utils.py 15.4 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
sdtblck's avatar
sdtblck committed
22

23
from lm_eval.logger import eval_logger
sdtblck's avatar
sdtblck committed
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
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
49
50
51
52
53
54
def simple_parse_args_string(args_string):
    """
    Parses something like
        args1=val1,arg2=val2
    Into a dictionary
    """
Jason Phang's avatar
Jason Phang committed
55
    args_string = args_string.strip()
Jason Phang's avatar
gpt3  
Jason Phang committed
56
57
58
    if not args_string:
        return {}
    arg_list = args_string.split(",")
Xingjian Shi's avatar
Xingjian Shi committed
59
    args_dict = OmegaConf.to_object(OmegaConf.from_dotlist(arg_list))
Jason Phang's avatar
gpt3  
Jason Phang committed
60
    return args_dict
Leo Gao's avatar
Leo Gao committed
61

Fabrizio Milo's avatar
Fabrizio Milo committed
62

Leo Gao's avatar
Leo Gao committed
63
64
def join_iters(iters):
    for iter in iters:
Leo Gao's avatar
Leo Gao committed
65
        yield from iter
Leo Gao's avatar
Leo Gao committed
66
67


68
def chunks(iter, n=0, fn=None):
Leo Gao's avatar
Leo Gao committed
69
    arr = []
70
    for i, x in enumerate(iter):
Leo Gao's avatar
Leo Gao committed
71
        arr.append(x)
72
        if len(arr) == (fn(i) if fn else n):
Leo Gao's avatar
Leo Gao committed
73
74
            yield arr
            arr = []
Fabrizio Milo's avatar
Fabrizio Milo committed
75
76
77
78

    if arr:
        yield arr

Leo Gao's avatar
Leo Gao committed
79

80
81
82
83
84
def group(arr, fn):
    res = collections.defaultdict(list)

    for ob in arr:
        res[fn(ob)].append(ob)
Fabrizio Milo's avatar
Fabrizio Milo committed
85

86
87
    return list(res.values())

Fabrizio Milo's avatar
Fabrizio Milo committed
88

gakada's avatar
gakada committed
89
90
91
92
93
94
95
class MultiChoice:
    def __init__(self, choices):
        self.choices = choices

    # Simple wildcard support (linux filename patterns)
    def __contains__(self, values):
        for value in values.split(","):
96
97
98
99
100
            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
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
        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
118
119
120
121
def general_detokenize(string):
    string = string.replace(" n't", "n't")
    string = string.replace(" )", ")")
    string = string.replace("( ", "(")
Fabrizio Milo's avatar
Fabrizio Milo committed
122
123
    string = string.replace('" ', '"')
    string = string.replace(' "', '"')
Leo Gao's avatar
Fix  
Leo Gao committed
124
    string = re.sub(r" (['.,])", r"\1", string)
125
126
127
    return string


Jason Phang's avatar
Jason Phang committed
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
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
155
    yield ([prefix_token] + token_list[: first_seq_len - 1], token_list[:first_seq_len])
Jason Phang's avatar
Jason Phang committed
156
157
158
159
160
    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
161

Jason Phang's avatar
Jason Phang committed
162
        yield (
lintangsutawika's avatar
lintangsutawika committed
163
164
            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
165
166
167
        )
        predicted += window_pred_len

Fabrizio Milo's avatar
Fabrizio Milo committed
168

Leo Gao's avatar
Leo Gao committed
169
def make_disjoint_window(pair):
Fabrizio Milo's avatar
Fabrizio Milo committed
170
    """Takes output from get_rolling_token_windows and makes the context not overlap with the continuation"""
Leo Gao's avatar
Leo Gao committed
171
    a, b = pair
172
    return a[: len(a) - (len(b) - 1)], b
Fabrizio Milo's avatar
Fabrizio Milo committed
173

Jason Phang's avatar
Jason Phang committed
174

175
176
177
178
179
class Reorderer:
    def __init__(self, arr, fn):
        self.size = len(arr)
        arr = list(enumerate(arr))
        arr = group(arr, lambda x: fn(x[1]))
180
181
182
        # 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]
183
184
185
        arr.sort(key=lambda x: fn(x[1]))

        self.arr = arr
Fabrizio Milo's avatar
Fabrizio Milo committed
186

187
188
    def get_reordered(self):
        return [x[1] for x in self.arr]
Fabrizio Milo's avatar
Fabrizio Milo committed
189

190
191
192
193
194
    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
195
            for ind in inds:
196
197
                res[ind] = v
                cov[ind] = True
Fabrizio Milo's avatar
Fabrizio Milo committed
198

199
        assert all(cov)
Fabrizio Milo's avatar
Fabrizio Milo committed
200

201
202
        return res

Fabrizio Milo's avatar
Fabrizio Milo committed
203

204
205
206
207
208
209
def make_table(result_dict):
    """Generate table of results."""
    from pytablewriter import MarkdownTableWriter, LatexTableWriter

    md_writer = MarkdownTableWriter()
    latex_writer = LatexTableWriter()
210
    md_writer.headers = ["Task", "Version", "Filter", "Metric", "Value", "", "Stderr"]
lintangsutawika's avatar
lintangsutawika committed
211
212
213
214
215
216
217
218
219
    latex_writer.headers = [
        "Task",
        "Version",
        "Filter",
        "Metric",
        "Value",
        "",
        "Stderr",
    ]
220
221
222
223
224

    values = []

    for k, dic in result_dict["results"].items():
        version = result_dict["versions"][k]
225
226
        for (mf), v in dic.items():
            m, _, f = mf.partition(",")
227
228
229
            if m.endswith("_stderr"):
                continue

230
231
232
            if m + "_stderr" + "," + f in dic:
                se = dic[m + "_stderr" + "," + f]
                values.append([k, version, f, m, "%.4f" % v, "±", "%.4f" % se])
233
            else:
234
                values.append([k, version, f, m, "%.4f" % v, "", ""])
235
236
237
238
239
240
241
242
243
244
245
            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()


246
247
def positional_deprecated(fn):
    """
Fabrizio Milo's avatar
Fabrizio Milo committed
248
    A decorator to nudge users into passing only keyword args (`kwargs`) to the
249
250
    wrapped function, `fn`.
    """
Fabrizio Milo's avatar
Fabrizio Milo committed
251

252
253
    @functools.wraps(fn)
    def _wrapper(*args, **kwargs):
Fabrizio Milo's avatar
Fabrizio Milo committed
254
255
256
        if len(args) != 1 if inspect.ismethod(fn) else 0:
            print(
                f"WARNING: using {fn.__name__} with positional arguments is "
257
                "deprecated and will be disallowed in a future version of "
Fabrizio Milo's avatar
Fabrizio Milo committed
258
259
                "lm-evaluation-harness!"
            )
260
        return fn(*args, **kwargs)
Fabrizio Milo's avatar
Fabrizio Milo committed
261

262
    return _wrapper
Stephen Hogg's avatar
Stephen Hogg committed
263

Fabrizio Milo's avatar
Fabrizio Milo committed
264

Stephen Hogg's avatar
Stephen Hogg committed
265
266
267
268
269
270
271
272
273
@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
274
        if (cur_path / "tests" / "test_version_stable.py").exists():
Stephen Hogg's avatar
Stephen Hogg committed
275
276
277
            return cur_path
        else:
            cur_path = cur_path.parent.resolve()
Fabrizio Milo's avatar
Fabrizio Milo committed
278
279
280
281
    raise FileNotFoundError(
        f"Unable to find package root within {max_layers} upwards" + f"of {start_path}"
    )

Stephen Hogg's avatar
Stephen Hogg committed
282
283

@positional_deprecated
284
def run_task_tests(task_list: List[str]):
Stephen Hogg's avatar
Stephen Hogg committed
285
286
287
    """
    Find the package root and run the tests for the given tasks
    """
jon-tow's avatar
jon-tow committed
288
289
    import pytest

290
    package_root = find_test_root(start_path=pathlib.Path(__file__))
Fabrizio Milo's avatar
Fabrizio Milo committed
291
292
293
294
295
296
297
    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
298
299
300
    sys.path.append(str(package_root))
    pytest_return_val = pytest.main(args)
    if pytest_return_val:
Fabrizio Milo's avatar
Fabrizio Milo committed
301
302
303
        raise ValueError(
            f"Not all tests for the specified tasks ({task_list}) ran successfully! Error code: {pytest_return_val}"
        )
304
305


306
307
308
309
310
311
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:
312
        git_hash = subprocess.check_output(["gt", "describe", "--always"]).strip()
313
        git_hash = git_hash.decode()
314
315
    except subprocess.CalledProcessError or FileNotFoundError:
        # FileNotFoundError occurs when git not installed on system
316
317
318
319
        git_hash = None
    return git_hash


lintangsutawika's avatar
lintangsutawika committed
320
321
322
323
324
def import_function(loader, node):

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

lintangsutawika's avatar
lintangsutawika committed
325
    module_name, function_name = function_name.split(".")
lintangsutawika's avatar
lintangsutawika committed
326
327
328
329
330
331
332
333
334
    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
335

lintangsutawika's avatar
lintangsutawika committed
336
# Add the import_function constructor to the YAML loader
lintangsutawika's avatar
lintangsutawika committed
337
yaml.add_constructor("!function", import_function)
lintangsutawika's avatar
lintangsutawika committed
338
339
340


def load_yaml_config(yaml_path):
lintangsutawika's avatar
lintangsutawika committed
341
    with open(yaml_path, "rb") as file:
lintangsutawika's avatar
lintangsutawika committed
342
343
        yaml_config = yaml.full_load(file)
        yaml_dir = os.path.dirname(yaml_path)
lintangsutawika's avatar
lintangsutawika committed
344
345
346
347

        if "include" in yaml_config:
            include_path = yaml_config["include"]
            del yaml_config["include"]
lintangsutawika's avatar
lintangsutawika committed
348
349
350

            if type(include_path) == str:
                include_path = [include_path]
lintangsutawika's avatar
lintangsutawika committed
351

lintangsutawika's avatar
lintangsutawika committed
352
353
354
355
356
357
            # 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
358
                # If not found, assume the included yaml
lintangsutawika's avatar
lintangsutawika committed
359
360
361
362
363
364
365
                # 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
366
                except Exception as ex:
lintangsutawika's avatar
lintangsutawika committed
367
                    # If failed to load, ignore
lintangsutawika's avatar
lintangsutawika committed
368
                    raise ex
lintangsutawika's avatar
lintangsutawika committed
369
370
371
372
373
374

            final_yaml_config.update(yaml_config)
            return final_yaml_config
        return yaml_config


375
376
377
def regex_replace(string, pattern, repl, count=0):
    """Implements the `re.sub` function as a custom Jinja filter."""
    return re.sub(pattern, repl, string, count=count)
lintangsutawika's avatar
lintangsutawika committed
378

lintangsutawika's avatar
lintangsutawika committed
379

380
env = Environment(loader=BaseLoader, undefined=StrictUndefined)
381
env.filters["regex_replace"] = regex_replace
382
383
384
385
386


def apply_template(template, doc):
    rtemplate = env.from_string(template)
    return rtemplate.render(**doc)
387
388


389
390
391
392
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
393
394
395
    among ranks in multigpu setting or only pulling a sample of documents
    """
    return islice(raw_iterator, rank, limit, world_size)
396
397


haileyschoelkopf's avatar
haileyschoelkopf committed
398
399
400
401
402
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.
lintangsutawika's avatar
lintangsutawika committed
403
    """
haileyschoelkopf's avatar
haileyschoelkopf committed
404
405
406
    assert (
        padding_side == "left" or padding_side == "right"
    ), f"Unrecognized padding type: '{padding_side}' not 'left' or 'right'"
haileyschoelkopf's avatar
haileyschoelkopf committed
407

lintangsutawika's avatar
lintangsutawika committed
408
409
410
    for i, tensor in enumerate(tensors):
        tensor_len = tensor.shape[0]
        if tensor_len < max_length:
haileyschoelkopf's avatar
haileyschoelkopf committed
411
412
413
            if padding_side == "right":
                # right-pad
                tensors[i] = torch.cat(
haileyschoelkopf's avatar
haileyschoelkopf committed
414
415
416
417
418
419
420
421
422
423
                    [
                        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
424
425
426
427
            else:
                # left-pad
                tensors[i] = torch.cat(
                    [
428
                        torch.zeros(
haileyschoelkopf's avatar
haileyschoelkopf committed
429
                            max_length - tensor_len,
430
431
                            dtype=torch.long,
                            device=tensor.device,
haileyschoelkopf's avatar
haileyschoelkopf committed
432
                        ),  # [padding_length - seq]
haileyschoelkopf's avatar
haileyschoelkopf committed
433
                        tensor,  # [seq]
haileyschoelkopf's avatar
haileyschoelkopf committed
434
435
436
                    ],
                    dim=0,
                ).unsqueeze(0)
lintangsutawika's avatar
lintangsutawika committed
437
438
439
        else:
            tensors[i] = tensor.unsqueeze(0)

haileyschoelkopf's avatar
haileyschoelkopf committed
440
    return torch.cat(tensors, dim=0)
haileyschoelkopf's avatar
haileyschoelkopf committed
441
442


443
444
445
def clear_torch_cache():
    gc.collect()
    torch.cuda.empty_cache()
haileyschoelkopf's avatar
haileyschoelkopf committed
446
447


lintangsutawika's avatar
lintangsutawika committed
448
449
450
451
452
453
454
455
456
457
def get_dtype(dtype: Union[str, torch.dtype]) -> torch.dtype:
    """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
    return _torch_dtype


haileyschoelkopf's avatar
haileyschoelkopf committed
458
# Multi-token stopping criteria
haileyschoelkopf's avatar
haileyschoelkopf committed
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
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
            ],
        ]
    )