File size: 14,203 Bytes
7e6c03a | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 | """
Multi-Hop task generator for temporal reasoning dataset.
Multi-hop inter-task: Explicit two-step or three-step temporal reasoning
over combinations of order, duration, volume, silence, and count.
Example chains:
- "What sound occurs after the longest sound?" (order + duration)
- "What sound occurs before the loudest sound?" (order + volume)
- "How many sounds occur after the longest silence?" (count + silence)
"""
import random
from collections import Counter
from pathlib import Path
from typing import Dict, List, Optional
from utils import (
setup_logger,
set_random_seed,
concatenate_to_target_duration,
generate_controlled_gap_durations,
get_lufs_loudness,
create_preprocessed_dataset,
)
from tasks.multihop_base import MultihopBaseGenerator
class MultiHopTaskGenerator(MultihopBaseGenerator):
"""Generates multi_hop task dataset samples."""
TASK_NAME = "multi_hop"
def __init__(self, config: dict, logger=None):
super().__init__(config, logger)
preprocessed_path = self.task_config.get(
"preprocessed_data_path",
config["tasks"].get("duration", {}).get("preprocessed_data_path", ""),
)
self.dataset = create_preprocessed_dataset(config, preprocessed_path=preprocessed_path)
def generate_sample(
self,
sample_id: int,
target_duration_seconds: float = None,
question_type: str = None,
) -> Optional[Dict]:
"""
Generate a single multi_hop sample.
Builds a rich scene with duration, volume, and silence variation,
then asks questions requiring chained reasoning.
"""
n_events = random.randint(
self.task_config.get("min_events", 5),
self.task_config.get("max_events", 8),
)
# Allow some repeats for count_before_loudest / count_after_longest
n_unique = random.randint(
max(3, n_events // 2), min(n_events, len(self.dataset.CATEGORIES))
)
categories_pool = self.dataset.sample_categories(n_unique)
categories = list(categories_pool)
while len(categories) < n_events:
categories.append(random.choice(categories_pool))
random.shuffle(categories)
# Volume variation
volume_range = self.task_config.get("volume_range_db", [-10, 6])
volume_levels = [
round(random.uniform(volume_range[0], volume_range[1]), 1)
for _ in range(n_events)
]
# Ensure min volume difference
min_diff = self.task_config.get("min_volume_diff_db", 4.0)
vmin, vmax = min(volume_levels), max(volume_levels)
if vmax - vmin < min_diff:
idx_min = volume_levels.index(vmin)
volume_levels[idx_min] = vmax - min_diff
from pydub import AudioSegment as PydubSegment
source_files = []
audio_segments = []
effective_durations_ms = []
measured_loudness = []
for i, cat in enumerate(categories):
fname, fpath, eff_dur = self.dataset.sample_file_from_category_with_duration(cat)
audio = self.audio_processor.load_audio(fpath)
audio = concatenate_to_target_duration(audio, max(self.source_clip_duration, eff_dur))
audio = audio.apply_gain(volume_levels[i])
loudness = get_lufs_loudness(audio)
audio_segments.append(audio)
source_files.append(fname)
effective_durations_ms.append(int(eff_dur * 1000))
measured_loudness.append(loudness)
# Gaps
num_gaps = n_events - 1
gap_min = self.task_config.get("min_gap_ms", 500)
gap_max = self.task_config.get("max_gap_ms", 3000)
if num_gaps > 0:
gap_durations = generate_controlled_gap_durations(
num_gaps, min_gap_ms=gap_min, max_gap_ms=gap_max,
gap_multiplier=2.5,
)
else:
gap_durations = []
# Assemble
result = audio_segments[0]
for i in range(1, n_events):
result = result + PydubSegment.silent(duration=gap_durations[i - 1])
result = result + audio_segments[i]
output_path = self.audio_output / f"{sample_id}.wav"
result.export(str(output_path), format="wav")
if question_type is None:
question_type = random.choice(self.task_config["question_types"])
mcq_data, open_data, q_meta = self._generate_question(
question_type, categories, effective_durations_ms,
measured_loudness, volume_levels, gap_durations
)
if mcq_data is None:
return None
metadata = {
"id": sample_id,
"audio_path": str(output_path.relative_to(self.output_base.parent)),
"n_events": n_events,
"categories": categories,
"source_files": source_files,
"volume_levels_db": volume_levels,
"effective_durations_ms": effective_durations_ms,
"measured_loudness_lufs": measured_loudness,
"gap_durations_ms": gap_durations,
"question_type": question_type,
"target_duration_s": target_duration_seconds,
"actual_duration_s": len(result) / 1000.0,
"mcq_question": mcq_data["question"],
"mcq_options": mcq_data["options"],
"mcq_correct_answer": mcq_data["correct_answer"],
"open_text_question": open_data["question"],
"open_text_answer": open_data["correct_answer"],
**q_meta,
}
self.logger.info(
f"Generated multi_hop sample {sample_id}: "
f"{n_events} events, type={question_type}"
)
return metadata
def _generate_question(
self, question_type, categories, durations_ms,
loudness_levels, volume_levels, gap_durations
):
"""Generate multi-hop chained reasoning question."""
n = len(categories)
# ---- Step 1: Identify the "pivot" event based on first property ----
if question_type in ("after_longest", "before_longest"):
pivot_idx = durations_ms.index(max(durations_ms))
pivot_label = "longest sound"
elif question_type == "after_shortest":
pivot_idx = durations_ms.index(min(durations_ms))
pivot_label = "shortest sound"
elif question_type == "before_loudest":
pivot_idx = loudness_levels.index(max(loudness_levels))
pivot_label = "loudest sound"
elif question_type == "after_longest_gap":
if not gap_durations:
return None, None, {}
longest_gap_idx = gap_durations.index(max(gap_durations))
# Events after the longest gap
after_idx = longest_gap_idx + 1
events_after = categories[after_idx:]
count_after = len(events_after)
# Pick template variant — count or identify
if random.random() < 0.5 and events_after:
# Identify first event after longest gap
correct = categories[after_idx]
mcq_text = self.task_config["mcq_questions"]["after_longest_gap"]
open_text = self.task_config["open_text_questions"]["after_longest_gap"]
# If template asks "how many" use count, else use category
if "how many" in mcq_text.lower():
mcq_data = self.question_generator.generate_count_mcq(
mcq_text, count_after, self.dataset.CATEGORIES
)
open_data = self.question_generator.generate_count_open_text(
open_text, count_after
)
else:
mcq_data = self.question_generator.generate_category_mcq(
mcq_text, correct, categories, self.dataset.CATEGORIES
)
open_data = self.question_generator.generate_category_open_text(
open_text, correct
)
q_meta = {"longest_gap_idx": longest_gap_idx, "correct_value": correct,
"count_after_longest_gap": count_after}
return mcq_data, open_data, q_meta
else:
mcq_text = f"How many sounds occur after the longest silence?"
open_text = f"How many sounds occur after the longest silence?"
mcq_data = self.question_generator.generate_count_mcq(
mcq_text, count_after, self.dataset.CATEGORIES
)
open_data = self.question_generator.generate_count_open_text(
open_text, count_after
)
q_meta = {"longest_gap_idx": longest_gap_idx,
"count_after_longest_gap": count_after}
return mcq_data, open_data, q_meta
elif question_type in ("count_before_loudest", "count_after_longest"):
if question_type == "count_before_loudest":
pivot_idx = loudness_levels.index(max(loudness_levels))
region = categories[:pivot_idx]
else:
pivot_idx = durations_ms.index(max(durations_ms))
region = categories[pivot_idx + 1:]
# Count occurrences of a target sound in region
unique_in_region = list(set(region))
if unique_in_region:
target_sound = random.choice(unique_in_region)
else:
target_sound = random.choice(list(set(categories)))
count = region.count(target_sound)
mcq_text = self.task_config["mcq_questions"][question_type].format(target_sound=target_sound)
open_text = self.task_config["open_text_questions"][question_type].format(target_sound=target_sound)
mcq_data = self.question_generator.generate_count_mcq(
mcq_text, count, self.dataset.CATEGORIES
)
open_data = self.question_generator.generate_count_open_text(
open_text, count
)
q_meta = {"target_sound": target_sound, "correct_count": count,
"pivot_index": pivot_idx}
return mcq_data, open_data, q_meta
elif question_type in ("overlap_after_anchor", "loudest_after_anchor",
"longest_before_anchor"):
# Pick anchor
anchor_idx = random.randint(1, n - 2) if n > 2 else 0
anchor_sound = categories[anchor_idx]
if question_type == "loudest_after_anchor":
subset = [(i, categories[i], loudness_levels[i])
for i in range(anchor_idx + 1, n)]
if not subset:
return None, None, {}
best = max(subset, key=lambda x: x[2])
correct = best[1]
elif question_type == "longest_before_anchor":
subset = [(i, categories[i], durations_ms[i])
for i in range(0, anchor_idx)]
if not subset:
return None, None, {}
best = max(subset, key=lambda x: x[2])
correct = best[1]
elif question_type == "overlap_after_anchor":
# In sequential audio, no true overlap — use event right after anchor
if anchor_idx + 1 < n:
correct = categories[anchor_idx + 1]
else:
return None, None, {}
else:
return None, None, {}
mcq_text = self.task_config["mcq_questions"][question_type].format(anchor_sound=anchor_sound)
open_text = self.task_config["open_text_questions"][question_type].format(anchor_sound=anchor_sound)
present = categories[anchor_idx + 1:] if "after" in question_type else categories[:anchor_idx]
mcq_data = self.question_generator.generate_category_mcq(
mcq_text, correct, present or categories, self.dataset.CATEGORIES
)
open_data = self.question_generator.generate_category_open_text(
open_text, correct
)
q_meta = {"anchor_sound": anchor_sound, "correct_category": correct}
return mcq_data, open_data, q_meta
else:
# Default: not one of the special cases above
return None, None, {}
# ---- Step 2: Get the event after/before the pivot ----
if "after" in question_type:
target_idx = pivot_idx + 1
if target_idx >= n:
return None, None, {}
else: # "before"
target_idx = pivot_idx - 1
if target_idx < 0:
return None, None, {}
correct = categories[target_idx]
mcq_text = self.task_config["mcq_questions"][question_type]
open_text = self.task_config["open_text_questions"][question_type]
mcq_data = self.question_generator.generate_category_mcq(
mcq_text, correct, categories, self.dataset.CATEGORIES
)
open_data = self.question_generator.generate_category_open_text(
open_text, correct
)
q_meta = {
"pivot_index": pivot_idx,
"pivot_category": categories[pivot_idx],
"correct_category": correct,
}
return mcq_data, open_data, q_meta
def main(config_path: str = None):
import yaml
if config_path is None:
config_path = Path(__file__).parent.parent / "config.yaml"
with open(config_path, "r") as f:
config = yaml.safe_load(f)
set_random_seed(config["random_seed"])
logger = setup_logger(
"multi_hop_task",
log_file=str(Path(config["output"]["base_path"]) / config["logging"]["log_file"]),
level=config["logging"]["level"],
console_output=config["logging"]["console_output"],
)
generator = MultiHopTaskGenerator(config, logger)
generator.generate_dataset()
if __name__ == "__main__":
main()
|