File size: 11,884 Bytes
6e802a4 | 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 | # dataset.py: Text + think block + speech tokens -> one training sequence with per-region loss weights; collator.
import json
import logging
import torch
from torch.utils.data import Dataset
from dataclasses import dataclass
from typing import List, Dict, Any
from tokenizer import TokenRegistry, NUM_SPEECH_TOKENS
logger = logging.getLogger(__name__)
IGNORE_INDEX = -100
class MambaCoTTTSDataset(Dataset):
def __init__(
self,
data_path: str,
tokenizer,
registry: TokenRegistry,
stage: int = 1,
max_seq_len: int = 2048,
cfg_dropout: float = 0.0,
cot_mode: str = "prose",
):
assert stage in (1, 2, "auto"), f"stage must be 1, 2 or 'auto', got {stage}"
self.tokenizer = tokenizer
self.registry = registry
self.stage = stage
self.max_seq_len = max_seq_len
self.cfg_dropout = cfg_dropout
self.data: List[dict] = []
with open(data_path, "r") as f:
for line in f:
line = line.strip()
if line:
self.data.append(json.loads(line))
assert cot_mode in ("prose", "tags", "none"), (
f"cot_mode={cot_mode!r} is not supported by the flat "
f"MambaCoTTTSDataset. The P²-CoT plan tiers (plan / tags+plan / "
f"full) carry a per-position injection schedule and run on the "
f"delay/depth path only — launch with --depth_module (or "
f"--delay_pattern), which routes to delay_dataset.DelayMimiDataset."
if cot_mode in ("plan", "tags+plan", "full") else cot_mode)
self.cot_mode = cot_mode
if cot_mode != "prose":
for row in self.data:
if cot_mode == "none":
row["reasoning"] = ""
else:
row["reasoning"] = (
f"EMO={row.get('emotion_label') or 'neutral'} "
f"PACE={row.get('pace') or 'normal'} "
f"PITCH={row.get('pitch') or 'normal'}")
self.bos_id = registry.bos_id
self.eos_id = registry.eos_id
self.user_prompt_id = registry.user_prompt_id
self.think_start_id = registry.think_start_id
self.think_end_id = registry.think_end_id
self.audio_start_id = registry.audio_start_id
self.audio_end_id = registry.audio_end_id
logger.info(
f"Loaded {len(self.data)} samples from {data_path} (stage={stage}, "
f"cot_mode={cot_mode}, max_seq_len={max_seq_len})"
)
def __len__(self) -> int:
return len(self.data)
def __getitem__(self, idx: int) -> dict:
item = self.data[idx]
if self.stage == "auto":
row_stage = 2 if str(item.get("reasoning", "") or "").strip() else 1
else:
row_stage = self.stage
prompt_text: str = item["prompt"]
speech_values: List[int] = item["speech_tokens"]
residual_codes = item.get("residual_codes", None)
speaker_id = item.get("speaker_id", None)
clamped = False
for i, v in enumerate(speech_values):
if v < 0 or v >= NUM_SPEECH_TOKENS:
speech_values[i] = max(0, min(v, NUM_SPEECH_TOKENS - 1))
clamped = True
if clamped:
logger.warning(f"Sample {idx}: speech_tokens contained out-of-range values, clamped to [0, {NUM_SPEECH_TOKENS - 1}]")
speech_ids = [self.registry.speech_value_to_id(v) for v in speech_values]
if self.cfg_dropout > 0.0 and torch.rand(1).item() < self.cfg_dropout:
prompt_ids = []
else:
prompt_ids = self.tokenizer.encode(prompt_text, add_special_tokens=False)
reasoning_ids = []
if row_stage == 2:
reasoning_text = item.get("reasoning", "")
if reasoning_text:
reasoning_clean = reasoning_text.strip()
if reasoning_clean.startswith("<THINK>"):
reasoning_clean = reasoning_clean[len("<THINK>"):].strip()
if reasoning_clean.endswith("</THINK>"):
reasoning_clean = reasoning_clean[:-len("</THINK>")].strip()
if reasoning_clean:
reasoning_ids = self.tokenizer.encode(reasoning_clean, add_special_tokens=False)
if row_stage == 1:
sequence = (
[self.bos_id, self.user_prompt_id]
+ prompt_ids
+ [self.audio_start_id]
+ speech_ids
+ [self.audio_end_id, self.eos_id]
)
prompt_end = 2 + len(prompt_ids)
reasoning_start = prompt_end
reasoning_end = prompt_end
speech_start = prompt_end + 1
speech_end = speech_start + len(speech_ids)
else:
sequence = (
[self.bos_id, self.user_prompt_id]
+ prompt_ids
+ [self.think_start_id]
+ reasoning_ids
+ [self.think_end_id]
+ [self.audio_start_id]
+ speech_ids
+ [self.audio_end_id, self.eos_id]
)
prompt_end = 2 + len(prompt_ids)
reasoning_start = prompt_end + 1
reasoning_end = reasoning_start + len(reasoning_ids)
speech_start = reasoning_end + 2
speech_end = speech_start + len(speech_ids)
if len(sequence) > self.max_seq_len:
overflow = len(sequence) - self.max_seq_len
max_speech_trim = len(speech_ids)
speech_trim = min(overflow, max_speech_trim)
speech_ids_trimmed = speech_ids[:len(speech_ids) - speech_trim]
overflow -= speech_trim
reasoning_ids_trimmed = reasoning_ids
if overflow > 0 and row_stage == 2:
max_reasoning_trim = len(reasoning_ids)
reasoning_trim = min(overflow, max_reasoning_trim)
reasoning_ids_trimmed = reasoning_ids[:len(reasoning_ids) - reasoning_trim]
overflow -= reasoning_trim
if overflow > 0:
logger.warning(
f"Sample {idx}: sequence still {overflow} tokens over max_seq_len "
f"after truncating speech and reasoning"
)
if row_stage == 1:
sequence = (
[self.bos_id, self.user_prompt_id]
+ prompt_ids
+ [self.audio_start_id]
+ speech_ids_trimmed
+ [self.audio_end_id, self.eos_id]
)
prompt_end = 2 + len(prompt_ids)
reasoning_start = prompt_end
reasoning_end = prompt_end
speech_start = prompt_end + 1
speech_end = speech_start + len(speech_ids_trimmed)
else:
sequence = (
[self.bos_id, self.user_prompt_id]
+ prompt_ids
+ [self.think_start_id]
+ reasoning_ids_trimmed
+ [self.think_end_id]
+ [self.audio_start_id]
+ speech_ids_trimmed
+ [self.audio_end_id, self.eos_id]
)
prompt_end = 2 + len(prompt_ids)
reasoning_start = prompt_end + 1
reasoning_end = reasoning_start + len(reasoning_ids_trimmed)
speech_start = reasoning_end + 2
speech_end = speech_start + len(speech_ids_trimmed)
return {
"input_ids": torch.tensor(sequence, dtype=torch.long),
"prompt_end": prompt_end,
"reasoning_start": reasoning_start,
"reasoning_end": reasoning_end,
"speech_start": speech_start,
"speech_end": speech_end,
"residual_codes": residual_codes,
"speaker_id": speaker_id,
}
@dataclass
class MambaCoTDataCollator:
pad_token_id: int
registry: TokenRegistry
stage: int = 1
lambda_r: float = 0.3
lambda_s: float = 1.0
semantic_weight: float = 1.0
eos_weight: float = 1.0
level_cycle: int = 0
def __call__(self, batch: List[Dict[str, Any]]) -> Dict[str, torch.Tensor]:
input_ids_list = [item["input_ids"] for item in batch]
input_ids = torch.nn.utils.rnn.pad_sequence(
input_ids_list, batch_first=True, padding_value=self.pad_token_id
)
B, T_max = input_ids.shape
labels = input_ids.clone()
loss_weights = torch.ones(B, T_max, dtype=torch.float32)
for i, item in enumerate(batch):
seq_len = len(item["input_ids"])
prompt_end = item["prompt_end"]
labels[i, :prompt_end] = IGNORE_INDEX
if seq_len < T_max:
labels[i, seq_len:] = IGNORE_INDEX
loss_weights[i, :prompt_end] = 0.0
if seq_len < T_max:
loss_weights[i, seq_len:] = 0.0
if self.stage == 2 or self.stage == "auto":
r_start = item["reasoning_start"]
r_end = item["reasoning_end"]
s_start = item["speech_start"]
s_end = item["speech_end"]
think_delim_start = max(prompt_end, r_start - 1)
think_delim_end = min(seq_len, r_end + 1)
loss_weights[i, think_delim_start:think_delim_end] = self.lambda_r
loss_weights[i, s_start - 1:seq_len] = self.lambda_s
if self.semantic_weight != 1.0 and self.level_cycle > 0:
s_start = item["speech_start"]
s_end = min(item["speech_end"], seq_len)
for pos in range(s_start, s_end, self.level_cycle):
loss_weights[i, pos] *= self.semantic_weight
if self.eos_weight != 1.0:
s_end = item["speech_end"]
for pos in (s_end, s_end + 1):
if pos < seq_len:
loss_weights[i, pos] *= self.eos_weight
speech_mask = self.registry.build_speech_mask(input_ids) & (labels != IGNORE_INDEX)
attention_mask = (input_ids != self.pad_token_id).long()
has_residual = any(item.get("residual_codes") is not None for item in batch)
residual_codes_tensor = None
if has_residual:
num_layers = 0
for item in batch:
if item.get("residual_codes") is not None:
num_layers = len(item["residual_codes"])
break
if num_layers > 0:
residual_codes_tensor = torch.full((B, num_layers, T_max), -100, dtype=torch.long)
for i, item in enumerate(batch):
rc = item.get("residual_codes")
if rc is not None:
s_start = item["speech_start"]
for layer_idx, layer_codes in enumerate(rc):
for t_idx, code in enumerate(layer_codes):
pos = s_start + t_idx
if pos < T_max:
residual_codes_tensor[i, layer_idx, pos] = code
speaker_ids = None
if any(item.get("speaker_id") is not None for item in batch):
speaker_ids = torch.tensor(
[item.get("speaker_id", 0) for item in batch], dtype=torch.long
)
return {
"input_ids": input_ids,
"labels": labels,
"attention_mask": attention_mask,
"speech_mask": speech_mask,
"loss_weights": loss_weights,
"residual_codes": residual_codes_tensor,
"speaker_ids": speaker_ids,
}
|