Repository navigation
Expand file tree
/
Copy pathevaluate_candidate_independent.py
More file actions
335 lines (311 loc) · 15.1 KB
/
Copy pathevaluate_candidate_independent.py
File metadata and controls
335 lines (311 loc) · 15.1 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
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
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
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
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
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
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
"""Evaluate quality and structural invariance of candidate-independent Gemma."""
from __future__ import annotations
import argparse
import inspect
import json
import math
import platform
import random
import sys
import time
from collections import Counter, defaultdict
from pathlib import Path
from typing import Any
import torch
from torch.utils.data import DataLoader
from data import option_id, permute_example
from data_candidate_independent import CandidateIndependentCollator, JsonlDecisionDataset
from modeling_candidate_independent_gemma import ARCHITECTURE, CandidateIndependentGemma
def _format_duration(seconds: float) -> str:
seconds = max(0, int(seconds))
hours, remainder = divmod(seconds, 3600)
minutes, seconds = divmod(remainder, 60)
if hours:
return f"{hours}h{minutes:02d}m"
if minutes:
return f"{minutes}m{seconds:02d}s"
return f"{seconds}s"
def _progress_batches(loader, description: str):
total = len(loader)
started = time.perf_counter()
for index, batch in enumerate(loader, start=1):
yield batch
elapsed = time.perf_counter() - started
rate = index / elapsed if elapsed > 0 else 0.0
remaining = (total - index) / rate if rate > 0 else 0.0
percent = 100 * index / total if total else 100.0
message = (
f"{description}: {index}/{total} ({percent:5.1f}%) "
f"elapsed {_format_duration(elapsed)}, "
f"ETA {_format_duration(remaining)}"
)
print(message, end="\r", file=sys.stderr, flush=True)
print(" " * 100, end="\r", file=sys.stderr, flush=True)
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser()
parser.add_argument("--checkpoint", required=True)
parser.add_argument("--data", required=True)
parser.add_argument("--output-dir", required=True)
parser.add_argument("--batch-size", type=int, default=4)
parser.add_argument("--max-length", type=int, default=1024)
parser.add_argument("--num-permutations", type=int, default=5)
parser.add_argument("--permutation-seed", type=int, default=123)
parser.add_argument("--invariance-tolerance", type=float, default=1e-3)
parser.add_argument(
"--audit-invariance",
action="store_true",
help="use FP32 and CUDA math SDPA for a strict numerical invariance audit",
)
parser.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu")
return parser.parse_args()
@torch.no_grad()
def predict(model, tokenizer, rows: list[dict[str, Any]], args, description: str):
collator_arguments = inspect.signature(CandidateIndependentCollator).parameters
collator_options = {}
if "reset_candidate_positions" in collator_arguments:
collator_options["reset_candidate_positions"] = getattr(
model, "reset_candidate_positions", True
)
if "isolate_candidates" in collator_arguments:
collator_options["isolate_candidates"] = getattr(model, "isolate_candidates", True)
collator = CandidateIndependentCollator(tokenizer, args.max_length, **collator_options)
loader = DataLoader(rows, batch_size=args.batch_size, collate_fn=collator)
probabilities, logits = {}, {}
if str(args.device).startswith("cuda"):
torch.cuda.synchronize()
started = time.perf_counter()
for batch in _progress_batches(loader, description):
example_ids = batch.pop("example_ids")
candidate_ids = batch.pop("candidate_ids")
output = model(**{key: value.to(args.device) for key, value in batch.items()})
batch_probabilities = output.log_probs.exp().cpu()
batch_logits = output.logits.cpu()
for index, example_id in enumerate(example_ids):
count = len(candidate_ids[index])
probabilities[example_id] = batch_probabilities[index, :count].tolist()
logits[example_id] = batch_logits[index, :count].tolist()
if str(args.device).startswith("cuda"):
torch.cuda.synchronize()
return probabilities, logits, time.perf_counter() - started, collator.rejected_overlength
def quality_metrics(rows, probabilities):
correct, nll, brier, confidence, entropy = [], [], [], [], []
by_count = defaultdict(list)
bins = [{"count": 0, "correct": 0.0, "confidence": 0.0} for _ in range(15)]
for row in rows:
probs, target = probabilities[row["id"]], row["target"]
predicted = max(range(len(probs)), key=probs.__getitem__)
reference = max(range(len(target)), key=target.__getitem__)
hit = float(predicted == reference)
row_nll = -sum(t * math.log(max(p, 1e-12)) for t, p in zip(target, probs))
correct.append(hit)
nll.append(row_nll)
brier.append(sum((p - t) ** 2 for p, t in zip(probs, target)))
confidence.append(max(probs))
entropy.append(-sum(t * math.log(t) for t in target if t > 0))
by_count[len(target)].append((hit, row_nll))
bin_index = min(int(max(probs) * 15), 14)
bins[bin_index]["count"] += 1
bins[bin_index]["correct"] += hit
bins[bin_index]["confidence"] += max(probs)
total = len(rows)
grouped = {
str(count): {
"accuracy": sum(value[0] for value in values) / len(values),
"nll": sum(value[1] for value in values) / len(values),
"count": len(values),
}
for count, values in sorted(by_count.items())
}
nonbinary = [value for count, values in by_count.items() if count > 2 for value in values]
ece = sum(
abs(item["correct"] / item["count"] - item["confidence"] / item["count"])
* item["count"]
/ total
for item in bins
if item["count"]
)
return {
"count": total,
"accuracy": sum(correct) / total,
"nll": sum(nll) / total,
"multiclass_brier": sum(brier) / total,
"top_label_ece_15": ece,
"average_confidence": sum(confidence) / total,
"mean_target_entropy": sum(entropy) / total,
"counts_by_candidates": dict(sorted(Counter(len(row["target"]) for row in rows).items())),
"by_candidate_count": grouped,
"macro_candidate_count_accuracy": sum(value["accuracy"] for value in grouped.values()) / len(grouped),
"macro_candidate_count_nll": sum(value["nll"] for value in grouped.values()) / len(grouped),
"nonbinary_accuracy": sum(value[0] for value in nonbinary) / len(nonbinary) if nonbinary else None,
"nonbinary_nll": sum(value[1] for value in nonbinary) / len(nonbinary) if nonbinary else None,
}
def js_divergence(first, second):
midpoint = [(a + b) / 2 for a, b in zip(first, second)]
def kl(left, right):
return sum(x * math.log(x / y) for x, y in zip(left, right) if x > 0)
return (kl(first, midpoint) + kl(second, midpoint)) / 2
def main(
model_class=CandidateIndependentGemma,
architecture: str = ARCHITECTURE,
) -> None:
args = parse_args()
if args.audit_invariance and str(args.device).startswith("cuda"):
# Optimized BF16 SDPA kernels may change reduction order when candidate
# blocks move physically. The audit path trades speed for determinism.
torch.backends.cuda.enable_flash_sdp(False)
torch.backends.cuda.enable_mem_efficient_sdp(False)
torch.backends.cuda.enable_math_sdp(True)
model, tokenizer = model_class.from_checkpoint(args.checkpoint, args.device)
if args.audit_invariance:
model = model.float()
model.eval()
rows = JsonlDecisionDataset(args.data).rows
original_probs, original_logits, elapsed, rejected = predict(
model, tokenizer, rows, args, "original"
)
restored_prob_runs, restored_logit_runs = [], []
permutation_accuracies = []
position_selections, position_available = Counter(), Counter()
for run_index in range(args.num_permutations):
permuted_rows, displayed_orders = [], {}
for row_index, row in enumerate(rows):
order = list(range(len(row["options"])))
permutation_row_index = row.get("_permutation_row_index", row_index)
random.Random(
f"{args.permutation_seed}:{run_index}:{permutation_row_index}:{row['id']}"
).shuffle(order)
displayed_orders[row["id"]] = [option_id(row["options"][i], row["id"], i) for i in order]
permuted_rows.append(permute_example(row, order))
predicted_probs, predicted_logits, duration, run_rejected = predict(
model,
tokenizer,
permuted_rows,
args,
f"permutation {run_index + 1}/{args.num_permutations}",
)
elapsed += duration
rejected += run_rejected
restored_probs, restored_logits, hits = {}, {}, []
for row in rows:
example_id = row["id"]
canonical_ids = [option_id(option, row["id"], index) for index, option in enumerate(row["options"])]
displayed_ids = displayed_orders[example_id]
probability_map = dict(zip(displayed_ids, predicted_probs[example_id]))
logit_map = dict(zip(displayed_ids, predicted_logits[example_id]))
restored_probs[example_id] = [probability_map[item] for item in canonical_ids]
restored_logits[example_id] = [logit_map[item] for item in canonical_ids]
display_probabilities = predicted_probs[example_id]
for position in range(len(display_probabilities)):
position_available[position] += 1
selected_position = max(
range(len(display_probabilities)), key=display_probabilities.__getitem__
)
position_selections[selected_position] += 1
prediction = max(
range(len(canonical_ids)), key=restored_probs[example_id].__getitem__
)
reference = max(range(len(row["target"])), key=row["target"].__getitem__)
hits.append(prediction == reference)
restored_prob_runs.append(restored_probs)
restored_logit_runs.append(restored_logits)
permutation_accuracies.append(sum(hits) / len(hits))
js_values, probability_changes, raw_logit_changes, centered_logit_changes, flips = [], [], [], [], []
output_rows = []
for row in rows:
example_id = row["id"]
base_probs, base_logits = original_probs[example_id], original_logits[example_id]
permutation_probs = [run[example_id] for run in restored_prob_runs]
permutation_logits = [run[example_id] for run in restored_logit_runs]
base_selected = max(range(len(base_probs)), key=base_probs.__getitem__)
selected = [max(range(len(values)), key=values.__getitem__) for values in permutation_probs]
flips.append(any(value != base_selected for value in selected))
js_values.extend(js_divergence(base_probs, values) for values in permutation_probs)
probability_changes.extend(
abs(a - b) for values in permutation_probs for a, b in zip(base_probs, values)
)
raw_logit_changes.extend(
abs(a - b) for values in permutation_logits for a, b in zip(base_logits, values)
)
base_mean = sum(base_logits) / len(base_logits)
base_centered = [value - base_mean for value in base_logits]
for values in permutation_logits:
run_mean = sum(values) / len(values)
centered = [value - run_mean for value in values]
centered_logit_changes.extend(
abs(a - b) for a, b in zip(base_centered, centered)
)
candidate_ids = [option_id(option, row["id"], index) for index, option in enumerate(row["options"])]
output_rows.append({
"id": example_id,
"group_id": str(row.get("group_id") or example_id),
"source": str(row.get("source") or "unknown"),
"candidate_ids": candidate_ids,
"targets": row["target"],
"original_probabilities": base_probs,
"original_logits": base_logits,
"permuted_probabilities": permutation_probs,
"permuted_logits": permutation_logits,
"selected_ids": [candidate_ids[base_selected], *[candidate_ids[i] for i in selected]],
})
metrics = quality_metrics(rows, original_probs)
maximum_probability_change = max(probability_changes, default=0.0)
maximum_raw_logit_drift = max(raw_logit_changes, default=0.0)
maximum_centered_logit_drift = max(centered_logit_changes, default=0.0)
metrics.update({
"architecture": architecture,
"examples_per_second": len(rows) * (1 + args.num_permutations) / elapsed,
"rejected_overlength": rejected,
"num_permutations": args.num_permutations,
"prediction_flip_rate": sum(flips) / len(flips) if flips else 0.0,
"mean_js_divergence": sum(js_values) / len(js_values) if js_values else 0.0,
"max_js_divergence": max(js_values, default=0.0),
"worst_permutation_accuracy": min(permutation_accuracies, default=metrics["accuracy"]),
"mean_absolute_probability_change": sum(probability_changes) / len(probability_changes) if probability_changes else 0.0,
"max_absolute_probability_change": maximum_probability_change,
"mean_absolute_raw_logit_drift": sum(raw_logit_changes) / len(raw_logit_changes) if raw_logit_changes else 0.0,
"max_absolute_raw_logit_drift": maximum_raw_logit_drift,
"mean_absolute_centered_logit_drift": sum(centered_logit_changes) / len(centered_logit_changes) if centered_logit_changes else 0.0,
"max_absolute_centered_logit_drift": maximum_centered_logit_drift,
"invariance_tolerance": args.invariance_tolerance,
"invariance_audit_mode": args.audit_invariance,
"invariance_passed": (
maximum_probability_change <= args.invariance_tolerance
and maximum_centered_logit_drift <= args.invariance_tolerance
if args.audit_invariance
else None
),
"position_wise_selection_frequency": {
str(key): value / max(sum(position_selections.values()), 1)
for key, value in sorted(position_selections.items())
},
"position_wise_selection_rate_when_available": {
str(key): position_selections[key] / value
for key, value in sorted(position_available.items())
},
})
output_dir = Path(args.output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
(output_dir / "metrics.json").write_text(json.dumps(metrics, indent=2) + "\n")
import transformers
(output_dir / "config.json").write_text(
json.dumps(
{
**vars(args),
"architecture": architecture,
"versions": {
"python": platform.python_version(),
"torch": torch.__version__,
"cuda": str(torch.version.cuda),
"transformers": transformers.__version__,
},
},
indent=2,
)
+ "\n"
)
with (output_dir / "predictions.jsonl").open("w") as handle:
for row in output_rows:
handle.write(json.dumps(row) + "\n")
print(json.dumps(metrics, indent=2))
if __name__ == "__main__":
main()