File size: 3,530 Bytes
32da3e8 | 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 | """Shared config dataclasses used across all training scripts."""
from __future__ import annotations
from dataclasses import dataclass, field
from typing import Any, Dict, List, Optional, Tuple
from omegaconf import MISSING
@dataclass
class ModelConfig:
"""Generic model configuration for instantiate_from_config().
Used for stage_1 (RAE) and stage_2 (DiT) model definitions.
The params dict is passed as kwargs to the target class constructor.
"""
target: str = ""
params: Dict[str, Any] = field(default_factory=dict)
ckpt: Optional[str] = None
@dataclass
class MiscConfig:
"""Miscellaneous model-related parameters."""
latent_size: List[int] = field(default_factory=lambda: [768, 16, 16]) # [C, H, W]
num_classes: int = 1000
time_dist_shift_dim: int = 196608 # 16*16*768
time_dist_shift_base: int = 4096
@dataclass
class OptimizerConfig:
"""Optimizer configuration (shared across all training)."""
type: str = "adamw" # "adamw", "gmuon"
lr: float = 2.0e-4
betas: Tuple[float, float] = (0.9, 0.95)
weight_decay: float = 0.0
eps: float = 1e-8
# GMuon-specific
momentum: float = 0.95
nesterov: bool = True
adamw_lr: Optional[float] = None
ns_use_kernels: bool = False
ns_coefficients_preset: str = "POLAR_EXPRESS_COEFFICIENTS"
@dataclass
class SchedulerConfig:
"""LR scheduler configuration."""
type: str = "cosine" # "cosine" or "linear"
warmup_epochs: float = 1.0
warmup_steps: Optional[int] = None
warmup_from_zero: bool = True
decay_end_epoch: float = 16.0
decay_end_steps: Optional[int] = None
base_lr: float = 2.0e-4
final_lr: float = 2.0e-5
@dataclass
class DatasetConfig:
"""Dataset configuration (shared across all training)."""
target: str = "imagenet"
type: str = "hf" # ["hf", "wds"]
data_dir: str = "./data"
split: Any = "train"
condition_type: Optional[str] = None # "label", "text", or "nwm"
shared_tmpdir: str = "~/tmp"
# WDS-specific
shuffle_buffer: int = 10000
seed: int = 42
# Free-form per-task params (e.g., nwm: context_size, len_traj_pred, ...)
params: Optional[Dict[str, Any]] = None
# Multi-subset wds (e.g. blip3o splits, or 'root' for flat 256 pools)
splits: Optional[List[str]] = None
subsets: Optional[List[str]] = None
# Multi-source mix: list of nested dataset configs each with a `weight`.
# Entries are passed verbatim to prepare_unified_dataloader recursively.
mix: Optional[List[Any]] = None
@dataclass
class EvalConfig:
"""Evaluation configuration.
eval.datasets.{name}.reference_npz, eval.datasets.{name}.metrics
"""
eval_interval: int = 5000
eval_model: bool = False # Eval non-EMA model too
eval_dir: str = MISSING # directory for eval CSVs, e.g. "experiments/<user>/evals/stage2"
datasets: Optional[Dict[str, Any]] = None
@dataclass
class TrainingConfig:
"""Base training configuration (shared across all)."""
epochs: int = 16
batch_size: int = 32
global_batch_size: Optional[int] = None
num_workers: int = 4
global_seed: int = 0
ema_decay: float = 0.9995
clip_grad: Optional[float] = None
log_interval: int = 100
checkpoint_interval: int = 4
sample_every: int = 2500
virtual_epoch_steps: Optional[int] = None
grad_accum_steps: int = 1
optimizer: OptimizerConfig = field(default_factory=OptimizerConfig)
scheduler: Optional[SchedulerConfig] = None
image_size: int = 256
|