File size: 8,520 Bytes
1f88cea | 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 | import datetime
import os
import warnings
from pathlib import Path
import time
from typing import Optional
import pytorch_lightning as pl
import torch
from omegaconf import OmegaConf, listconfig
from pytorch_lightning import LightningModule
from pytorch_lightning.callbacks.model_checkpoint import ModelCheckpoint
from pytorch_lightning.loggers import WandbLogger
from pytorch_lightning.strategies import DDPStrategy
from boltzgen.task.task import Task
from boltzgen.task.train.data import DataConfig, TrainingDataModule
class Training(Task):
"""Training configuration."""
def __init__(
self,
data: DataConfig,
model: LightningModule,
output: str,
name: str,
trainer: Optional[dict] = None,
resume: Optional[str] = None,
pretrained: Optional[str] = None,
wandb: Optional[dict] = None,
disable_checkpoint: bool = False,
slurm: bool = False,
matmul_precision: Optional[str] = None,
find_unused_parameters: Optional[bool] = False,
ddp_timeout_seconds: Optional[int] = 6400,
save_every_n_train_steps: Optional[int] = None,
save_top_k: Optional[int] = 1,
validation_only: bool = False,
debug: bool = False,
strict_loading: bool = True,
metric_mode: str = "max",
) -> None:
"""Initialize training configuration.
Parameters
----------
data : DataConfig
The data configuration.
model : LightningModule
The model.
output : str
The output directory.
name : str
The output subdirectory and wandb run name.
trainer : Optional[dict], optional
The trainer configuration, by default None.
resume : Optional[str], optional
The resume checkpoint, by default None
pretrained : Optional[str], optional
The pretrained model, by default None
wandb : Optional[dict], optional
The wandb configuration, by default None
disable_checkpoint : bool, optional
Disable checkpoint, by default False
"""
if not isinstance(data, DataConfig):
data = DataConfig(**data)
self.data = data
self.model = model
self.output = output
self.name = name
self.trainer = trainer
self.resume = resume
self.pretrained = pretrained
self.wandb = wandb
self.disable_checkpoint = disable_checkpoint
self.slurm = slurm
self.matmul_precision = matmul_precision
self.find_unused_parameters = find_unused_parameters
self.ddp_timeout_seconds = ddp_timeout_seconds
self.save_every_n_train_steps = save_every_n_train_steps
self.save_top_k = save_top_k
self.validation_only = validation_only
self.debug = debug
self.strict_loading = strict_loading
self.metric_mode = metric_mode
def run(self, config: OmegaConf) -> None:
"""Run training.
Parameters
----------
config : OmegaConf
The configuration for the task, for bookkeeping.
"""
# Disable some warnings
warnings.filterwarnings(
"ignore", ".*when logging on epoch level in distributed setting.*"
)
# Experiment with this during training (high or medium)
if self.matmul_precision is not None:
torch.set_float32_matmul_precision(self.matmul_precision)
# Create trainer dict
if self.trainer is None:
self.trainer = {}
# Flip some arguments in debug mode
devices = self.trainer.get("devices", 1)
if self.debug:
self.slurm = False
if isinstance(devices, int):
devices = 1
elif isinstance(devices, (list, listconfig.ListConfig)):
devices = [devices[0]]
self.trainer["devices"] = devices
self.data.num_workers = 0
if self.wandb:
self.wandb = None
# slurm
if self.slurm:
self.trainer["devices"] = int(
os.environ.get("SLURM_NTASKS_PER_NODE", "auto")
)
self.trainer["num_nodes"] = int(os.environ.get("SLURM_NNODES", 1))
# Create objects
data_module = TrainingDataModule(self.data)
model_module = self.model
if self.pretrained and not self.resume:
file_path = self.pretrained
print(f"Loading model from {file_path}")
model_module = type(model_module).load_from_checkpoint(
file_path, map_location="cpu", strict=False, weights_only=False, **(model_module.hparams)
)
# Create checkpoint callback
callbacks = self.trainer.get("callbacks", [])
dirpath = f"{self.output}/{self.name}"
os.makedirs(dirpath, exist_ok=True)
if not self.disable_checkpoint:
if self.slurm:
jobid = os.environ.get("SLURM_JOB_ID", "")
dirpath = Path(self.output) / jobid
dirpath.mkdir(parents=True, exist_ok=True)
print(
"Configuring ModelCheckpoint for SLURM to directory: ", str(dirpath)
)
mc = ModelCheckpoint(
monitor="val/lddt",
save_top_k=self.save_top_k,
save_last=True,
mode=self.metric_mode,
every_n_epochs=1,
)
callbacks.append(mc)
# Create wandb logger
loggers = []
if self.wandb:
extra = {}
if self.slurm:
ckpt_path = (
"hpc"
if len([f for f in os.listdir(dirpath) if f.startswith("hpc")])
else None
)
extra = {"resume": "must"} if ckpt_path else {}
wdb_logger = WandbLogger(
name=self.name,
group=self.wandb["group"],
save_dir=dirpath,
dir=f".tmp_wandb/{time.time()}",
project=self.wandb["project"],
entity=self.wandb["entity"],
log_model=False,
id=os.environ.get("SLURM_JOB_ID", None) if self.slurm else None,
**extra,
)
loggers.append(wdb_logger)
# Save the config to wandb
# This fails when not on rank0, so just catching
try:
config_out = Path(wdb_logger.experiment.dir) / "run.yaml"
with Path.open(config_out, "w") as f:
OmegaConf.save(config, f)
wdb_logger.experiment.save(str(config_out))
except Exception: # noqa: BLE001, S110
pass
# Set up trainer
strategy = "auto"
if (isinstance(devices, int) and devices > 1) or (
isinstance(devices, (list, listconfig.ListConfig)) and len(devices) > 1
):
strategy = DDPStrategy(
find_unused_parameters=self.find_unused_parameters,
timeout=datetime.timedelta(seconds=self.ddp_timeout_seconds),
)
trainer = pl.Trainer(
default_root_dir=str(dirpath),
strategy=strategy,
callbacks=callbacks,
logger=loggers,
enable_checkpointing=not self.disable_checkpoint,
reload_dataloaders_every_n_epochs=1,
**self.trainer,
)
# Run training
if self.slurm:
# If we are on SLURM, we need to check if we have a checkpoint for this job
ckpt_path = (
"hpc"
if len([f for f in os.listdir(dirpath) if f.startswith("hpc")])
else None
)
# If none check with we have a resume checkpoint passed in
if ckpt_path is None:
ckpt_path = self.resume
else:
ckpt_path = self.resume
if not self.strict_loading:
model_module.strict_loading = False
if self.validation_only:
trainer.validate(
model_module,
datamodule=data_module,
ckpt_path=ckpt_path,
)
else:
trainer.fit(
model_module,
datamodule=data_module,
ckpt_path=ckpt_path,
)
|