File size: 20,938 Bytes
62977bb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
#
# Pyserini: Reproducible IR research with sparse and dense representations
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#

import argparse
import math
import os
import sys
import time
import subprocess
import pkg_resources
from collections import defaultdict, OrderedDict
from string import Template

import yaml

from ._base import run_eval_and_return_metric, ok_str, okish_str, fail_str

languages = [
    ['ar', 'arabic'],
    ['bn', 'bengali'],
    ['en', 'english'],
    ['es', 'spanish'],
    ['fa', 'persian'],
    ['fi', 'finnish'],
    ['fr', 'french'],
    ['hi', 'hindi'],
    ['id', 'indonesian'],
    ['ja', 'japanese'],
    ['ko', 'korean'],
    ['ru', 'russian'],
    ['sw', 'swahili'],
    ['te', 'telugu'],
    ['th', 'thai'],
    ['zh', 'chinese'],
    ['de', 'german'],
    ['yo', 'yoruba']
]

html_display = OrderedDict()
html_display['bm25'] = 'BM25'
html_display['mdpr-tied-pft-msmarco'] = 'mDPR (tied encoders), pre-FT w/ MS MARCO'
html_display['mdpr-tied-pft-msmarco-ft-all'] = 'mDPR (tied encoders), pre-FT w/ MS MARCO then FT w/ all Mr. TyDi'
html_display['bm25-mdpr-tied-pft-msmarco-hybrid'] = 'Hybrid of `bm25` and `mdpr-tied-pft-msmarco`'
html_display['mdpr-tied-pft-msmarco-ft-miracl'] = 'mDPR (tied encoders), pre-FT w/ MS MARCO then in-lang FT w/ MIRACL'
html_display['mcontriever-tied-pft-msmarco'] = 'mContriever (tied encoders), pre-FT w/ MS MARCO'

models = list(html_display)

trec_eval_metric_definitions = {
    'nDCG@10': '-c -M 100 -m ndcg_cut.10',
    'R@100': '-c -m recall.100',
}


def format_run_command(raw):
    return raw.replace('--lang', '\\\n  --lang') \
        .replace('--encoder', '\\\n  --encoder') \
        .replace('--topics', '\\\n  --topics') \
        .replace('--index', '\\\n  --index') \
        .replace('--output ', '\\\n  --output ') \
        .replace('--runs', '\\\n  --runs ') \
        .replace('--batch ', '\\\n  --batch ') \
        .replace('--threads 12', '--threads 12 \\\n ')


def format_eval_command(raw):
    return raw.replace('-c ', '\\\n  -c ') \
        .replace(raw.split()[-1], f'\\\n  {raw.split()[-1]}')


def read_file(f):
    fin = open(f, 'r')
    text = fin.read()
    fin.close()

    return text


def list_conditions():
    print('Conditions:\n-----------')
    for condition, _ in html_display.items():
        print(condition)
    print('\nLanguages\n---------')
    for language in languages:
        print(language[0])


def generate_table_rows(table, row_template, commands, eval_commands, table_id, split, metric):
    row_cnt = 1
    html_rows = []

    for model in models:
        s = Template(row_template)

        keys = {}
        used_langs = 0
        for lang in languages:
            keys[lang[0]] = f'{model}.{lang[0]}'
            used_langs += 1 if table[keys[lang[0]]][split][metric] != 0 else 0

        sum = table[keys["ar"]][split][metric] + \
              table[keys["bn"]][split][metric] + \
              table[keys["en"]][split][metric] + \
              table[keys["es"]][split][metric] + \
              table[keys["fa"]][split][metric] + \
              table[keys["fi"]][split][metric] + \
              table[keys["fr"]][split][metric] + \
              table[keys["hi"]][split][metric] + \
              table[keys["id"]][split][metric] + \
              table[keys["ja"]][split][metric] + \
              table[keys["ko"]][split][metric] + \
              table[keys["ru"]][split][metric] + \
              table[keys["sw"]][split][metric] + \
              table[keys["te"]][split][metric] + \
              table[keys["th"]][split][metric] + \
              table[keys["zh"]][split][metric] + \
              table[keys["de"]][split][metric] + \
              table[keys["yo"]][split][metric]
        avg = sum / used_langs

        s = s.substitute(table_cnt=table_id,
                         row_cnt=row_cnt,
                         model=html_display[model],
                         ar=f'{table[keys["ar"]][split][metric]:.3f}',
                         bn=f'{table[keys["bn"]][split][metric]:.3f}',
                         en=f'{table[keys["en"]][split][metric]:.3f}',
                         es=f'{table[keys["es"]][split][metric]:.3f}',
                         fa=f'{table[keys["fa"]][split][metric]:.3f}',
                         fi=f'{table[keys["fi"]][split][metric]:.3f}',
                         fr=f'{table[keys["fr"]][split][metric]:.3f}',
                         hi=f'{table[keys["hi"]][split][metric]:.3f}',
                         id=f'{table[keys["id"]][split][metric]:.3f}',
                         ja=f'{table[keys["ja"]][split][metric]:.3f}',
                         ko=f'{table[keys["ko"]][split][metric]:.3f}',
                         ru=f'{table[keys["ru"]][split][metric]:.3f}',
                         sw=f'{table[keys["sw"]][split][metric]:.3f}',
                         te=f'{table[keys["te"]][split][metric]:.3f}',
                         th=f'{table[keys["th"]][split][metric]:.3f}',
                         zh=f'{table[keys["zh"]][split][metric]:.3f}',
                         de=f'{table[keys["de"]][split][metric]:.3f}',
                         yo=f'{table[keys["yo"]][split][metric]:.3f}',
                         avg=f'{avg:.3f}',
                         cmd1=f'{commands[keys["ar"]]}',
                         cmd2=f'{commands[keys["bn"]]}',
                         cmd3=f'{commands[keys["en"]]}',
                         cmd4=f'{commands[keys["es"]]}',
                         cmd5=f'{commands[keys["fa"]]}',
                         cmd6=f'{commands[keys["fi"]]}',
                         cmd7=f'{commands[keys["fr"]]}',
                         cmd8=f'{commands[keys["hi"]]}',
                         cmd9=f'{commands[keys["id"]]}',
                         cmd10=f'{commands[keys["ja"]]}',
                         cmd11=f'{commands[keys["ko"]]}',
                         cmd12=f'{commands[keys["ru"]]}',
                         cmd13=f'{commands[keys["sw"]]}',
                         cmd14=f'{commands[keys["te"]]}',
                         cmd15=f'{commands[keys["th"]]}',
                         cmd16=f'{commands[keys["zh"]]}',
                         cmd17=f'{commands[keys["de"]]}',
                         cmd18=f'{commands[keys["yo"]]}',
                         eval_cmd1=f'{eval_commands[keys["ar"]][metric]}',
                         eval_cmd2=f'{eval_commands[keys["bn"]][metric]}',
                         eval_cmd3=f'{eval_commands[keys["en"]][metric]}',
                         eval_cmd4=f'{eval_commands[keys["es"]][metric]}',
                         eval_cmd5=f'{eval_commands[keys["fa"]][metric]}',
                         eval_cmd6=f'{eval_commands[keys["fi"]][metric]}',
                         eval_cmd7=f'{eval_commands[keys["fr"]][metric]}',
                         eval_cmd8=f'{eval_commands[keys["hi"]][metric]}',
                         eval_cmd9=f'{eval_commands[keys["id"]][metric]}',
                         eval_cmd10=f'{eval_commands[keys["ja"]][metric]}',
                         eval_cmd11=f'{eval_commands[keys["ko"]][metric]}',
                         eval_cmd12=f'{eval_commands[keys["ru"]][metric]}',
                         eval_cmd13=f'{eval_commands[keys["sw"]][metric]}',
                         eval_cmd14=f'{eval_commands[keys["te"]][metric]}',
                         eval_cmd15=f'{eval_commands[keys["th"]][metric]}',
                         eval_cmd16=f'{eval_commands[keys["zh"]][metric]}',
                         eval_cmd17=f'{eval_commands[keys["de"]][metric]}',
                         eval_cmd18=f'{eval_commands[keys["yo"]][metric]}'
                         )

        s = s.replace("0.000", "--")
        html_rows.append(s)
        row_cnt += 1

    return html_rows


def print_results(table, metric, split):
    print(f'Metric = {metric}, Split = {split}')
    print(' ' * 35, end='')
    for lang in languages:
        print(f'{lang[0]:3}    ', end='')
    print('')
    for model in models:
        print(f'{model:33}', end='')
        for lang in languages:
            key = f'{model}.{lang[0]}'
            print(f'{table[key][split][metric]:7.3f}', end='')
        print('')
    print('')


def extract_topic_fn_from_cmd(cmd):
    cmd = cmd.split()
    topic_idx = cmd.index('--topics')
    return cmd[topic_idx + 1]


def generate_report(args):
    table = defaultdict(lambda: defaultdict(lambda: defaultdict(lambda: 0.0)))
    commands = defaultdict(lambda: '')
    eval_commands = defaultdict(lambda: defaultdict(lambda: ''))

    html_template = read_file(pkg_resources.resource_filename(__name__, 'miracl_html.template'))
    table_template = read_file(pkg_resources.resource_filename(__name__, 'miracl_html_table.template'))
    row_template = read_file(pkg_resources.resource_filename(__name__, 'miracl_html_table_row.template'))

    with open(pkg_resources.resource_filename(__name__, 'miracl.yaml')) as f:
        yaml_data = yaml.safe_load(f)
        for condition in yaml_data['conditions']:
            name = condition['name']
            eval_key = condition['eval_key']
            cmd_template = condition['command']
            cmd_lst = cmd_template.split()
            lang = name.split('.')[-1]
            is_hybrid_run = 'hybrid' in name

            for splits in condition['splits']:
                split = splits['split']
                if is_hybrid_run:
                    hits = int(cmd_lst[cmd_lst.index('--k') + 1])
                else:
                    hits = int(cmd_lst[cmd_lst.index('--hits') + 1])

                runfile = os.path.join(args.directory, f'run.miracl.{name}.{split}.txt')
                if is_hybrid_run:
                    bm25_output = os.path.join(args.directory,
                                               f'run.miracl.bm25.{lang}.{split}.top{hits}.txt')
                    mdpr_output = os.path.join(args.directory,
                                               f'run.miracl.mdpr-tied-pft-msmarco.{lang}.{split}.top{hits}.txt')
                    expected_args = dict(output=runfile, bm25_output=bm25_output, mdpr_output=mdpr_output)
                else:
                    expected_args = dict(split=split, output=runfile)

                if not all([f"${k}" in cmd_template or f"${{{k}}}" in cmd_template for k in expected_args]):
                    raise ValueError(f"Not all arguements {list(expected_args)} detected from inputs: {cmd_template}.")
                cmd = Template(cmd_template).substitute(**expected_args)
                commands[name] = format_run_command(cmd)

                for expected in splits['scores']:
                    for metric in expected:
                        if str(expected[metric])[-1] == "5":
                            # without adding espilon, there is a chance that f-string would round 0.5 to 0 rather than 1
                            # e.g., 0.8885 -> 0.888 rather than 0.889
                            # add a espilon to the expected score to avoid rounding error
                            expected[metric] += 1e-5
                        table[name][split][metric] = expected[metric]

                        eval_cmd = f'python -m pyserini.eval.trec_eval ' + \
                                   f'{trec_eval_metric_definitions[metric]} {eval_key}-{split} {runfile}'
                        eval_commands[name][metric] = format_eval_command(eval_cmd)

        tables_html = []

        split = 'dev'

        # Build the table for MRR@100, test queries
        html_rows = generate_table_rows(table, row_template, commands, eval_commands, 1, split, 'nDCG@10')
        all_rows = '\n'.join(html_rows)
        tables_html.append(Template(table_template).substitute(desc=f'nDCG@10, {split} queries', rows=all_rows))

        # Build the table for R@100, test queries
        html_rows = generate_table_rows(table, row_template, commands, eval_commands, 2, split, 'R@100')
        all_rows = '\n'.join(html_rows)
        tables_html.append(Template(table_template).substitute(desc=f'Recall@100, {split} queries', rows=all_rows))

    with open(args.output, 'w') as out:
        out.write(Template(html_template).substitute(title='MIRACL', tables=' '.join(tables_html)))


def run_conditions(args):
    if args.condition == 'mdpr-tied-pft-msmarco-ft-miracl' and args.language in ['de', 'yo']:
        print('MIRACL de and yo datasets do not have train splits to finetune with')
        return

    start = time.time()

    table = defaultdict(lambda: defaultdict(lambda: defaultdict(lambda: 0.0)))

    with open(pkg_resources.resource_filename(__name__, 'miracl.yaml')) as f:
        yaml_data = yaml.safe_load(f)
        for condition in yaml_data['conditions']:
            name = condition['name']
            encoder = name.split('.')[0]
            lang = name.split('.')[-1]
            if args.all:
                pass
            elif args.condition != encoder:
                continue
            elif args.language and args.language != lang:
                continue
            eval_key = condition['eval_key']
            cmd_template = condition['command']
            cmd_lst = cmd_template.split()

            print(f'condition {name}:')
            is_hybrid_run = 'hybrid' in name

            for splits in condition['splits']:
                split = splits['split']
                if is_hybrid_run:
                    hits = int(cmd_lst[cmd_lst.index('--k') + 1])
                else:
                    hits = int(cmd_lst[cmd_lst.index('--hits') + 1])

                print(f'  - split: {split}')

                runfile = os.path.join(args.directory, f'run.miracl.{name}.{split}.top{hits}.txt')
                if is_hybrid_run:
                    bm25_output = os.path.join(args.directory,
                                               f'run.miracl.bm25.{lang}.{split}.top{hits}.txt')
                    mdpr_output = os.path.join(args.directory,
                                               f'run.miracl.mdpr-tied-pft-msmarco.{lang}.{split}.top{hits}.txt')
                    if not os.path.exists(bm25_output):
                        print(f'Missing BM25 file: {bm25_output}')
                        continue
                    if not os.path.exists(mdpr_output):
                        print(f'Missing mDPR file: {mdpr_output}')
                        continue
                    cmd = Template(cmd_template).substitute(split=split, output=runfile, bm25_output=bm25_output,
                                                            mdpr_output=mdpr_output)
                else:
                    cmd = Template(cmd_template).substitute(split=split, output=runfile)

                # In the yaml file, the topics are written as something like '--topics miracl-v1.0-ar-${split}'
                # This works for the dev split because the topics are directly included in Anserini/Pyserini.
                # For this training split, we have to map the symbol into a file in tools/topics-and-qrels/
                # Here, we assume that the developer has cloned the miracl repo and placed the topics there.
                if split == 'train':
                    cmd = cmd.replace(f'--topics miracl-v1.0-{lang}-{split}',
                                      f'--topics tools/topics-and-qrels/topics.miracl-v1.0-{lang}-{split}.tsv')

                if args.display_commands:
                    print(f'\n```bash\n{format_run_command(cmd)}\n```\n')

                if not os.path.exists(runfile):
                    if not args.dry_run:
                        rtn = subprocess.run(cmd.split(), capture_output=True)
                        stderr = rtn.stderr.decode()
                        if '--topics' in cmd:
                            topic_fn = extract_topic_fn_from_cmd(cmd)
                            if f'ValueError: Topic {topic_fn} Not Found' in stderr:
                                print(f'Skipping {topic_fn}: file not found.')
                                continue

                for expected in splits['scores']:
                    for metric in expected:
                        if not args.skip_eval:
                            # We have the translate the training qrels into a file located in tools/topics-and-qrels/
                            # because they are not included with Anserini/Pyserini by default.
                            # Here, we assume that the developer has cloned the miracl repo and placed the qrels there.
                            if split == 'train':
                                qrels = f'tools/topics-and-qrels/qrels.{eval_key}-train.tsv'
                            else:
                                qrels = f'{eval_key}-{split}'
                            score = float(run_eval_and_return_metric(metric, qrels,
                                                                     trec_eval_metric_definitions[metric], runfile))
                            if math.isclose(score, float(expected[metric])):
                                result_str = ok_str
                            # Flaky tests
                            elif (name == 'mdpr-tied-pft-msmarco.hi' and split == 'train'
                                  and math.isclose(score, float(expected[metric]), abs_tol=2e-4)) or \
                                 (name == 'mdpr-tied-pft-msmarco-ft-all.ru'
                                  and split == 'dev' and metric == 'nDCG@10'
                                  and math.isclose(score, float(expected[metric]), abs_tol=2e-4)) or \
                                 (name == 'bm25-mdpr-tied-pft-msmarco-hybrid.te'
                                  and split == 'train' and metric == 'nDCG@10'
                                  and math.isclose(score, float(expected[metric]), abs_tol=2e-4)) or \
                                 (name == 'bm25-mdpr-tied-pft-msmarco-hybrid.zh'
                                  and split == 'dev' and metric == 'nDCG@10'
                                  and math.isclose(score, float(expected[metric]), abs_tol=2e-4)):
                                result_str = okish_str
                            else:
                                result_str = fail_str + f' expected {expected[metric]:.4f}'
                            print(f'      {metric:7}: {score:.4f} {result_str}')
                            table[name][split][metric] = score
                        else:
                            table[name][split][metric] = expected[metric]

            print('')

    for metric in ['nDCG@10', 'R@100']:
        for split in ['dev', 'train']:
            print_results(table, metric, split)

    end = time.time()
    print(f'Total elapsed time: {end - start:.0f}s')


if __name__ == '__main__':
    parser = argparse.ArgumentParser(description='Generate regression matrix for MIRACL.')
    parser.add_argument('--condition', type=str,
                        help='Condition to run', required=False)
    # To list all conditions
    parser.add_argument('--list-conditions', action='store_true', default=False, help='List available conditions.')
    # For generating reports
    parser.add_argument('--generate-report', action='store_true', default=False, help='Generate report.')
    parser.add_argument('--output', type=str, help='File to store report.', required=False)
    # For actually running the experimental conditions
    parser.add_argument('--all', action='store_true', default=False, help='Run using all languages.')
    parser.add_argument('--language', type=str, help='Language to run.', required=False)
    parser.add_argument('--directory', type=str, help='Base directory.', default='', required=False)
    parser.add_argument('--dry-run', action='store_true', default=False, help='Print out commands but do not execute.')
    parser.add_argument('--skip-eval', action='store_true', default=False, help='Skip running trec_eval.')
    parser.add_argument('--display-commands', action='store_true', default=False, help='Display command.')
    args = parser.parse_args()

    if args.list_conditions:
        list_conditions()
        sys.exit()

    if args.generate_report:
        if not args.output:
            print(f'Must specify report filename with --output.')
            sys.exit()

        generate_report(args)
        sys.exit()

    if args.all and (args.condition or args.language):
        print('Specifying --all will run all conditions and languages')
        sys.exit()

    run_conditions(args)