Jon Gauthier
commited on
Commit
·
2b2f744
1
Parent(s):
d254185
stage working data loader, no metric/eval
Browse files- requirements.txt +1 -0
- syntaxgym/__init__.py +0 -0
- syntaxgym/prediction.py +228 -0
- syntaxgym/syntaxgym.py +94 -0
- test.py +4 -0
requirements.txt
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
pyparsing
|
syntaxgym/__init__.py
ADDED
File without changes
|
syntaxgym/prediction.py
ADDED
@@ -0,0 +1,228 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Union, Optional as TOptional, List as TList
|
2 |
+
|
3 |
+
|
4 |
+
from pyparsing import *
|
5 |
+
import numpy as np
|
6 |
+
|
7 |
+
from syntaxgym.utils import METRICS
|
8 |
+
|
9 |
+
|
10 |
+
# Enable parser packrat (caching)
|
11 |
+
ParserElement.enablePackrat()
|
12 |
+
|
13 |
+
# Relative and absolute tolerance thresholds for surprisal equality
|
14 |
+
EQUALITY_RTOL = 1e-5
|
15 |
+
EQUALITY_ATOL = 1e-3
|
16 |
+
|
17 |
+
|
18 |
+
#######
|
19 |
+
# Define a grammar for prediction formulae.
|
20 |
+
|
21 |
+
# References a surprisal region
|
22 |
+
lpar = Suppress("(")
|
23 |
+
rpar = Suppress(")")
|
24 |
+
region = lpar + (Word(nums) | "*") + Suppress(";%") + Word(alphanums + "_-") + Suppress("%") + rpar
|
25 |
+
literal_float = pyparsing_common.number
|
26 |
+
|
27 |
+
class Region(object):
|
28 |
+
def __init__(self, tokens):
|
29 |
+
self.region_number = tokens[0]
|
30 |
+
self.condition_name = tokens[1]
|
31 |
+
|
32 |
+
def __str__(self):
|
33 |
+
return "(%s;%%%s%%)" % (self.region_number, self.condition_name)
|
34 |
+
|
35 |
+
def __repr__(self):
|
36 |
+
return "Region(%s,%s)" % (self.condition_name, self.region_number)
|
37 |
+
|
38 |
+
def __call__(self, surprisal_dict):
|
39 |
+
if self.region_number == "*":
|
40 |
+
return sum(value for (condition, region), value in surprisal_dict.items()
|
41 |
+
if condition == self.condition_name)
|
42 |
+
|
43 |
+
return surprisal_dict[self.condition_name, int(self.region_number)]
|
44 |
+
|
45 |
+
class LiteralFloat(object):
|
46 |
+
def __init__(self, tokens):
|
47 |
+
self.value = float(tokens[0])
|
48 |
+
|
49 |
+
def __str__(self):
|
50 |
+
return "%f" % (self.value,)
|
51 |
+
|
52 |
+
def __repr__(self):
|
53 |
+
return "LiteralFloat(%f)" % (self.value,)
|
54 |
+
|
55 |
+
def __call__(self, surprisal_dict):
|
56 |
+
return self.value
|
57 |
+
|
58 |
+
class BinaryOp(object):
|
59 |
+
operators: TOptional[TList[str]]
|
60 |
+
|
61 |
+
def __init__(self, tokens):
|
62 |
+
self.operator = tokens[0][1]
|
63 |
+
if self.operators is not None and self.operator not in self.operators:
|
64 |
+
raise ValueError("Invalid %s operator %s" % (self.__class__.__name__,
|
65 |
+
self.operator))
|
66 |
+
self.operands = [tokens[0][0], tokens[0][2]]
|
67 |
+
|
68 |
+
def __str__(self):
|
69 |
+
return "(%s %s %s)" % (self.operands[0], self.operator, self.operands[1])
|
70 |
+
|
71 |
+
def __repr__(self):
|
72 |
+
return "%s(%s)(%s)" % (self.__class__.__name__, self.operator, ",".join(map(repr, self.operands)))
|
73 |
+
|
74 |
+
def __call__(self, surprisal_dict):
|
75 |
+
op_vals = [op(surprisal_dict) for op in self.operands]
|
76 |
+
return self._evaluate(op_vals, surprisal_dict)
|
77 |
+
|
78 |
+
def _evaluate(self, evaluated_operands, surprisal_dict):
|
79 |
+
raise NotImplementedError()
|
80 |
+
|
81 |
+
class BoolOp(BinaryOp):
|
82 |
+
operators = ["&", "|"]
|
83 |
+
def _evaluate(self, op_vals, surprisal_dict):
|
84 |
+
if self.operator == "&":
|
85 |
+
return op_vals[0] and op_vals[1]
|
86 |
+
elif self.operator == "|":
|
87 |
+
return op_vals[0] or op_vals[1]
|
88 |
+
|
89 |
+
class FloatOp(BinaryOp):
|
90 |
+
operators = ["-", "+"]
|
91 |
+
def _evaluate(self, op_vals, surprisal_dict):
|
92 |
+
if self.operator == "-":
|
93 |
+
return op_vals[0] - op_vals[1]
|
94 |
+
elif self.operator == "+":
|
95 |
+
return op_vals[0] + op_vals[1]
|
96 |
+
|
97 |
+
class ComparatorOp(BinaryOp):
|
98 |
+
operators = ["<", ">", "="]
|
99 |
+
def _evaluate(self, op_vals, surprisal_dict):
|
100 |
+
if self.operator == "<":
|
101 |
+
return op_vals[0] < op_vals[1]
|
102 |
+
elif self.operator == ">":
|
103 |
+
return op_vals[0] > op_vals[1]
|
104 |
+
elif self.operator == "=":
|
105 |
+
return np.isclose(op_vals[0], op_vals[1],
|
106 |
+
rtol=EQUALITY_RTOL,
|
107 |
+
atol=EQUALITY_ATOL)
|
108 |
+
|
109 |
+
def Chain(op_cls, left_assoc=True):
|
110 |
+
def chainer(tokens):
|
111 |
+
"""
|
112 |
+
Create a binary tree of BinaryOps from the given repeated application
|
113 |
+
of the op.
|
114 |
+
"""
|
115 |
+
operators = tokens[0][1::2]
|
116 |
+
args = tokens[0][0::2]
|
117 |
+
if not left_assoc:
|
118 |
+
raise NotImplementedError
|
119 |
+
|
120 |
+
arg1 = args.pop(0)
|
121 |
+
while len(args) > 0:
|
122 |
+
operator = operators.pop(0)
|
123 |
+
arg2 = args.pop(0)
|
124 |
+
arg1 = op_cls([[arg1, operator, arg2]])
|
125 |
+
|
126 |
+
return arg1
|
127 |
+
|
128 |
+
return chainer
|
129 |
+
|
130 |
+
atom = region.setParseAction(Region) | literal_float.setParseAction(LiteralFloat)
|
131 |
+
|
132 |
+
prediction_expr = infixNotation(
|
133 |
+
atom,
|
134 |
+
[
|
135 |
+
(oneOf("- +"), 2, opAssoc.LEFT, Chain(FloatOp)),
|
136 |
+
(oneOf("< > ="), 2, opAssoc.LEFT, ComparatorOp),
|
137 |
+
(oneOf("& |"), 2, opAssoc.LEFT, Chain(BoolOp)),
|
138 |
+
],
|
139 |
+
lpar=lpar, rpar=rpar
|
140 |
+
)
|
141 |
+
|
142 |
+
|
143 |
+
class Prediction(object):
|
144 |
+
"""
|
145 |
+
Predictions state expected relations between language model surprisal
|
146 |
+
measures in different regions and conditions of a test suite. For more
|
147 |
+
information, see :ref:`architecture`.
|
148 |
+
"""
|
149 |
+
|
150 |
+
def __init__(self, idx: int, formula: Union[str, BinaryOp], metric: str):
|
151 |
+
"""
|
152 |
+
Args:
|
153 |
+
idx: A unique prediction ID. This is only relevant for
|
154 |
+
serialization.
|
155 |
+
formula: A string representation of the prediction formula, or an
|
156 |
+
already parsed formula. For more information, see
|
157 |
+
:ref:`architecture`.
|
158 |
+
metric: Metric for aggregating surprisals within regions.
|
159 |
+
"""
|
160 |
+
if isinstance(formula, str):
|
161 |
+
try:
|
162 |
+
formula = prediction_expr.parseString(formula, parseAll=True)[0]
|
163 |
+
except ParseException as e:
|
164 |
+
raise ValueError("Invalid formula expression %r" % (formula,)) from e
|
165 |
+
|
166 |
+
self.idx = idx
|
167 |
+
self.formula = formula
|
168 |
+
|
169 |
+
if metric not in METRICS.keys():
|
170 |
+
raise ValueError("Unknown metric %s. Supported metrics: %s" %
|
171 |
+
(metric, " ".join(METRICS.keys())))
|
172 |
+
self.metric = metric
|
173 |
+
|
174 |
+
def __call__(self, item):
|
175 |
+
"""
|
176 |
+
Evaluate the prediction on the given item dict representation. For more
|
177 |
+
information on item representations, see :ref:`suite_json`.
|
178 |
+
"""
|
179 |
+
# Prepare relevant surprisal dict
|
180 |
+
surps = {(c["condition_name"], r["region_number"]): r["metric_value"][self.metric]
|
181 |
+
for c in item["conditions"]
|
182 |
+
for r in c["regions"]}
|
183 |
+
return self.formula(surps)
|
184 |
+
|
185 |
+
@classmethod
|
186 |
+
def from_dict(cls, pred_dict, idx: int, metric: str):
|
187 |
+
"""
|
188 |
+
Parse from a prediction dictionary representation (see
|
189 |
+
:ref:`suite_json`).
|
190 |
+
"""
|
191 |
+
if not pred_dict["type"] == "formula":
|
192 |
+
raise ValueError("Unknown prediction type %s" % (pred_dict["type"],))
|
193 |
+
|
194 |
+
return cls(formula=pred_dict["formula"], idx=idx, metric=metric)
|
195 |
+
|
196 |
+
@property
|
197 |
+
def referenced_regions(self):
|
198 |
+
"""
|
199 |
+
Get a set of the regions referenced by this formula.
|
200 |
+
Each item is a tuple of the form ``(condition_name, region_number)``.
|
201 |
+
"""
|
202 |
+
def traverse(x, acc):
|
203 |
+
if isinstance(x, BinaryOp):
|
204 |
+
for val in x.operands:
|
205 |
+
traverse(val, acc)
|
206 |
+
elif isinstance(x, Region):
|
207 |
+
acc.add((x.condition_name, int(x.region_number)))
|
208 |
+
|
209 |
+
return acc
|
210 |
+
|
211 |
+
return traverse(self.formula, set())
|
212 |
+
|
213 |
+
def as_dict(self):
|
214 |
+
"""
|
215 |
+
Serialize as a prediction dictionary representation (see
|
216 |
+
:ref:`suite_json`).
|
217 |
+
"""
|
218 |
+
return dict(type="formula", formula=str(self.formula))
|
219 |
+
|
220 |
+
def __str__(self):
|
221 |
+
return "Prediction(%s)" % (self.formula,)
|
222 |
+
__repr__ = __str__
|
223 |
+
|
224 |
+
def __hash__(self):
|
225 |
+
return hash(self.formula)
|
226 |
+
|
227 |
+
def __eq__(self, other):
|
228 |
+
return isinstance(other, Prediction) and hash(self) == hash(other)
|
syntaxgym/syntaxgym.py
ADDED
@@ -0,0 +1,94 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
|
3 |
+
"""
|
4 |
+
SyntaxGym dataset as used in Hu et al. (2020).
|
5 |
+
"""
|
6 |
+
|
7 |
+
|
8 |
+
import json
|
9 |
+
from pathlib import Path
|
10 |
+
from typing import List
|
11 |
+
|
12 |
+
import datasets
|
13 |
+
|
14 |
+
|
15 |
+
_CITATION = """
|
16 |
+
@inproceedings{Hu:et-al:2020,
|
17 |
+
author = {Hu, Jennifer and Gauthier, Jon and Qian, Peng and Wilcox, Ethan and Levy, Roger},
|
18 |
+
title = {A systematic assessment of syntactic generalization in neural language models},
|
19 |
+
booktitle = {Proceedings of the Association of Computational Linguistics},
|
20 |
+
year = {2020}
|
21 |
+
}
|
22 |
+
"""
|
23 |
+
|
24 |
+
_DESCRIPTION = "" # TODO
|
25 |
+
|
26 |
+
|
27 |
+
_PROJECT_URL = "https://syntaxgym.org"
|
28 |
+
_DOWNLOAD_URL = "https://github.com/cpllab/syntactic-generalization"
|
29 |
+
|
30 |
+
|
31 |
+
print("_____HERE")
|
32 |
+
SUITE_JSONS = []
|
33 |
+
for suite_f in Path("test_suites").glob("*.json"):
|
34 |
+
with suite_f.open() as f:
|
35 |
+
SUITE_JSONS.append(json.load(f))
|
36 |
+
|
37 |
+
|
38 |
+
class SyntaxGymSuiteConfig(datasets.BuilderConfig):
|
39 |
+
|
40 |
+
def __init__(self, suite_json, version=datasets.Version("1.0.0"), **kwargs):
|
41 |
+
self.meta = suite_json["meta"]
|
42 |
+
name = self.meta["name"]
|
43 |
+
description = f"SyntaxGym test suite {name}.\n" + _DESCRIPTION
|
44 |
+
|
45 |
+
super().__init__(name=name, description=description, version=version,
|
46 |
+
**kwargs)
|
47 |
+
|
48 |
+
self.features = list(suite_json["region_meta"].values())
|
49 |
+
|
50 |
+
|
51 |
+
class SyntaxGym(datasets.GeneratorBasedBuilder):
|
52 |
+
|
53 |
+
BUILDER_CONFIGS = [SyntaxGymSuiteConfig(suite_json)
|
54 |
+
for suite_json in SUITE_JSONS]
|
55 |
+
|
56 |
+
def _info(self):
|
57 |
+
condition_spec = {
|
58 |
+
"condition_name": datasets.Value("string"),
|
59 |
+
"regions": datasets.Sequence({
|
60 |
+
"region_number": datasets.Value("int32"),
|
61 |
+
"content": datasets.Value("string")
|
62 |
+
})
|
63 |
+
}
|
64 |
+
|
65 |
+
features = {
|
66 |
+
"item_number": datasets.Value("string"),
|
67 |
+
"conditions": datasets.Sequence(condition_spec)
|
68 |
+
}
|
69 |
+
|
70 |
+
citation = ""
|
71 |
+
if self.config.meta["reference"]:
|
72 |
+
citation = f"Test suite citation: {self.meta['reference']}\n"
|
73 |
+
citation += f"SyntaxGym citation:\n{_CITATION}"
|
74 |
+
|
75 |
+
return datasets.DatasetInfo(
|
76 |
+
description=_DESCRIPTION,
|
77 |
+
features=datasets.Features(features),
|
78 |
+
homepage=_PROJECT_URL,
|
79 |
+
citation=citation,
|
80 |
+
)
|
81 |
+
|
82 |
+
def _split_generators(self, dl_manager: datasets.DownloadManager) -> List[datasets.SplitGenerator]:
|
83 |
+
return [datasets.SplitGenerator(name=datasets.Split.TEST,
|
84 |
+
gen_kwargs={"name": self.config.name})]
|
85 |
+
|
86 |
+
def _generate_examples(self, name):
|
87 |
+
# DEV: NB suite jsons already loaded because BUILDER_CONFIGS is static
|
88 |
+
suite_jsons = SUITE_JSONS
|
89 |
+
|
90 |
+
suite_json = next(suite for suite in SUITE_JSONS
|
91 |
+
if suite["meta"]["name"] == name)
|
92 |
+
|
93 |
+
for item in suite_json["items"]:
|
94 |
+
yield item["item_number"], item
|
test.py
ADDED
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import datasets
|
2 |
+
|
3 |
+
|
4 |
+
dataset = datasets.load_dataset("syntaxgym", "mvrr_mod")
|