utils.py 7.2 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

Xingjian Shi's avatar
Xingjian Shi committed
10
from omegaconf import OmegaConf
11
from jinja2 import BaseLoader, Environment, StrictUndefined
12
from itertools import islice
sdtblck's avatar
sdtblck committed
13
14
15
16
17
18
19
20
21
22

class ExitCodeError(Exception):
    pass


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


Jason Phang's avatar
gpt3  
Jason Phang committed
23
24
25
26
27
28
def simple_parse_args_string(args_string):
    """
    Parses something like
        args1=val1,arg2=val2
    Into a dictionary
    """
Jason Phang's avatar
Jason Phang committed
29
    args_string = args_string.strip()
Jason Phang's avatar
gpt3  
Jason Phang committed
30
31
32
    if not args_string:
        return {}
    arg_list = args_string.split(",")
Xingjian Shi's avatar
Xingjian Shi committed
33
    args_dict = OmegaConf.to_object(OmegaConf.from_dotlist(arg_list))
Jason Phang's avatar
gpt3  
Jason Phang committed
34
    return args_dict
Leo Gao's avatar
Leo Gao committed
35

Fabrizio Milo's avatar
Fabrizio Milo committed
36

Leo Gao's avatar
Leo Gao committed
37
38
def join_iters(iters):
    for iter in iters:
Leo Gao's avatar
Leo Gao committed
39
        yield from iter
Leo Gao's avatar
Leo Gao committed
40
41
42
43
44
45
46
47
48


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
49
50
51
52

    if arr:
        yield arr

Leo Gao's avatar
Leo Gao committed
53

54
55
56
57
58
def group(arr, fn):
    res = collections.defaultdict(list)

    for ob in arr:
        res[fn(ob)].append(ob)
Fabrizio Milo's avatar
Fabrizio Milo committed
59

60
61
    return list(res.values())

Fabrizio Milo's avatar
Fabrizio Milo committed
62

Leo Gao's avatar
Leo Gao committed
63
64
65
66
def general_detokenize(string):
    string = string.replace(" n't", "n't")
    string = string.replace(" )", ")")
    string = string.replace("( ", "(")
Fabrizio Milo's avatar
Fabrizio Milo committed
67
68
    string = string.replace('" ', '"')
    string = string.replace(' "', '"')
Leo Gao's avatar
Fix  
Leo Gao committed
69
    string = re.sub(r" (['.,])", r"\1", string)
70
71
72
    return string


Jason Phang's avatar
Jason Phang committed
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
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
100
    yield ([prefix_token] + token_list[: first_seq_len - 1], token_list[:first_seq_len])
Jason Phang's avatar
Jason Phang committed
101
102
103
104
105
    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
106

Jason Phang's avatar
Jason Phang committed
107
        yield (
Fabrizio Milo's avatar
Fabrizio Milo committed
108
109
            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
110
111
112
        )
        predicted += window_pred_len

Fabrizio Milo's avatar
Fabrizio Milo committed
113

Leo Gao's avatar
Leo Gao committed
114
def make_disjoint_window(pair):
Fabrizio Milo's avatar
Fabrizio Milo committed
115
    """Takes output from get_rolling_token_windows and makes the context not overlap with the continuation"""
Leo Gao's avatar
Leo Gao committed
116
    a, b = pair
117
    return a[: len(a) - (len(b) - 1)], b
Fabrizio Milo's avatar
Fabrizio Milo committed
118

Jason Phang's avatar
Jason Phang committed
119

120
121
122
123
124
class Reorderer:
    def __init__(self, arr, fn):
        self.size = len(arr)
        arr = list(enumerate(arr))
        arr = group(arr, lambda x: fn(x[1]))
125
126
127
        # 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]
128
129
130
        arr.sort(key=lambda x: fn(x[1]))

        self.arr = arr
Fabrizio Milo's avatar
Fabrizio Milo committed
131

132
133
    def get_reordered(self):
        return [x[1] for x in self.arr]
Fabrizio Milo's avatar
Fabrizio Milo committed
134

135
136
137
138
139
    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
140
            for ind in inds:
141
142
                res[ind] = v
                cov[ind] = True
Fabrizio Milo's avatar
Fabrizio Milo committed
143

144
        assert all(cov)
Fabrizio Milo's avatar
Fabrizio Milo committed
145

146
147
        return res

Fabrizio Milo's avatar
Fabrizio Milo committed
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

def make_table(result_dict):
    """Generate table of results."""
    from pytablewriter import MarkdownTableWriter, LatexTableWriter

    md_writer = MarkdownTableWriter()
    latex_writer = LatexTableWriter()
    md_writer.headers = ["Task", "Version", "Metric", "Value", "", "Stderr"]
    latex_writer.headers = ["Task", "Version", "Metric", "Value", "", "Stderr"]

    values = []

    for k, dic in result_dict["results"].items():
        version = result_dict["versions"][k]
        for m, v in dic.items():
            if m.endswith("_stderr"):
                continue

            if m + "_stderr" in dic:
                se = dic[m + "_stderr"]
                values.append([k, version, m, "%.4f" % v, "±", "%.4f" % se])
            else:
                values.append([k, version, m, "%.4f" % v, "", ""])
            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()


183
184
def positional_deprecated(fn):
    """
Fabrizio Milo's avatar
Fabrizio Milo committed
185
    A decorator to nudge users into passing only keyword args (`kwargs`) to the
186
187
    wrapped function, `fn`.
    """
Fabrizio Milo's avatar
Fabrizio Milo committed
188

189
190
    @functools.wraps(fn)
    def _wrapper(*args, **kwargs):
Fabrizio Milo's avatar
Fabrizio Milo committed
191
192
193
        if len(args) != 1 if inspect.ismethod(fn) else 0:
            print(
                f"WARNING: using {fn.__name__} with positional arguments is "
194
                "deprecated and will be disallowed in a future version of "
Fabrizio Milo's avatar
Fabrizio Milo committed
195
196
                "lm-evaluation-harness!"
            )
197
        return fn(*args, **kwargs)
Fabrizio Milo's avatar
Fabrizio Milo committed
198

199
    return _wrapper
Stephen Hogg's avatar
Stephen Hogg committed
200

Fabrizio Milo's avatar
Fabrizio Milo committed
201

Stephen Hogg's avatar
Stephen Hogg committed
202
203
204
205
206
207
208
209
210
@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
211
        if (cur_path / "tests" / "test_version_stable.py").exists():
Stephen Hogg's avatar
Stephen Hogg committed
212
213
214
            return cur_path
        else:
            cur_path = cur_path.parent.resolve()
Fabrizio Milo's avatar
Fabrizio Milo committed
215
216
217
218
    raise FileNotFoundError(
        f"Unable to find package root within {max_layers} upwards" + f"of {start_path}"
    )

Stephen Hogg's avatar
Stephen Hogg committed
219
220

@positional_deprecated
221
def run_task_tests(task_list: List[str]):
Stephen Hogg's avatar
Stephen Hogg committed
222
223
224
    """
    Find the package root and run the tests for the given tasks
    """
jon-tow's avatar
jon-tow committed
225
226
    import pytest

227
    package_root = find_test_root(start_path=pathlib.Path(__file__))
Fabrizio Milo's avatar
Fabrizio Milo committed
228
229
230
231
232
233
234
    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
235
236
237
    sys.path.append(str(package_root))
    pytest_return_val = pytest.main(args)
    if pytest_return_val:
Fabrizio Milo's avatar
Fabrizio Milo committed
238
239
240
        raise ValueError(
            f"Not all tests for the specified tasks ({task_list}) ran successfully! Error code: {pytest_return_val}"
        )
241
242


243
env = Environment(loader=BaseLoader, undefined=StrictUndefined)
244
245
246
247
248


def apply_template(template, doc):
    rtemplate = env.from_string(template)
    return rtemplate.render(**doc)
249
250
251
252
253
254
255
256
257


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 
    among ranks in multigpu setting or only pulling a sample of documents
    """
    return islice(raw_iterator, rank, limit, world_size)