_id)
        else:   #We only have negatives, use two negs
            pos_id = query['neg'].pop(0)    #Pop negative and add at end
            pos_text = self.corpus[pos_id]
            query['neg'].append(pos_id)

        #Get a negative passage
        neg_id = query['neg'].pop(0)    #Pop negative and add at end
        neg_text = self.corpus[neg_id]
        query['neg'].append(neg_id)
        
        pos_score = self.ce_scores[qid][pos_id]
        neg_score = self.ce_scores[qid][neg_id]
        
        return InputExample(texts=[query_text, pos_text, neg_text], label=pos_score-neg_score)

    def __len__(self):
        return len(self.queries)

# For training the SentenceTransformer model, we need a dataset, a dataloader, and a loss used for training.
train_dataset = MSMARCODataset(queries=train_queries, corpus=corpus, ce_scores=ce_scores)
train_dataloader = DataLoader(train_dataset, shuffle=True, batch_size=train_batch_size, drop_last=True)
train_loss = losses.MarginMSELoss(model=model)

# Train the model
model.fit(train_objectives=[(train_dataloader, train_loss)],
          epochs=num_epochs,
          warmup_steps=args.warmup_steps,
          use_amp=True,
          checkpoint_path=model_save_path,
          checkpoint_save_steps=10000,
          checkpoint_save_total_limit = 0,
          optimizer_params = {'lr': args.lr},
          )

# Train latest model
model.save(model_save_path)


# Script was called via:
#python train_bi-encoder-margin_mse-en.py --model final-models/co-condenser-margin_mse-sym_mnrl-mean-v1 --lr=1e-5 --warmup_steps=10000 --negs_to_use=co-condenser-margin_mse-sym_mnrl-mean-v1 --num_negs_per_system=10 --epochs=30 --name=cnt_with_mined_negs_mean --use_pre_trained_model --train_batch_size 32