utils.py 16.5 KB
Newer Older
sdtblck's avatar
sdtblck committed
1
import os
Leo Gao's avatar
Leo Gao committed
2
import re
Stephen Hogg's avatar
Stephen Hogg committed
3
import sys
4
5
6
7
8
9
import yaml
import inspect
import pathlib
import functools
import subprocess
import collections
lintangsutawika's avatar
lintangsutawika committed
10
import importlib.util
gakada's avatar
gakada committed
11
import fnmatch
12

13
14
from typing import List, Union

15
import gc
16
import torch
haileyschoelkopf's avatar
haileyschoelkopf committed
17
import transformers
sdtblck's avatar
sdtblck committed
18

Xingjian Shi's avatar
Xingjian Shi committed
19
from omegaconf import OmegaConf
20
from jinja2 import BaseLoader, Environment, StrictUndefined
21
from itertools import islice
haileyschoelkopf's avatar
haileyschoelkopf committed
22
23
24
25
26
<<<<<<< HEAD
=======

import transformers
>>>>>>> more pre-commit
sdtblck's avatar
sdtblck committed
27

28
from lm_eval.logger import eval_logger
sdtblck's avatar
sdtblck committed
29
30
31
32
33
34
35
36
37
38
39


class ExitCodeError(Exception):
    pass


def sh(x):
    if os.system(x):
        raise ExitCodeError()


40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
def escaped_split(text, sep_char, maxsplit=-1):
    """Split text into a list on occurrences of the given separation
    character `sep_char`. The separation character may be escaped by a
    backslash to avoid splitting at that location.

    The separation character must be a string of size 1.

    If `maxsplit` is given, at most `maxsplit` splits are done (thus,
    the list will have at most `maxsplit + 1` elements). If `maxsplit`
    is not specified or less than 0, then there is no limit on the
    number of splits (all possible splits are made).
    """
    assert (
        len(sep_char) == 1
    ), "separation string must be a single character for escaped splitting"

    if maxsplit == 0:
        return text
    maxsplit = max(0, maxsplit)

    return re.split(r"(?<!\\)" + sep_char, text, maxsplit)


Jason Phang's avatar
gpt3  
Jason Phang committed
63
64
65
66
67
68
def simple_parse_args_string(args_string):
    """
    Parses something like
        args1=val1,arg2=val2
    Into a dictionary
    """
Jason Phang's avatar
Jason Phang committed
69
    args_string = args_string.strip()
Jason Phang's avatar
gpt3  
Jason Phang committed
70
71
72
    if not args_string:
        return {}
    arg_list = args_string.split(",")
Xingjian Shi's avatar
Xingjian Shi committed
73
    args_dict = OmegaConf.to_object(OmegaConf.from_dotlist(arg_list))
Jason Phang's avatar
gpt3  
Jason Phang committed
74
    return args_dict
Leo Gao's avatar
Leo Gao committed
75

Fabrizio Milo's avatar
Fabrizio Milo committed
76

Leo Gao's avatar
Leo Gao committed
77
78
def join_iters(iters):
    for iter in iters:
Leo Gao's avatar
Leo Gao committed
79
        yield from iter
Leo Gao's avatar
Leo Gao committed
80
81


82
def chunks(iter, n=0, fn=None):
Leo Gao's avatar
Leo Gao committed
83
    arr = []
84
    for i, x in enumerate(iter):
Leo Gao's avatar
Leo Gao committed
85
        arr.append(x)
86
        if len(arr) == (fn(i) if fn else n):
Leo Gao's avatar
Leo Gao committed
87
88
            yield arr
            arr = []
Fabrizio Milo's avatar
Fabrizio Milo committed
89
90
91
92

    if arr:
        yield arr

Leo Gao's avatar
Leo Gao committed
93

94
95
96
97
98
def group(arr, fn):
    res = collections.defaultdict(list)

    for ob in arr:
        res[fn(ob)].append(ob)
Fabrizio Milo's avatar
Fabrizio Milo committed
99

100
101
    return list(res.values())

Fabrizio Milo's avatar
Fabrizio Milo committed
102

gakada's avatar
gakada committed
103
104
105
106
107
108
109
class MultiChoice:
    def __init__(self, choices):
        self.choices = choices

    # Simple wildcard support (linux filename patterns)
    def __contains__(self, values):
        for value in values.split(","):
110
111
112
113
114
            if len(fnmatch.filter(self.choices, value)) == 0:
                eval_logger.warning("{} is not in task list.".format(value))
                eval_logger.info(f"Available tasks to choose:")
                for choice in self.choices:
                    eval_logger.info(f"  - {choice}")
gakada's avatar
gakada committed
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
        return True

    def __iter__(self):
        for choice in self.choices:
            yield choice


# Returns a list containing all values of the source_list that
# match at least one of the patterns
def pattern_match(patterns, source_list):
    task_names = set()
    for pattern in patterns:
        for matching in fnmatch.filter(source_list, pattern):
            task_names.add(matching)
    return sorted(list(task_names))


Leo Gao's avatar
Leo Gao committed
132
133
134
135
def general_detokenize(string):
    string = string.replace(" n't", "n't")
    string = string.replace(" )", ")")
    string = string.replace("( ", "(")
Fabrizio Milo's avatar
Fabrizio Milo committed
136
137
    string = string.replace('" ', '"')
    string = string.replace(' "', '"')
Leo Gao's avatar
Fix  
Leo Gao committed
138
    string = re.sub(r" (['.,])", r"\1", string)
139
140
141
    return string


Jason Phang's avatar
Jason Phang committed
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
def get_rolling_token_windows(token_list, prefix_token, max_seq_len, context_len):
    """
    - context_len allows for a rolling window context, allowing each prediction window to potentially
      condition on some context

    :param token_list: list
        List of tokens to be PREDICTED
    :param max_seq_len: int
        max_seq_len of model (or max_seq_len we want to use)
    :param context_len: int
        Amount of desired token context for prediction. Needs to be at least 1.
    :param prefix_token: token
        Dummy token like <eos> so the first token has something to condition on
    :return: generator
        Generator of tuples
            (input_tokens, pred_tokens)
        Note: Score only the last len(pred_tokens) logits of the LM
    """
    assert 1 <= context_len <= max_seq_len
    if not token_list:
        return
    # +1 offset, going from input->preds
    pred_len = max_seq_len - context_len + 1
    predicted = 0

    # Special handling for first window: predict all tokens
    first_seq_len = min(max_seq_len, len(token_list))
Fabrizio Milo's avatar
Fabrizio Milo committed
169
    yield ([prefix_token] + token_list[: first_seq_len - 1], token_list[:first_seq_len])
Jason Phang's avatar
Jason Phang committed
170
171
172
173
174
    predicted += first_seq_len

    while predicted < len(token_list):
        window_pred_len = min(len(token_list) - predicted, pred_len)
        window_end = predicted + window_pred_len
Leo Gao's avatar
Leo Gao committed
175

Jason Phang's avatar
Jason Phang committed
176
        yield (
lintangsutawika's avatar
lintangsutawika committed
177
178
            token_list[window_end - max_seq_len - 1 : window_end - 1],
            token_list[window_end - window_pred_len : window_end],
Jason Phang's avatar
Jason Phang committed
179
180
181
        )
        predicted += window_pred_len

Fabrizio Milo's avatar
Fabrizio Milo committed
182

Leo Gao's avatar
Leo Gao committed
183
def make_disjoint_window(pair):
Fabrizio Milo's avatar
Fabrizio Milo committed
184
    """Takes output from get_rolling_token_windows and makes the context not overlap with the continuation"""
Leo Gao's avatar
Leo Gao committed
185
    a, b = pair
186
    return a[: len(a) - (len(b) - 1)], b
Fabrizio Milo's avatar
Fabrizio Milo committed
187

Jason Phang's avatar
Jason Phang committed
188

189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
def select_continuation_from_batch_left_padding(
    generations: Union[List[List[int]], torch.Tensor], max_context_size: int
):
    """Select the continuation from the batch, removing prompts of different lengths.
    Args:
        generations (Union[List[List[int]], torch.Tensor]):
            A tensor or list-of-lists of shape [batch_size, sequence length].
        max_context_size (int):
            The size of the biggest context; generations will proceed from that
            index.
    Example:
        PAD     PAD Continue : The dog chased the cat  [every       day of the week]
        Riddle  me    this   : The  dog chased the  cat [yesterday] PAD PAD PAD PAD
    Output:
        [every day of the week]
        [yesterday]  PAD PAD PAD PAD
    """
    return generations[:, max_context_size:]


209
210
211
212
213
class Reorderer:
    def __init__(self, arr, fn):
        self.size = len(arr)
        arr = list(enumerate(arr))
        arr = group(arr, lambda x: fn(x[1]))
214
215
216
        # arr = [([y[0] for y in x], x[0][1]) for x in arr]
        # TODO: overhaul reorderer. It currently grouped requests by content but we don't want this
        arr = [([y[0]], x[0][1]) for x in arr for y in x]
217
218
219
        arr.sort(key=lambda x: fn(x[1]))

        self.arr = arr
Fabrizio Milo's avatar
Fabrizio Milo committed
220

221
222
    def get_reordered(self):
        return [x[1] for x in self.arr]
Fabrizio Milo's avatar
Fabrizio Milo committed
223

224
225
226
227
228
    def get_original(self, newarr):
        res = [None] * self.size
        cov = [False] * self.size

        for (inds, _), v in zip(self.arr, newarr):
Fabrizio Milo's avatar
Fabrizio Milo committed
229
            for ind in inds:
230
231
                res[ind] = v
                cov[ind] = True
Fabrizio Milo's avatar
Fabrizio Milo committed
232

233
        assert all(cov)
Fabrizio Milo's avatar
Fabrizio Milo committed
234

235
236
        return res

Fabrizio Milo's avatar
Fabrizio Milo committed
237

238
239
240
241
242
243
def make_table(result_dict):
    """Generate table of results."""
    from pytablewriter import MarkdownTableWriter, LatexTableWriter

    md_writer = MarkdownTableWriter()
    latex_writer = LatexTableWriter()
244
    md_writer.headers = ["Task", "Version", "Filter", "Metric", "Value", "", "Stderr"]
lintangsutawika's avatar
lintangsutawika committed
245
246
247
248
249
250
251
252
253
    latex_writer.headers = [
        "Task",
        "Version",
        "Filter",
        "Metric",
        "Value",
        "",
        "Stderr",
    ]
254
255
256
257
258

    values = []

    for k, dic in result_dict["results"].items():
        version = result_dict["versions"][k]
259
260
        for (mf), v in dic.items():
            m, _, f = mf.partition(",")
261
262
263
            if m.endswith("_stderr"):
                continue

264
265
266
            if m + "_stderr" + "," + f in dic:
                se = dic[m + "_stderr" + "," + f]
                values.append([k, version, f, m, "%.4f" % v, "±", "%.4f" % se])
267
            else:
268
                values.append([k, version, f, m, "%.4f" % v, "", ""])
269
270
271
272
273
274
275
276
277
278
279
            k = ""
            version = ""
    md_writer.value_matrix = values
    latex_writer.value_matrix = values

    # todo: make latex table look good
    # print(latex_writer.dumps())

    return md_writer.dumps()


280
281
def positional_deprecated(fn):
    """
Fabrizio Milo's avatar
Fabrizio Milo committed
282
    A decorator to nudge users into passing only keyword args (`kwargs`) to the
283
284
    wrapped function, `fn`.
    """
Fabrizio Milo's avatar
Fabrizio Milo committed
285

286
287
    @functools.wraps(fn)
    def _wrapper(*args, **kwargs):
Fabrizio Milo's avatar
Fabrizio Milo committed
288
289
290
        if len(args) != 1 if inspect.ismethod(fn) else 0:
            print(
                f"WARNING: using {fn.__name__} with positional arguments is "
291
                "deprecated and will be disallowed in a future version of "
Fabrizio Milo's avatar
Fabrizio Milo committed
292
293
                "lm-evaluation-harness!"
            )
294
        return fn(*args, **kwargs)
Fabrizio Milo's avatar
Fabrizio Milo committed
295

296
    return _wrapper
Stephen Hogg's avatar
Stephen Hogg committed
297

Fabrizio Milo's avatar
Fabrizio Milo committed
298

Stephen Hogg's avatar
Stephen Hogg committed
299
300
301
302
303
304
305
306
307
@positional_deprecated
def find_test_root(start_path: pathlib.Path) -> pathlib.Path:
    """
    Search upward in the directory tree to a maximum of three layers
    to find and return the package root (containing the 'tests' folder)
    """
    cur_path = start_path.resolve()
    max_layers = 3
    for _ in range(max_layers):
Fabrizio Milo's avatar
Fabrizio Milo committed
308
        if (cur_path / "tests" / "test_version_stable.py").exists():
Stephen Hogg's avatar
Stephen Hogg committed
309
310
311
            return cur_path
        else:
            cur_path = cur_path.parent.resolve()
Fabrizio Milo's avatar
Fabrizio Milo committed
312
313
314
315
    raise FileNotFoundError(
        f"Unable to find package root within {max_layers} upwards" + f"of {start_path}"
    )

Stephen Hogg's avatar
Stephen Hogg committed
316
317

@positional_deprecated
318
def run_task_tests(task_list: List[str]):
Stephen Hogg's avatar
Stephen Hogg committed
319
320
321
    """
    Find the package root and run the tests for the given tasks
    """
jon-tow's avatar
jon-tow committed
322
323
    import pytest

324
    package_root = find_test_root(start_path=pathlib.Path(__file__))
Fabrizio Milo's avatar
Fabrizio Milo committed
325
326
327
328
329
330
331
    task_string = " or ".join(task_list)
    args = [
        f"{package_root}/tests/test_version_stable.py",
        f"--rootdir={package_root}",
        "-k",
        f"{task_string}",
    ]
Stephen Hogg's avatar
Stephen Hogg committed
332
333
334
    sys.path.append(str(package_root))
    pytest_return_val = pytest.main(args)
    if pytest_return_val:
Fabrizio Milo's avatar
Fabrizio Milo committed
335
336
337
        raise ValueError(
            f"Not all tests for the specified tasks ({task_list}) ran successfully! Error code: {pytest_return_val}"
        )
338
339


340
341
342
343
344
345
346
347
348
349
350
351
352
def get_git_commit_hash():
    """
    Gets the git commit hash of your current repo (if it exists).
    Source: https://github.com/EleutherAI/gpt-neox/blob/b608043be541602170bfcfb8ec9bf85e8a0799e0/megatron/neox_arguments/neox_args.py#L42
    """
    try:
        git_hash = subprocess.check_output(["git", "describe", "--always"]).strip()
        git_hash = git_hash.decode()
    except subprocess.CalledProcessError:
        git_hash = None
    return git_hash


lintangsutawika's avatar
lintangsutawika committed
353
354
355
356
357
def import_function(loader, node):

    function_name = loader.construct_scalar(node)
    yaml_path = os.path.dirname(loader.name)

lintangsutawika's avatar
lintangsutawika committed
358
    module_name, function_name = function_name.split(".")
lintangsutawika's avatar
lintangsutawika committed
359
360
361
362
363
364
365
366
367
    module_path = os.path.join(yaml_path, "{}.py".format(module_name))

    spec = importlib.util.spec_from_file_location(module_name, module_path)
    module = importlib.util.module_from_spec(spec)
    spec.loader.exec_module(module)

    function = getattr(module, function_name)
    return function

lintangsutawika's avatar
lintangsutawika committed
368

lintangsutawika's avatar
lintangsutawika committed
369
# Add the import_function constructor to the YAML loader
lintangsutawika's avatar
lintangsutawika committed
370
yaml.add_constructor("!function", import_function)
lintangsutawika's avatar
lintangsutawika committed
371
372
373


def load_yaml_config(yaml_path):
lintangsutawika's avatar
lintangsutawika committed
374
    with open(yaml_path, "rb") as file:
lintangsutawika's avatar
lintangsutawika committed
375
376
        yaml_config = yaml.full_load(file)
        yaml_dir = os.path.dirname(yaml_path)
lintangsutawika's avatar
lintangsutawika committed
377
378
379
380

        if "include" in yaml_config:
            include_path = yaml_config["include"]
            del yaml_config["include"]
lintangsutawika's avatar
lintangsutawika committed
381
382
383

            if type(include_path) == str:
                include_path = [include_path]
lintangsutawika's avatar
lintangsutawika committed
384

lintangsutawika's avatar
lintangsutawika committed
385
386
387
388
389
390
            # Load from the last one first
            include_path.reverse()
            final_yaml_config = {}
            for path in include_path:

                # Assumes that path is a full path.
lintangsutawika's avatar
lintangsutawika committed
391
                # If not found, assume the included yaml
lintangsutawika's avatar
lintangsutawika committed
392
393
394
395
396
397
398
                # is in the same dir as the original yaml
                if not os.path.isfile(path):
                    path = os.path.join(yaml_dir, path)

                try:
                    included_yaml_config = load_yaml_config(path)
                    final_yaml_config.update(included_yaml_config)
lintangsutawika's avatar
lintangsutawika committed
399
                except Exception as ex:
lintangsutawika's avatar
lintangsutawika committed
400
                    # If failed to load, ignore
lintangsutawika's avatar
lintangsutawika committed
401
                    raise ex
lintangsutawika's avatar
lintangsutawika committed
402
403
404
405
406
407

            final_yaml_config.update(yaml_config)
            return final_yaml_config
        return yaml_config


408
env = Environment(loader=BaseLoader, undefined=StrictUndefined)
409
410
411
412
413


def apply_template(template, doc):
    rtemplate = env.from_string(template)
    return rtemplate.render(**doc)
414
415


416
417
418
419
def create_iterator(raw_iterator, rank, world_size, limit=None):
    """
    Method for creating a (potentially) sliced and limited
    iterator from a raw document iterator. Used for splitting data
420
421
422
    among ranks in multigpu setting or only pulling a sample of documents
    """
    return islice(raw_iterator, rank, limit, world_size)
423
424


haileyschoelkopf's avatar
haileyschoelkopf committed
425
<<<<<<< HEAD
426
427
428
def clear_torch_cache():
    gc.collect()
    torch.cuda.empty_cache()
haileyschoelkopf's avatar
haileyschoelkopf committed
429
430


lintangsutawika's avatar
lintangsutawika committed
431
def get_dtype(dtype: Union[str, torch.dtype]) -> torch.dtype:
haileyschoelkopf's avatar
haileyschoelkopf committed
432
433
434
435
436
437
    """Converts `dtype` from `str` to torch.dtype when possible. Does not use an instantiated HF AutoConfig"""
    if isinstance(dtype, str) and dtype != "auto":
        # Convert `str` args torch dtype: `float16` -> `torch.float16`
        _torch_dtype = getattr(torch, dtype)
    else:
        _torch_dtype = dtype
lintangsutawika's avatar
lintangsutawika committed
438
    return _torch_dtype
lintangsutawika's avatar
lintangsutawika committed
439
440


haileyschoelkopf's avatar
haileyschoelkopf committed
441
442
443
444
445
def pad_and_concat(max_length:int, tensors: List[torch.Tensor], padding_side="right"):
    """ 
    Method for padding a list of tensors given the maximum tensor 
    length in the batch. Used for batching inputs and continuations in 
    seq2seq models. 
haileyschoelkopf's avatar
haileyschoelkopf committed
446
447
448
449
450
451
452
=======
def pad_and_concat(max_length: int, tensors: List[torch.Tensor], padding_side="right"):
    """
    Method for padding a list of tensors given the maximum tensor
    length in the batch. Used for batching inputs and continuations in
    seq2seq models.
>>>>>>> more pre-commit
lintangsutawika's avatar
lintangsutawika committed
453
    """
haileyschoelkopf's avatar
haileyschoelkopf committed
454
455
456
    assert (
        padding_side == "left" or padding_side == "right"
    ), f"Unrecognized padding type: '{padding_side}' not 'left' or 'right'"
haileyschoelkopf's avatar
haileyschoelkopf committed
457

lintangsutawika's avatar
lintangsutawika committed
458
459
460
    for i, tensor in enumerate(tensors):
        tensor_len = tensor.shape[0]
        if tensor_len < max_length:
haileyschoelkopf's avatar
haileyschoelkopf committed
461
462
463
            if padding_side == "right":
                # right-pad
                tensors[i] = torch.cat(
haileyschoelkopf's avatar
haileyschoelkopf committed
464
465
466
467
468
469
470
471
472
473
                    [
                        tensor,  # [seq]
                        torch.zeros(
                            max_length - tensor_len,
                            dtype=torch.long,
                            device=tensor.device,
                        ),  # [padding_length - seq]
                    ],
                    dim=0,
                ).unsqueeze(0)
haileyschoelkopf's avatar
haileyschoelkopf committed
474
475
476
477
            else:
                # left-pad
                tensors[i] = torch.cat(
                    [
478
                        torch.zeros(
haileyschoelkopf's avatar
haileyschoelkopf committed
479
                            max_length - tensor_len,
480
481
                            dtype=torch.long,
                            device=tensor.device,
haileyschoelkopf's avatar
haileyschoelkopf committed
482
                        ),  # [padding_length - seq]
haileyschoelkopf's avatar
haileyschoelkopf committed
483
                        tensor,  # [seq]
haileyschoelkopf's avatar
haileyschoelkopf committed
484
485
486
                    ],
                    dim=0,
                ).unsqueeze(0)
lintangsutawika's avatar
lintangsutawika committed
487
488
489
        else:
            tensors[i] = tensor.unsqueeze(0)

haileyschoelkopf's avatar
haileyschoelkopf committed
490
    return torch.cat(tensors, dim=0)
haileyschoelkopf's avatar
haileyschoelkopf committed
491
492


haileyschoelkopf's avatar
haileyschoelkopf committed
493
<<<<<<< HEAD
haileyschoelkopf's avatar
haileyschoelkopf committed
494
# Multi-token stopping criteria 
haileyschoelkopf's avatar
haileyschoelkopf committed
495
496
497
498
499
500
501
502
=======
def clear_torch_cache():
    gc.collect()
    torch.cuda.empty_cache()


# Multi-token stopping criteria
>>>>>>> more pre-commit
haileyschoelkopf's avatar
haileyschoelkopf committed
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
class MultiTokenEOSCriteria(transformers.StoppingCriteria):
    """Criteria to stop on the specified multi-token sequence."""

    def __init__(
        self,
        sequence: str,
        tokenizer: transformers.PreTrainedTokenizer,
        initial_decoder_input_length: int,
        batch_size: int,
    ):
        self.initial_decoder_input_length = initial_decoder_input_length
        self.done_tracker = [False] * batch_size
        self.sequence = sequence
        self.sequence_ids = tokenizer.encode(sequence, add_special_tokens=False)
        self.sequence_id_len = len(self.sequence_ids)
        self.tokenizer = tokenizer

    def __call__(self, input_ids, scores, **kwargs) -> bool:
        # For efficiency, we compare the last n tokens where n is the number of tokens in the stop_sequence
        lookback_ids_batch = input_ids[:, self.initial_decoder_input_length :][
            :, -self.sequence_id_len :
        ]

        lookback_tokens_batch = self.tokenizer.batch_decode(lookback_ids_batch)

        for i, done in enumerate(self.done_tracker):
            if not done:
                self.done_tracker[i] = self.sequence in lookback_tokens_batch[i]
        return False not in self.done_tracker


def stop_sequences_criteria(
    tokenizer: transformers.PreTrainedTokenizer,
    stop_sequences: List[str],
    initial_decoder_input_length: int,
    batch_size: int,
) -> transformers.StoppingCriteriaList:
    return transformers.StoppingCriteriaList(
        [
            *[
                MultiTokenEOSCriteria(
                    sequence, tokenizer, initial_decoder_input_length, batch_size
                )
                for sequence in stop_sequences
            ],
        ]
    )