File size: 7,920 Bytes
262fa3f | 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 | """Training configuration dataclass for lr_whisper.
All hyperparameters are read from environment variables via ``TrainingConfig.from_env()``.
Use ``get_config()`` to obtain a process-wide singleton (lazy, never called at import time).
"""
import os
from dataclasses import dataclass, field
from typing import List, Optional
from utils import env_flag, normalize_train_mode
@dataclass
class TrainingConfig:
# Model / dataset
model_id: str = "Qwen/Qwen3-ASR-0.6B"
dataset_name: str = "openslr/librispeech_asr"
dataset_config: str = "clean"
# Training batch / epoch
batch_size: int = 4
grad_accum_steps: int = 4
num_epochs: int = 10
eval_samples: int = 1000
pretrain_eval_samples: int = 10
# Latent tokens
n_latent: int = 4
is_deq_train: bool = False
deq_train_iters: int = 15
halt_threshold: float = 0.0
latent_drop_prob: float = 0.15
latent_input_noise_std: float = 0.5
value_forced_neg_prob: float = 0.3
latent_use_bounded_delta: bool = True
latent_use_injection_gate: bool = True
latent_use_embedding_anchor: bool = True
train_max_samples: int = 0
checkpoint_prefix: str = ""
# Thought layout
thought_mode: str = "prefix" # "prefix" | "interleaved"
thought_group_size: int = 1 # words per NT token in interleaved mode
# Training mode
train_mode: str = "latent"
# Learning rates
lr_adapter: float = 1e-4
lr_scale: float = 5e-5
lr_baseline_ft: float = 1e-5
lr_prompt_tuning: float = 5e-4
lr_lora_r16: float = 1e-4
# Prompt text
user_prompt_text: str = "Transcribe the audio into text."
# Prompt-tuning hyperparameters
prompt_tuning_init_mode: str = "text"
prompt_tuning_num_virtual_tokens: int = 4
prompt_tuning_init_text: str = "Transcribe the audio into text."
# LoRA hyperparameters
lora_rank: int = 16
lora_alpha: int = 32
lora_dropout: float = 0.05
lora_target_modules: List[str] = field(
default_factory=lambda: [
"q_proj", "k_proj", "v_proj", "o_proj",
"gate_proj", "up_proj", "down_proj",
]
)
# Delta stabilisation / scale
delta_tanh_c: float = 5.0
scale_max: float = 3.0
scale_init: float = 0.2
# Regularisation weights
w_smooth: float = 0.0 # Disabled — replaced by w_cycle
w_cycle: float = 0.1 # Cycle consistency loss (vocabulary-independent) - main regularizer
w_value: float = 3.0
# Logging intervals
log_every: int = 50
grad_log_every: int = 50
# Derived flags (set in __post_init__)
use_latent_reasoning: bool = False
use_prompt_tuning: bool = False
use_lora_r16: bool = False
use_peft_mode: bool = False
def __post_init__(self) -> None:
valid_modes = {"latent", "baseline", "prompt_tuning", "lora_r16"}
if self.train_mode not in valid_modes:
raise ValueError(
f"Unsupported train_mode={self.train_mode!r}. "
f"Use one of: {sorted(valid_modes)}"
)
valid_thought_modes = {"prefix", "interleaved"}
if self.thought_mode not in valid_thought_modes:
raise ValueError(
f"Unsupported thought_mode={self.thought_mode!r}. "
f"Use one of: {sorted(valid_thought_modes)}"
)
if self.thought_group_size < 1:
raise ValueError(f"thought_group_size must be >= 1, got {self.thought_group_size}")
self.use_latent_reasoning = self.train_mode == "latent"
self.use_prompt_tuning = self.train_mode == "prompt_tuning"
self.use_lora_r16 = self.train_mode == "lora_r16"
self.use_peft_mode = self.use_lora_r16
@classmethod
def from_env(cls) -> "TrainingConfig":
"""Construct a ``TrainingConfig`` by reading environment variables."""
lora_modules_raw = os.getenv(
"LORA_TARGET_MODULES",
"q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj",
)
lora_target_modules = [x.strip() for x in lora_modules_raw.split(",") if x.strip()]
train_mode = normalize_train_mode(os.getenv("TRAIN_MODE", "latent"))
default_dataset = "openslr/librispeech_asr"
default_config = "clean"
if train_mode == "latent":
default_dataset = "SpeechTest/extreme_asr_pony"
default_config = "default"
return cls(
model_id=os.getenv("MODEL_ID", "Qwen/Qwen3-ASR-0.6B"),
dataset_name=os.getenv("DATASET_NAME", default_dataset),
dataset_config=os.getenv("DATASET_CONFIG", default_config),
batch_size=int(os.getenv("BATCH_SIZE", "4")),
grad_accum_steps=int(os.getenv("GRAD_ACCUM_STEPS", "4")),
num_epochs=int(os.getenv("NUM_EPOCHS", "10")),
eval_samples=int(os.getenv("EVAL_SAMPLES", "1000")),
pretrain_eval_samples=int(os.getenv("PRETRAIN_EVAL_SAMPLES", "10")),
n_latent=int(os.getenv("N_LATENT", "4")),
is_deq_train=os.getenv("IS_DEQ_TRAIN", "true").strip().lower() == "true",
deq_train_iters=int(os.getenv("DEQ_TRAIN_ITERS", "15")),
halt_threshold=float(os.getenv("HALT_THRESHOLD", "0.0")),
latent_drop_prob=float(os.getenv("LATENT_DROP_PROB", "0.15")),
latent_input_noise_std=float(os.getenv("LATENT_INPUT_NOISE_STD", "0.5")),
value_forced_neg_prob=float(os.getenv("VALUE_FORCED_NEG_PROB", "0.3")),
latent_use_bounded_delta=env_flag("LATENT_USE_BOUNDED_DELTA", True),
latent_use_injection_gate=env_flag("LATENT_USE_INJECTION_GATE", True),
latent_use_embedding_anchor=env_flag("LATENT_USE_EMBEDDING_ANCHOR", True),
train_max_samples=int(os.getenv("TRAIN_MAX_SAMPLES", "0")),
checkpoint_prefix=os.getenv("CHECKPOINT_PREFIX", "").strip(),
thought_mode=os.getenv("THOUGHT_MODE", "prefix").strip().lower(),
thought_group_size=int(os.getenv("THOUGHT_GROUP_SIZE", "1")),
train_mode=train_mode,
lr_adapter=float(os.getenv("LR_ADAPTER", "1e-4")),
lr_scale=float(os.getenv("LR_SCALE", "5e-5")),
lr_baseline_ft=float(os.getenv("LR_BASELINE_FT", "1e-5")),
lr_prompt_tuning=float(os.getenv("LR_PROMPT_TUNING", "5e-4")),
lr_lora_r16=float(os.getenv("LR_LORA_R16", "1e-4")),
user_prompt_text=os.getenv(
"USER_PROMPT_TEXT", "Transcribe the audio into text."
).strip(),
prompt_tuning_init_mode=os.getenv("PROMPT_TUNING_INIT_MODE", "text").strip().lower(),
prompt_tuning_num_virtual_tokens=int(
os.getenv("PROMPT_TUNING_NUM_VIRTUAL_TOKENS", "4")
),
prompt_tuning_init_text=os.getenv(
"PROMPT_TUNING_INIT_TEXT", "Transcribe the audio into text."
).strip(),
lora_rank=int(os.getenv("LORA_RANK", "16")),
lora_alpha=int(os.getenv("LORA_ALPHA", "32")),
lora_dropout=float(os.getenv("LORA_DROPOUT", "0.05")),
lora_target_modules=lora_target_modules,
delta_tanh_c=float(os.getenv("DELTA_TANH_C", "5.0")),
scale_max=float(os.getenv("SCALE_MAX", "3.0")),
scale_init=float(os.getenv("SCALE_INIT", "0.2")),
w_smooth=float(os.getenv("W_SMOOTH", "0.0")),
w_cycle=float(os.getenv("W_CYCLE", "0.1")),
w_value=float(os.getenv("W_VALUE", "3.0")),
log_every=int(os.getenv("LOG_EVERY", "50")),
grad_log_every=int(os.getenv("GRAD_LOG_EVERY", "50")),
)
_config: Optional[TrainingConfig] = None
def get_config() -> TrainingConfig:
"""Return the process-wide ``TrainingConfig`` singleton (lazy init from env)."""
global _config
if _config is None:
_config = TrainingConfig.from_env()
return _config
|