Jon Gauthier
commited on
Commit
·
70f4226
1
Parent(s):
aab1635
Get token aggregation working; just need to support predictions
Browse files- syntaxgym.py +123 -17
- test.py +18 -6
syntaxgym.py
CHANGED
@@ -5,13 +5,16 @@ SyntaxGym dataset as used in Hu et al. (2020).
|
|
5 |
"""
|
6 |
|
7 |
|
|
|
8 |
from copy import deepcopy
|
9 |
import json
|
10 |
from pathlib import Path
|
11 |
import re
|
12 |
-
from typing import List
|
13 |
|
14 |
import datasets
|
|
|
|
|
15 |
|
16 |
from .prediction import Prediction
|
17 |
|
@@ -39,6 +42,15 @@ for suite_f in Path("test_suites").glob("*.json"):
|
|
39 |
SUITE_JSONS = {suite["meta"]["name"]: suite for suite in SUITE_JSONS}
|
40 |
|
41 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
42 |
class SyntaxGymSuiteConfig(datasets.BuilderConfig):
|
43 |
|
44 |
def __init__(self, suite_json, version=datasets.Version("1.0.0"), **kwargs):
|
@@ -68,7 +80,7 @@ class SyntaxGym(datasets.GeneratorBasedBuilder):
|
|
68 |
}
|
69 |
|
70 |
features = {
|
71 |
-
"item_number": datasets.Value("
|
72 |
"conditions": datasets.Sequence(condition_spec)
|
73 |
}
|
74 |
|
@@ -95,10 +107,7 @@ class SyntaxGym(datasets.GeneratorBasedBuilder):
|
|
95 |
for item in suite_json["items"]:
|
96 |
# Convert to sentence input.
|
97 |
for cond in item["conditions"]:
|
98 |
-
cond["content"] =
|
99 |
-
for region in cond["regions"]
|
100 |
-
if region["content"].strip() != ""])
|
101 |
-
cond["content"] = re.sub(r"\s+,", ",", cond["content"])
|
102 |
|
103 |
yield item["item_number"], item
|
104 |
|
@@ -117,16 +126,18 @@ class SyntaxGymMetric(datasets.Metric):
|
|
117 |
]
|
118 |
|
119 |
def _info(self):
|
|
|
120 |
features = datasets.Features({
|
121 |
-
|
122 |
-
|
123 |
-
|
124 |
-
|
125 |
-
|
126 |
-
|
127 |
-
|
128 |
-
|
129 |
-
|
|
|
130 |
})
|
131 |
return datasets.MetricInfo(
|
132 |
description="TODO",
|
@@ -135,5 +146,100 @@ class SyntaxGymMetric(datasets.Metric):
|
|
135 |
features=features,
|
136 |
)
|
137 |
|
138 |
-
def _compute(self,
|
139 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
5 |
"""
|
6 |
|
7 |
|
8 |
+
from collections import defaultdict
|
9 |
from copy import deepcopy
|
10 |
import json
|
11 |
from pathlib import Path
|
12 |
import re
|
13 |
+
from typing import List, Tuple
|
14 |
|
15 |
import datasets
|
16 |
+
import numpy as np
|
17 |
+
import torch
|
18 |
|
19 |
from .prediction import Prediction
|
20 |
|
|
|
42 |
SUITE_JSONS = {suite["meta"]["name"]: suite for suite in SUITE_JSONS}
|
43 |
|
44 |
|
45 |
+
def condition_to_string(cond):
|
46 |
+
ret = " ".join([region["content"].lstrip()
|
47 |
+
for region in cond["regions"]
|
48 |
+
if region["content"].strip() != ""])
|
49 |
+
ret = re.sub(r"\s+,", ",", ret)
|
50 |
+
|
51 |
+
return ret
|
52 |
+
|
53 |
+
|
54 |
class SyntaxGymSuiteConfig(datasets.BuilderConfig):
|
55 |
|
56 |
def __init__(self, suite_json, version=datasets.Version("1.0.0"), **kwargs):
|
|
|
80 |
}
|
81 |
|
82 |
features = {
|
83 |
+
"item_number": datasets.Value("int32"),
|
84 |
"conditions": datasets.Sequence(condition_spec)
|
85 |
}
|
86 |
|
|
|
107 |
for item in suite_json["items"]:
|
108 |
# Convert to sentence input.
|
109 |
for cond in item["conditions"]:
|
110 |
+
cond["content"] = condition_to_string(cond)
|
|
|
|
|
|
|
111 |
|
112 |
yield item["item_number"], item
|
113 |
|
|
|
126 |
]
|
127 |
|
128 |
def _info(self):
|
129 |
+
seq = datasets.Sequence
|
130 |
features = datasets.Features({
|
131 |
+
# surprisals: 3d float array
|
132 |
+
"surprisals": seq(seq(datasets.Value("float"))),
|
133 |
+
|
134 |
+
# TODO necessary? can assume it remains sorted?
|
135 |
+
"condition_names": datasets.Value("string"),
|
136 |
+
|
137 |
+
"input_ids": seq(datasets.Value("int32")),
|
138 |
+
|
139 |
+
# offset mapping: 3d int array
|
140 |
+
"offset_mapping": seq(seq(datasets.Value("int32"))),
|
141 |
})
|
142 |
return datasets.MetricInfo(
|
143 |
description="TODO",
|
|
|
146 |
features=features,
|
147 |
)
|
148 |
|
149 |
+
def _compute(self, surprisals, item_number, condition_names,
|
150 |
+
input_ids, offset_mapping):
|
151 |
+
# surprisals: B * T * V
|
152 |
+
surprisals = torch.tensor(surprisals)
|
153 |
+
assert surprisals.ndim == 3
|
154 |
+
|
155 |
+
# input_ids: B * T
|
156 |
+
input_ids = torch.tensor(input_ids)
|
157 |
+
assert input_ids.ndim == 2
|
158 |
+
|
159 |
+
# Get surprisals of expected words.
|
160 |
+
surps_shifted = surprisals[:, :-1, :]
|
161 |
+
expected_ids = input_ids[:, 1:]
|
162 |
+
|
163 |
+
# TODO: check this logic
|
164 |
+
tt = expected_ids.unsqueeze(2)
|
165 |
+
surprisals = torch.gather(surps_shifted, 2, expected_ids.unsqueeze(2)) \
|
166 |
+
.squeeze(2)
|
167 |
+
# This is the original, which works but not with multiple axes in expected_ids
|
168 |
+
# surprisals = surps_shifted[range(surps_shifted.shape[0]), expected_ids]
|
169 |
+
|
170 |
+
# surprisals is now B * (T - 1)
|
171 |
+
|
172 |
+
#### aggregate
|
173 |
+
region_totals = {condition_name: defaultdict(float)
|
174 |
+
for condition_name in condition_names}
|
175 |
+
region2tokens = self.compute_region_token_mapping(
|
176 |
+
item_number, condition_names, input_ids, offset_mapping)
|
177 |
+
|
178 |
+
for i, (i_cond, i_inputs) in enumerate(zip(condition_names, input_ids)):
|
179 |
+
for region_number, region_tokens in region2tokens[i_cond].items():
|
180 |
+
for token in region_tokens:
|
181 |
+
if token < surprisals.shape[1]:
|
182 |
+
region_totals[i_cond][region_number] += surprisals[i, token]
|
183 |
+
else:
|
184 |
+
# TODO don't think this is an issue, just should clean
|
185 |
+
# up the aggregation output
|
186 |
+
print("Warning: exceeded ", token)
|
187 |
+
|
188 |
+
region_totals = {c: dict(totals) for c, totals in region_totals.items()}
|
189 |
+
|
190 |
+
def get_region_edges(self, item_number, condition_name):
|
191 |
+
"""
|
192 |
+
Get left edge of each region as a character index.
|
193 |
+
"""
|
194 |
+
# NB this is coupled with `condition_to_string` logic of course
|
195 |
+
|
196 |
+
# DEV bad, just reindex
|
197 |
+
item = next(item for item in self.suite["items"]
|
198 |
+
if item["item_number"] == item_number)
|
199 |
+
cond = next(cond for cond in item["conditions"]
|
200 |
+
if cond["condition_name"] == condition_name)
|
201 |
+
|
202 |
+
idx = 0
|
203 |
+
ret = []
|
204 |
+
for r_idx, region in enumerate(cond["regions"]):
|
205 |
+
ret.append(idx)
|
206 |
+
|
207 |
+
content = region["content"]
|
208 |
+
region_size = len(content)
|
209 |
+
if content.strip() != "" and r_idx != 0 and not content.startswith(","):
|
210 |
+
# Add joining space
|
211 |
+
region_size += 1
|
212 |
+
|
213 |
+
idx += region_size
|
214 |
+
|
215 |
+
return ret
|
216 |
+
|
217 |
+
def compute_region_token_mapping(self, item_number, condition_names, input_ids,
|
218 |
+
offset_mapping: List[Tuple[int, int]]):
|
219 |
+
# input_ids: B * T
|
220 |
+
# offset_mapping: B * T * 2
|
221 |
+
|
222 |
+
region2tokens = {cond: defaultdict(list) for cond in condition_names}
|
223 |
+
|
224 |
+
input_ids = input_ids.detach()
|
225 |
+
for i_cond, i_tokens, i_offsets in zip(condition_names, input_ids, offset_mapping):
|
226 |
+
region_edges = self.get_region_edges(item_number, i_cond)
|
227 |
+
|
228 |
+
t_cursor, r_cursor = 0, 0
|
229 |
+
while t_cursor < i_tokens.shape[0]:
|
230 |
+
# token = i_tokens[t_cursor]
|
231 |
+
token_char_start, token_char_end = i_offsets[t_cursor]
|
232 |
+
|
233 |
+
region_start = region_edges[r_cursor]
|
234 |
+
region_end = region_edges[r_cursor + 1] \
|
235 |
+
if r_cursor + 1 < len(region_edges) else np.inf
|
236 |
+
|
237 |
+
# NB region boundaries are left edges, hence the >= here.
|
238 |
+
if token_char_start >= region_end:
|
239 |
+
r_cursor += 1
|
240 |
+
continue
|
241 |
+
|
242 |
+
region2tokens[i_cond][r_cursor + 1].append(t_cursor)
|
243 |
+
t_cursor += 1
|
244 |
+
|
245 |
+
return region2tokens
|
test.py
CHANGED
@@ -1,4 +1,5 @@
|
|
1 |
import datasets
|
|
|
2 |
import transformers
|
3 |
import torch
|
4 |
|
@@ -6,25 +7,36 @@ import torch
|
|
6 |
dataset = datasets.load_dataset("syntaxgym.py", "mvrr_mod")
|
7 |
metric = datasets.load_metric("syntaxgym.py", "mvrr_mod")
|
8 |
|
9 |
-
|
|
|
|
|
|
|
10 |
# DEV
|
11 |
tokenizer.pad_token = tokenizer.eos_token
|
12 |
|
13 |
-
model = transformers.AutoModelForCausalLM.from_pretrained(
|
14 |
model.eval()
|
15 |
|
16 |
|
17 |
for item in dataset["test"]:
|
18 |
# TODO full preprocessing setup
|
|
|
19 |
tokenized = tokenizer(item["conditions"]["content"], return_tensors="pt",
|
20 |
-
padding=True)
|
21 |
|
22 |
print(item)
|
23 |
print(tokenized)
|
|
|
24 |
|
25 |
with torch.no_grad():
|
26 |
-
|
27 |
-
|
28 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
29 |
|
30 |
break
|
|
|
1 |
import datasets
|
2 |
+
import numpy as np
|
3 |
import transformers
|
4 |
import torch
|
5 |
|
|
|
7 |
dataset = datasets.load_dataset("syntaxgym.py", "mvrr_mod")
|
8 |
metric = datasets.load_metric("syntaxgym.py", "mvrr_mod")
|
9 |
|
10 |
+
# model_ref = "gpt2"
|
11 |
+
model_ref = "hf-internal-testing/tiny-random-gpt_neo"
|
12 |
+
|
13 |
+
tokenizer = transformers.AutoTokenizer.from_pretrained(model_ref)
|
14 |
# DEV
|
15 |
tokenizer.pad_token = tokenizer.eos_token
|
16 |
|
17 |
+
model = transformers.AutoModelForCausalLM.from_pretrained(model_ref)
|
18 |
model.eval()
|
19 |
|
20 |
|
21 |
for item in dataset["test"]:
|
22 |
# TODO full preprocessing setup
|
23 |
+
condition_names = item["conditions"]["condition_name"]
|
24 |
tokenized = tokenizer(item["conditions"]["content"], return_tensors="pt",
|
25 |
+
padding=True, return_offsets_mapping=True)
|
26 |
|
27 |
print(item)
|
28 |
print(tokenized)
|
29 |
+
print(tokenized["offset_mapping"].shape)
|
30 |
|
31 |
with torch.no_grad():
|
32 |
+
# Pre-softmax predictive distribution (shape B * T * V)
|
33 |
+
output = model(tokenized["input_ids"])[0]
|
34 |
+
surprisals = -output.log_softmax(dim=2) / np.log(2)
|
35 |
+
|
36 |
+
result = metric.compute(surprisals=surprisals,
|
37 |
+
item_number=item["item_number"],
|
38 |
+
condition_names=condition_names,
|
39 |
+
input_ids=tokenized["input_ids"],
|
40 |
+
offset_mapping=tokenized["offset_mapping"])
|
41 |
|
42 |
break
|