utils.py 9.09 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
11

Stephen Hogg's avatar
Stephen Hogg committed
12
from typing import List
sdtblck's avatar
sdtblck committed
13

Xingjian Shi's avatar
Xingjian Shi committed
14
from omegaconf import OmegaConf
15
from jinja2 import BaseLoader, Environment, StrictUndefined
Xingjian Shi's avatar
Xingjian Shi committed
16

sdtblck's avatar
sdtblck committed
17
18
19
20
21
22
23
24
25
26

class ExitCodeError(Exception):
    pass


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


Jason Phang's avatar
gpt3  
Jason Phang committed
27
28
29
30
31
32
def simple_parse_args_string(args_string):
    """
    Parses something like
        args1=val1,arg2=val2
    Into a dictionary
    """
Jason Phang's avatar
Jason Phang committed
33
    args_string = args_string.strip()
Jason Phang's avatar
gpt3  
Jason Phang committed
34
35
36
    if not args_string:
        return {}
    arg_list = args_string.split(",")
Xingjian Shi's avatar
Xingjian Shi committed
37
    args_dict = OmegaConf.to_object(OmegaConf.from_dotlist(arg_list))
Jason Phang's avatar
gpt3  
Jason Phang committed
38
    return args_dict
Leo Gao's avatar
Leo Gao committed
39

Fabrizio Milo's avatar
Fabrizio Milo committed
40

Leo Gao's avatar
Leo Gao committed
41
42
def join_iters(iters):
    for iter in iters:
Leo Gao's avatar
Leo Gao committed
43
        yield from iter
Leo Gao's avatar
Leo Gao committed
44
45
46
47
48
49
50
51
52


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
53
54
55
56

    if arr:
        yield arr

Leo Gao's avatar
Leo Gao committed
57

58
59
60
61
62
def group(arr, fn):
    res = collections.defaultdict(list)

    for ob in arr:
        res[fn(ob)].append(ob)
Fabrizio Milo's avatar
Fabrizio Milo committed
63

64
65
    return list(res.values())

Fabrizio Milo's avatar
Fabrizio Milo committed
66

Leo Gao's avatar
Leo Gao committed
67
68
69
70
def general_detokenize(string):
    string = string.replace(" n't", "n't")
    string = string.replace(" )", ")")
    string = string.replace("( ", "(")
Fabrizio Milo's avatar
Fabrizio Milo committed
71
72
    string = string.replace('" ', '"')
    string = string.replace(' "', '"')
Leo Gao's avatar
Fix  
Leo Gao committed
73
    string = re.sub(r" (['.,])", r"\1", string)
74
75
76
    return string


Jason Phang's avatar
Jason Phang committed
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
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
104
    yield ([prefix_token] + token_list[: first_seq_len - 1], token_list[:first_seq_len])
Jason Phang's avatar
Jason Phang committed
105
106
107
108
109
    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
110

Jason Phang's avatar
Jason Phang committed
111
        yield (
Fabrizio Milo's avatar
Fabrizio Milo committed
112
113
            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
114
115
116
        )
        predicted += window_pred_len

Fabrizio Milo's avatar
Fabrizio Milo committed
117

Leo Gao's avatar
Leo Gao committed
118
def make_disjoint_window(pair):
Fabrizio Milo's avatar
Fabrizio Milo committed
119
    """Takes output from get_rolling_token_windows and makes the context not overlap with the continuation"""
Leo Gao's avatar
Leo Gao committed
120
    a, b = pair
121
    return a[: len(a) - (len(b) - 1)], b
Fabrizio Milo's avatar
Fabrizio Milo committed
122

Jason Phang's avatar
Jason Phang committed
123

124
125
126
127
128
class Reorderer:
    def __init__(self, arr, fn):
        self.size = len(arr)
        arr = list(enumerate(arr))
        arr = group(arr, lambda x: fn(x[1]))
129
130
131
        # 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]
132
133
134
        arr.sort(key=lambda x: fn(x[1]))

        self.arr = arr
Fabrizio Milo's avatar
Fabrizio Milo committed
135

136
137
    def get_reordered(self):
        return [x[1] for x in self.arr]
Fabrizio Milo's avatar
Fabrizio Milo committed
138

139
140
141
142
143
    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
144
            for ind in inds:
145
146
                res[ind] = v
                cov[ind] = True
Fabrizio Milo's avatar
Fabrizio Milo committed
147

148
        assert all(cov)
Fabrizio Milo's avatar
Fabrizio Milo committed
149

150
151
        return res

Fabrizio Milo's avatar
Fabrizio Milo committed
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

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()


187
188
def positional_deprecated(fn):
    """
Fabrizio Milo's avatar
Fabrizio Milo committed
189
    A decorator to nudge users into passing only keyword args (`kwargs`) to the
190
191
    wrapped function, `fn`.
    """
Fabrizio Milo's avatar
Fabrizio Milo committed
192

193
194
    @functools.wraps(fn)
    def _wrapper(*args, **kwargs):
Fabrizio Milo's avatar
Fabrizio Milo committed
195
196
197
        if len(args) != 1 if inspect.ismethod(fn) else 0:
            print(
                f"WARNING: using {fn.__name__} with positional arguments is "
198
                "deprecated and will be disallowed in a future version of "
Fabrizio Milo's avatar
Fabrizio Milo committed
199
200
                "lm-evaluation-harness!"
            )
201
        return fn(*args, **kwargs)
Fabrizio Milo's avatar
Fabrizio Milo committed
202

203
    return _wrapper
Stephen Hogg's avatar
Stephen Hogg committed
204

Fabrizio Milo's avatar
Fabrizio Milo committed
205

Stephen Hogg's avatar
Stephen Hogg committed
206
207
208
209
210
211
212
213
214
@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
215
        if (cur_path / "tests" / "test_version_stable.py").exists():
Stephen Hogg's avatar
Stephen Hogg committed
216
217
218
            return cur_path
        else:
            cur_path = cur_path.parent.resolve()
Fabrizio Milo's avatar
Fabrizio Milo committed
219
220
221
222
    raise FileNotFoundError(
        f"Unable to find package root within {max_layers} upwards" + f"of {start_path}"
    )

Stephen Hogg's avatar
Stephen Hogg committed
223
224

@positional_deprecated
225
def run_task_tests(task_list: List[str]):
Stephen Hogg's avatar
Stephen Hogg committed
226
227
228
    """
    Find the package root and run the tests for the given tasks
    """
jon-tow's avatar
jon-tow committed
229
230
    import pytest

231
    package_root = find_test_root(start_path=pathlib.Path(__file__))
Fabrizio Milo's avatar
Fabrizio Milo committed
232
233
234
235
236
237
238
    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
239
240
241
    sys.path.append(str(package_root))
    pytest_return_val = pytest.main(args)
    if pytest_return_val:
Fabrizio Milo's avatar
Fabrizio Milo committed
242
243
244
        raise ValueError(
            f"Not all tests for the specified tasks ({task_list}) ran successfully! Error code: {pytest_return_val}"
        )
245
246


247
248
249
250
251
252
253
254
255
256
257
258
259
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
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
def import_function(loader, node):

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

    module_name, function_name = function_name.split('.')
    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

# Add the import_function constructor to the YAML loader
yaml.add_constructor('!function', import_function)


def load_yaml_config(yaml_path):
    with open(yaml_path, 'rb') as file:
        yaml_config = yaml.full_load(file)
        yaml_dir = os.path.dirname(yaml_path)
        
        if 'include' in yaml_config:
            include_path = yaml_config['include']
            del yaml_config['include']

            if type(include_path) == str:
                include_path = [include_path]
            
            # Load from the last one first
            include_path.reverse()
            final_yaml_config = {}
            for path in include_path:

                # Assumes that path is a full path.
                # If not found, assume the included yaml 
                # 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)
                except:
                    # If failed to load, ignore
                    pass

            final_yaml_config.update(yaml_config)
            return final_yaml_config
        return yaml_config


314
env = Environment(loader=BaseLoader, undefined=StrictUndefined)
315
316
317
318

def apply_template(template, doc):
    rtemplate = env.from_string(template)
    return rtemplate.render(**doc)
319