Datasets:

Modalities:
Text
Libraries:
Datasets
Jon Gauthier commited on
Commit
535608c
·
1 Parent(s): 9d3d980

Implementation nice and clean, but need to verify results w/ reference syntaxgym-core still

Browse files
Files changed (2) hide show
  1. syntaxgym.py +86 -57
  2. test.py +6 -39
syntaxgym.py CHANGED
@@ -13,8 +13,10 @@ 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
 
@@ -64,26 +66,28 @@ class SyntaxGymSuiteConfig(datasets.BuilderConfig):
64
  self.features = list(suite_json["region_meta"].values())
65
 
66
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
67
  class SyntaxGym(datasets.GeneratorBasedBuilder):
68
 
69
  BUILDER_CONFIGS = [SyntaxGymSuiteConfig(suite_json)
70
  for suite_json in SUITE_JSONS.values()]
71
 
72
  def _info(self):
73
- condition_spec = {
74
- "condition_name": datasets.Value("string"),
75
- "content": datasets.Value("string"),
76
- "regions": datasets.Sequence({
77
- "region_number": datasets.Value("int32"),
78
- "content": datasets.Value("string")
79
- })
80
- }
81
-
82
- features = {
83
- "item_number": datasets.Value("int32"),
84
- "conditions": datasets.Sequence(condition_spec)
85
- }
86
-
87
  citation = ""
88
  if self.config.meta["reference"]:
89
  citation = f"Test suite citation: {self.meta['reference']}\n"
@@ -91,7 +95,7 @@ class SyntaxGym(datasets.GeneratorBasedBuilder):
91
 
92
  return datasets.DatasetInfo(
93
  description=_DESCRIPTION,
94
- features=datasets.Features(features),
95
  homepage=_PROJECT_URL,
96
  citation=citation,
97
  )
@@ -103,12 +107,15 @@ class SyntaxGym(datasets.GeneratorBasedBuilder):
103
  def _generate_examples(self, name):
104
  # DEV: NB suite jsons already loaded because BUILDER_CONFIGS is static
105
  suite_json = SUITE_JSONS[name]
 
106
 
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
 
114
 
@@ -117,27 +124,10 @@ class SyntaxGymMetric(datasets.Metric):
117
  SyntaxGym prediction evaluation metric.
118
  """
119
 
120
- def __init__(self, *args, **kwargs):
121
- super().__init__(*args, **kwargs)
122
- self.suite = SUITE_JSONS[self.config_name]
123
- self.predictions = [
124
- Prediction(idx, p["formula"], "sum")
125
- for idx, p in enumerate(self.suite["predictions"])
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",
@@ -146,16 +136,53 @@ class SyntaxGymMetric(datasets.Metric):
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:]
@@ -170,10 +197,11 @@ class SyntaxGymMetric(datasets.Metric):
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():
@@ -183,7 +211,8 @@ class SyntaxGymMetric(datasets.Metric):
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 = {(condition_name, region_number): float(total)
189
  for condition_name, totals in region_totals.items()
@@ -191,33 +220,29 @@ class SyntaxGymMetric(datasets.Metric):
191
 
192
  results = {
193
  "prediction_results": [
194
- p.formula(region_totals) for p in self.predictions
 
195
  ],
196
 
197
  "region_totals": region_totals
198
  }
199
  return results
200
 
201
- def get_region_edges(self, item_number, condition_name):
202
  """
203
  Get left edge of each region as a character index.
204
  """
205
  # NB this is coupled with `condition_to_string` logic of course
206
 
207
- # DEV bad, just reindex
208
- item = next(item for item in self.suite["items"]
209
- if item["item_number"] == item_number)
210
- cond = next(cond for cond in item["conditions"]
211
- if cond["condition_name"] == condition_name)
212
 
213
  idx = 0
214
  ret = []
215
- for r_idx, region in enumerate(cond["regions"]):
216
  ret.append(idx)
217
 
218
- content = region["content"]
219
- region_size = len(content)
220
- if content.strip() != "" and r_idx != 0 and not content.startswith(","):
221
  # Add joining space
222
  region_size += 1
223
 
@@ -225,16 +250,20 @@ class SyntaxGymMetric(datasets.Metric):
225
 
226
  return ret
227
 
228
- def compute_region_token_mapping(self, item_number, condition_names, input_ids,
229
  offset_mapping: List[Tuple[int, int]]):
230
  # input_ids: B * T
231
  # offset_mapping: B * T * 2
 
232
 
 
233
  region2tokens = {cond: defaultdict(list) for cond in condition_names}
234
 
 
 
235
  input_ids = input_ids.detach()
236
- for i_cond, i_tokens, i_offsets in zip(condition_names, input_ids, offset_mapping):
237
- region_edges = self.get_region_edges(item_number, i_cond)
238
 
239
  t_cursor, r_cursor = 0, 0
240
  while t_cursor < i_tokens.shape[0]:
@@ -243,14 +272,14 @@ class SyntaxGymMetric(datasets.Metric):
243
 
244
  region_start = region_edges[r_cursor]
245
  region_end = region_edges[r_cursor + 1] \
246
- if r_cursor + 1 < len(region_edges) else np.inf
247
 
248
  # NB region boundaries are left edges, hence the >= here.
249
  if token_char_start >= region_end:
250
  r_cursor += 1
251
  continue
252
 
253
- region2tokens[i_cond][r_cursor + 1].append(t_cursor)
254
  t_cursor += 1
255
 
256
  return region2tokens
 
13
  from typing import List, Tuple
14
 
15
  import datasets
16
+ from datasets import logging
17
  import numpy as np
18
  import torch
19
+ from transformers import AutoModelForCausalLM, AutoTokenizer
20
 
21
  from .prediction import Prediction
22
 
 
66
  self.features = list(suite_json["region_meta"].values())
67
 
68
 
69
+ SUITE_DATASET_CONDITION_SPEC = {
70
+ "condition_name": datasets.Value("string"),
71
+ "content": datasets.Value("string"),
72
+ "regions": datasets.Sequence({
73
+ "region_number": datasets.Value("int32"),
74
+ "content": datasets.Value("string")
75
+ })
76
+ }
77
+
78
+ SUITE_DATASET_SPEC = {
79
+ "item_number": datasets.Value("int32"),
80
+ "conditions": datasets.Sequence(SUITE_DATASET_CONDITION_SPEC),
81
+ "predictions": datasets.Sequence(datasets.Value("string")),
82
+ }
83
+
84
+
85
  class SyntaxGym(datasets.GeneratorBasedBuilder):
86
 
87
  BUILDER_CONFIGS = [SyntaxGymSuiteConfig(suite_json)
88
  for suite_json in SUITE_JSONS.values()]
89
 
90
  def _info(self):
 
 
 
 
 
 
 
 
 
 
 
 
 
 
91
  citation = ""
92
  if self.config.meta["reference"]:
93
  citation = f"Test suite citation: {self.meta['reference']}\n"
 
95
 
96
  return datasets.DatasetInfo(
97
  description=_DESCRIPTION,
98
+ features=datasets.Features(SUITE_DATASET_SPEC),
99
  homepage=_PROJECT_URL,
100
  citation=citation,
101
  )
 
107
  def _generate_examples(self, name):
108
  # DEV: NB suite jsons already loaded because BUILDER_CONFIGS is static
109
  suite_json = SUITE_JSONS[name]
110
+ predictions = [p["formula"] for p in suite_json["predictions"]]
111
 
112
  for item in suite_json["items"]:
113
  # Convert to sentence input.
114
  for cond in item["conditions"]:
115
  cond["content"] = condition_to_string(cond)
116
 
117
+ item["predictions"] = predictions
118
+
119
  yield item["item_number"], item
120
 
121
 
 
124
  SyntaxGym prediction evaluation metric.
125
  """
126
 
 
 
 
 
 
 
 
 
127
  def _info(self):
128
  seq = datasets.Sequence
129
  features = datasets.Features({
130
+ "suite": SUITE_DATASET_SPEC
 
 
 
 
 
 
 
 
 
131
  })
132
  return datasets.MetricInfo(
133
  description="TODO",
 
136
  features=features,
137
  )
138
 
139
+ def _compute(self, suite, model_id, batch_size: int = 16,
140
+ add_start_token=True, device=None):
141
+ if device is not None:
142
+ assert device in ["gpu", "cpu", "cuda"]
143
+ if device == "gpu":
144
+ device = "cuda"
145
+ else:
146
+ device = "cuda" if torch.cuda.is_available() else "cpu"
147
+
148
+ model = AutoModelForCausalLM.from_pretrained(model_id)
149
+ model = model.to(device)
150
+ model.eval()
151
+
152
+ tokenizer = AutoTokenizer.from_pretrained(model_id)
153
+ # TODO copy from perplexity metric
154
+ tokenizer.pad_token = tokenizer.eos_token
155
+
156
+ results = {"prediction_results": [], "region_totals": []}
157
+ # TODO batch all items together
158
+ for item in logging.tqdm(suite):
159
+ result_single = self._compute_single(item, tokenizer, model, device)
160
+
161
+ for k in ["prediction_results", "region_totals"]:
162
+ results[k].append(result_single[k])
163
+
164
+ return results
165
+
166
+
167
+ def _compute_single(self, item, tokenizer, model, device):
168
+ tokenized = tokenizer(item["conditions"]["content"],
169
+ padding=True,
170
+ return_tensors="pt",
171
+ return_offsets_mapping=True).to(device)
172
 
173
  # input_ids: B * T
174
+ input_ids = tokenized["input_ids"]
175
  assert input_ids.ndim == 2
176
 
177
+ # Compute sentence level surprisals.
178
+ with torch.no_grad():
179
+ # Pre-softmax predictive distribution B * T * V
180
+ logits = model(input_ids).logits
181
+ surprisals = -logits.log_softmax(dim=2) / np.log(2)
182
+
183
+ # surprisals: B * T * V
184
+ assert surprisals.ndim == 3
185
+
186
  # Get surprisals of expected words.
187
  surps_shifted = surprisals[:, :-1, :]
188
  expected_ids = input_ids[:, 1:]
 
197
  # surprisals is now B * (T - 1)
198
 
199
  #### aggregate
200
+ condition_names = item["conditions"]["condition_name"]
201
  region_totals = {condition_name: defaultdict(float)
202
  for condition_name in condition_names}
203
  region2tokens = self.compute_region_token_mapping(
204
+ item, input_ids, tokenized["offset_mapping"])
205
 
206
  for i, (i_cond, i_inputs) in enumerate(zip(condition_names, input_ids)):
207
  for region_number, region_tokens in region2tokens[i_cond].items():
 
211
  else:
212
  # TODO don't think this is an issue, just should clean
213
  # up the aggregation output
214
+ assert token == surprisals.shape[1], \
215
+ "%s %s" % (token, surprisals.shape[1])
216
 
217
  region_totals = {(condition_name, region_number): float(total)
218
  for condition_name, totals in region_totals.items()
 
220
 
221
  results = {
222
  "prediction_results": [
223
+ Prediction(i, formula, "sum").formula(region_totals)
224
+ for i, formula in enumerate(item["predictions"])
225
  ],
226
 
227
  "region_totals": region_totals
228
  }
229
  return results
230
 
231
+ def get_region_edges(self, item, condition_idx):
232
  """
233
  Get left edge of each region as a character index.
234
  """
235
  # NB this is coupled with `condition_to_string` logic of course
236
 
237
+ regions = item["conditions"]["regions"][condition_idx]
 
 
 
 
238
 
239
  idx = 0
240
  ret = []
241
+ for r_idx, region_content in enumerate(regions["content"]):
242
  ret.append(idx)
243
 
244
+ region_size = len(region_content)
245
+ if region_content.strip() != "" and r_idx != 0 and not region_content.startswith(","):
 
246
  # Add joining space
247
  region_size += 1
248
 
 
250
 
251
  return ret
252
 
253
+ def compute_region_token_mapping(self, item, input_ids: torch.LongTensor,
254
  offset_mapping: List[Tuple[int, int]]):
255
  # input_ids: B * T
256
  # offset_mapping: B * T * 2
257
+ # assumes batch is sorted according to item's condition_name order
258
 
259
+ condition_names = item["conditions"]["condition_name"]
260
  region2tokens = {cond: defaultdict(list) for cond in condition_names}
261
 
262
+ max_long = torch.iinfo(torch.int64).max
263
+
264
  input_ids = input_ids.detach()
265
+ for i_cond, (i_tokens, i_offsets) in enumerate(zip(input_ids, offset_mapping)):
266
+ region_edges = self.get_region_edges(item, i_cond)
267
 
268
  t_cursor, r_cursor = 0, 0
269
  while t_cursor < i_tokens.shape[0]:
 
272
 
273
  region_start = region_edges[r_cursor]
274
  region_end = region_edges[r_cursor + 1] \
275
+ if r_cursor + 1 < len(region_edges) else max_long
276
 
277
  # NB region boundaries are left edges, hence the >= here.
278
  if token_char_start >= region_end:
279
  r_cursor += 1
280
  continue
281
 
282
+ region2tokens[condition_names[i_cond]][r_cursor + 1].append(t_cursor)
283
  t_cursor += 1
284
 
285
  return region2tokens
test.py CHANGED
@@ -7,44 +7,11 @@ import transformers
7
  import torch
8
 
9
 
10
- dataset = datasets.load_dataset("syntaxgym.py", "mvrr_mod")
11
- metric = datasets.load_metric("syntaxgym.py", "mvrr_mod")
12
 
13
- # model_ref = "gpt2"
14
- model_ref = "hf-internal-testing/tiny-random-gpt_neo"
15
 
16
- tokenizer = transformers.AutoTokenizer.from_pretrained(model_ref)
17
- # DEV
18
- tokenizer.pad_token = tokenizer.eos_token
19
-
20
- model = transformers.AutoModelForCausalLM.from_pretrained(model_ref)
21
- model.eval()
22
-
23
-
24
- # all_sentences: List[List[str]] = [item["conditions"]["content"] for item in dataset["test"]]
25
- # all_sentences_flat = list(itertools.chain.from_iterable(all_sentences))
26
-
27
- tokenized = tokenizer(all_sentences_flat,
28
- return_tensors="pt", padding=True, return_offsets_mapping=True)
29
- for item in dataset["test"]:
30
- # TODO full preprocessing setup
31
- condition_names = item["conditions"]["condition_name"]
32
- tokenized = tokenizer(item["conditions"]["content"], return_tensors="pt",
33
- padding=True, return_offsets_mapping=True)
34
-
35
- print(item)
36
- print(tokenized)
37
- print(tokenized["offset_mapping"].shape)
38
-
39
- with torch.no_grad():
40
- # Pre-softmax predictive distribution (shape B * T * V)
41
- output = model(tokenized["input_ids"])[0]
42
- surprisals = -output.log_softmax(dim=2) / np.log(2)
43
-
44
- result = metric.compute(surprisals=surprisals,
45
- item_number=item["item_number"],
46
- condition_names=condition_names,
47
- input_ids=tokenized["input_ids"],
48
- offset_mapping=tokenized["offset_mapping"])
49
-
50
- break
 
7
  import torch
8
 
9
 
10
+ dataset = datasets.load_dataset("syntaxgym.py", "subordination_src-src")
11
+ metric = datasets.load_metric("syntaxgym.py")
12
 
13
+ model_ref = "gpt2"
14
+ # model_ref = "hf-internal-testing/tiny-random-gpt_neo"
15
 
16
+ result = metric.compute(suite=dataset["test"], model_id=model_ref)
17
+ print(result)