utils.py 6.22 KB
Newer Older
sdtblck's avatar
sdtblck committed
1
import os
Stephen Hogg's avatar
Stephen Hogg committed
2
import pathlib
Leo Gao's avatar
Leo Gao committed
3
import re
4
import collections
5
import functools
Leo Gao's avatar
Leo Gao committed
6
import inspect
Stephen Hogg's avatar
Stephen Hogg committed
7
8
import sys
from typing import List
sdtblck's avatar
sdtblck committed
9
10
11
12
13
14
15
16
17
18
19


class ExitCodeError(Exception):
    pass


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


20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
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
43
44
45
46
47
48
def simple_parse_args_string(args_string):
    """
    Parses something like
        args1=val1,arg2=val2
    Into a dictionary
    """
Jason Phang's avatar
Jason Phang committed
49
    args_string = args_string.strip()
Jason Phang's avatar
gpt3  
Jason Phang committed
50
51
52
53
    if not args_string:
        return {}
    arg_list = args_string.split(",")
    args_dict = {}
Jason Phang's avatar
Jason Phang committed
54
    for arg in arg_list:
Jason Phang's avatar
gpt3  
Jason Phang committed
55
56
57
        k, v = arg.split("=")
        args_dict[k] = v
    return args_dict
Leo Gao's avatar
Leo Gao committed
58

Fabrizio Milo's avatar
Fabrizio Milo committed
59

Leo Gao's avatar
Leo Gao committed
60
61
def join_iters(iters):
    for iter in iters:
Leo Gao's avatar
Leo Gao committed
62
        yield from iter
Leo Gao's avatar
Leo Gao committed
63
64
65
66
67
68
69
70
71


def chunks(iter, n):
    arr = []
    for x in iter:
        arr.append(x)
        if len(arr) == n:
            yield arr
            arr = []
Fabrizio Milo's avatar
Fabrizio Milo committed
72
73
74
75

    if arr:
        yield arr

Leo Gao's avatar
Leo Gao committed
76

77
78
79
80
81
def group(arr, fn):
    res = collections.defaultdict(list)

    for ob in arr:
        res[fn(ob)].append(ob)
Fabrizio Milo's avatar
Fabrizio Milo committed
82

83
84
    return list(res.values())

Fabrizio Milo's avatar
Fabrizio Milo committed
85

Leo Gao's avatar
Leo Gao committed
86
87
88
89
def general_detokenize(string):
    string = string.replace(" n't", "n't")
    string = string.replace(" )", ")")
    string = string.replace("( ", "(")
Fabrizio Milo's avatar
Fabrizio Milo committed
90
91
    string = string.replace('" ', '"')
    string = string.replace(' "', '"')
Leo Gao's avatar
Fix  
Leo Gao committed
92
    string = re.sub(r" (['.,])", r"\1", string)
93
94
95
    return string


Jason Phang's avatar
Jason Phang committed
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
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
123
    yield ([prefix_token] + token_list[: first_seq_len - 1], token_list[:first_seq_len])
Jason Phang's avatar
Jason Phang committed
124
125
126
127
128
    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
129

Jason Phang's avatar
Jason Phang committed
130
        yield (
Fabrizio Milo's avatar
Fabrizio Milo committed
131
132
            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
133
134
135
        )
        predicted += window_pred_len

Fabrizio Milo's avatar
Fabrizio Milo committed
136

Leo Gao's avatar
Leo Gao committed
137
def make_disjoint_window(pair):
Fabrizio Milo's avatar
Fabrizio Milo committed
138
    """Takes output from get_rolling_token_windows and makes the context not overlap with the continuation"""
Leo Gao's avatar
Leo Gao committed
139
    a, b = pair
140
    return a[: len(a) - (len(b) - 1)], b
Fabrizio Milo's avatar
Fabrizio Milo committed
141

Jason Phang's avatar
Jason Phang committed
142

143
144
145
146
147
class Reorderer:
    def __init__(self, arr, fn):
        self.size = len(arr)
        arr = list(enumerate(arr))
        arr = group(arr, lambda x: fn(x[1]))
Fabrizio Milo's avatar
Fabrizio Milo committed
148
        arr = [([y[0] for y in x], x[0][1]) for x in arr]
149
150
151
        arr.sort(key=lambda x: fn(x[1]))

        self.arr = arr
Fabrizio Milo's avatar
Fabrizio Milo committed
152

153
154
    def get_reordered(self):
        return [x[1] for x in self.arr]
Fabrizio Milo's avatar
Fabrizio Milo committed
155

156
157
158
159
160
    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
161
            for ind in inds:
162
163
                res[ind] = v
                cov[ind] = True
Fabrizio Milo's avatar
Fabrizio Milo committed
164

165
        assert all(cov)
Fabrizio Milo's avatar
Fabrizio Milo committed
166

167
168
        return res

Fabrizio Milo's avatar
Fabrizio Milo committed
169

170
171
def positional_deprecated(fn):
    """
Fabrizio Milo's avatar
Fabrizio Milo committed
172
    A decorator to nudge users into passing only keyword args (`kwargs`) to the
173
174
    wrapped function, `fn`.
    """
Fabrizio Milo's avatar
Fabrizio Milo committed
175

176
177
    @functools.wraps(fn)
    def _wrapper(*args, **kwargs):
Fabrizio Milo's avatar
Fabrizio Milo committed
178
179
180
        if len(args) != 1 if inspect.ismethod(fn) else 0:
            print(
                f"WARNING: using {fn.__name__} with positional arguments is "
181
                "deprecated and will be disallowed in a future version of "
Fabrizio Milo's avatar
Fabrizio Milo committed
182
183
                "lm-evaluation-harness!"
            )
184
        return fn(*args, **kwargs)
Fabrizio Milo's avatar
Fabrizio Milo committed
185

186
    return _wrapper
Stephen Hogg's avatar
Stephen Hogg committed
187

Fabrizio Milo's avatar
Fabrizio Milo committed
188

Stephen Hogg's avatar
Stephen Hogg committed
189
190
191
192
193
194
195
196
197
@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
198
        if (cur_path / "tests" / "test_version_stable.py").exists():
Stephen Hogg's avatar
Stephen Hogg committed
199
200
201
            return cur_path
        else:
            cur_path = cur_path.parent.resolve()
Fabrizio Milo's avatar
Fabrizio Milo committed
202
203
204
205
    raise FileNotFoundError(
        f"Unable to find package root within {max_layers} upwards" + f"of {start_path}"
    )

Stephen Hogg's avatar
Stephen Hogg committed
206
207

@positional_deprecated
208
def run_task_tests(task_list: List[str]):
Stephen Hogg's avatar
Stephen Hogg committed
209
210
211
    """
    Find the package root and run the tests for the given tasks
    """
jon-tow's avatar
jon-tow committed
212
213
    import pytest

214
    package_root = find_test_root(start_path=pathlib.Path(__file__))
Fabrizio Milo's avatar
Fabrizio Milo committed
215
216
217
218
219
220
221
    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
222
223
224
    sys.path.append(str(package_root))
    pytest_return_val = pytest.main(args)
    if pytest_return_val:
Fabrizio Milo's avatar
Fabrizio Milo committed
225
226
227
        raise ValueError(
            f"Not all tests for the specified tasks ({task_list}) ran successfully! Error code: {pytest_return_val}"
        )