ai-content-maker/.venv/Lib/site-packages/trainer/trainer.py

2169 lines
88 KiB
Python
Raw Blame History

This file contains invisible Unicode characters

This file contains invisible Unicode characters that are indistinguishable to humans but may be processed differently by a computer. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

# -*- coding: utf-8 -*-
import gc
import importlib
import logging
import os
import platform
import shutil
import sys
import time
import traceback
from contextlib import nullcontext
from dataclasses import dataclass, field
from inspect import signature
from typing import Callable, Dict, List, Tuple, Union
import torch
import torch.distributed as dist
from coqpit import Coqpit
from torch import nn
from torch.nn.parallel import DistributedDataParallel as DDP_th
from torch.utils.data import DataLoader
from trainer.analytics import ping_training_run
from trainer.callbacks import TrainerCallback
from trainer.generic_utils import (
KeepAverage,
count_parameters,
get_experiment_folder_path,
get_git_branch,
isimplemented,
remove_experiment_folder,
set_partial_state_dict,
to_cuda,
)
from trainer.io import (
copy_model_files,
get_last_checkpoint,
load_fsspec,
save_best_model,
save_checkpoint,
)
from trainer.logging import ConsoleLogger, DummyLogger, logger_factory
from trainer.trainer_utils import (
get_optimizer,
get_scheduler,
is_apex_available,
print_training_env,
setup_torch_training_env,
)
from trainer.utils.cuda_memory import cuda_meminfo, should_reduce_batch_size
from trainer.utils.distributed import (
get_rank,
init_distributed,
rank_zero_logger_info,
rank_zero_only,
)
logger = logging.getLogger("trainer")
if is_apex_available():
from apex import amp # pylint: disable=import-error
@dataclass
class TrainerConfig(Coqpit):
"""Config fields tweaking the Trainer for a model.
A ````ModelConfig```, by inheriting ```TrainerConfig``` must be defined for using 👟.
Inherit this by a new model config and override the fields as needed.
All the fields can be overridden from comman-line as ```--coqpit.arg_name=value```.
Example::
Run the training code by overriding the ```lr``` and ```plot_step``` fields.
>>> python train.py --coqpit.plot_step=22 --coqpit.lr=0.001
Defining a model using ```TrainerConfig```.
>>> from trainer import TrainerConfig
>>> class MyModelConfig(TrainerConfig):
... optimizer: str = "Adam"
... lr: float = 0.001
... epochs: int = 1
... ...
>>> class MyModel(nn.module):
... def __init__(self, config):
... ...
>>> model = MyModel(MyModelConfig())
"""
# Fields for the run
output_path: str = field(default="output")
logger_uri: str = field(
default=None,
metadata={
"help": "URI to save training artifacts by the logger. If not set, logs will be saved in the output_path. Defaults to None"
},
)
run_name: str = field(default="run", metadata={"help": "Name of the run. Defaults to 'run'"})
project_name: str = field(default=None, metadata={"help": "Name of the project. Defaults to None"})
run_description: str = field(
default="🐸Coqui trainer run.",
metadata={"help": "Notes and description about the run. Defaults to '🐸Coqui trainer run.'"},
)
# Fields for logging
print_step: int = field(
default=25, metadata={"help": "Print training stats on the terminal every print_step steps. Defaults to 25"}
)
plot_step: int = field(
default=100, metadata={"help": "Plot training stats on the logger every plot_step steps. Defaults to 100"}
)
model_param_stats: bool = field(
default=False, metadata={"help": "Log model parameters stats on the logger dashboard. Defaults to False"}
)
wandb_entity: str = field(default=None, metadata={"help": "Wandb entity to log the run. Defaults to None"})
dashboard_logger: str = field(
default="tensorboard", metadata={"help": "Logger to use for the tracking dashboard. Defaults to 'tensorboard'"}
)
# Fields for checkpointing
save_on_interrupt: bool = field(
default=True, metadata={"help": "Save checkpoint on interrupt (Ctrl+C). Defaults to True"}
)
log_model_step: int = field(
default=None,
metadata={
"help": "Save checkpoint to the logger every log_model_step steps. If not defined `save_step == log_model_step`."
},
)
save_step: int = field(
default=10000, metadata={"help": "Save local checkpoint every save_step steps. Defaults to 10000"}
)
save_n_checkpoints: int = field(default=5, metadata={"help": "Keep n local checkpoints. Defaults to 5"})
save_checkpoints: bool = field(default=True, metadata={"help": "Save checkpoints locally. Defaults to True"})
save_all_best: bool = field(
default=False, metadata={"help": "Save all best checkpoints and keep the older ones. Defaults to False"}
)
save_best_after: int = field(
default=0, metadata={"help": "Wait N steps to save best checkpoints. Defaults to 0"}
)
target_loss: str = field(
default=None, metadata={"help": "Target loss name to select the best model. Defaults to None"}
)
# Fields for eval and test run
print_eval: bool = field(default=False, metadata={"help": "Print eval steps on the terminal. Defaults to False"})
test_delay_epochs: int = field(default=0, metadata={"help": "Wait N epochs before running the test. Defaults to 0"})
run_eval: bool = field(
default=True, metadata={"help": "Run evalulation epoch after training epoch. Defaults to True"}
)
run_eval_steps: int = field(
default=None,
metadata={
"help": "Run evalulation epoch after N steps. If None, waits until training epoch is completed. Defaults to None"
},
)
# Fields for distributed training
distributed_backend: str = field(
default="nccl", metadata={"help": "Distributed backend to use. Defaults to 'nccl'"}
)
distributed_url: str = field(
default="tcp://localhost:54321",
metadata={"help": "Distributed url to use. Defaults to 'tcp://localhost:54321'"},
)
# Fields for training specs
mixed_precision: bool = field(default=False, metadata={"help": "Use mixed precision training. Defaults to False"})
precision: str = field(
default="fp16",
metadata={
"help": "Precision to use in mixed precision training. `fp16` for float16 and `bf16` for bfloat16. Defaults to 'f16'"
},
)
epochs: int = field(default=1000, metadata={"help": "Number of epochs to train. Defaults to 1000"})
batch_size: int = field(default=32, metadata={"help": "Batch size to use. Defaults to 32"})
eval_batch_size: int = field(default=16, metadata={"help": "Batch size to use for eval. Defaults to 16"})
grad_clip: float = field(
default=0.0, metadata={"help": "Gradient clipping value. Disabled if <= 0. Defaults to 0.0"}
)
scheduler_after_epoch: bool = field(
default=True,
metadata={"help": "Step the scheduler after each epoch else step after each iteration. Defaults to True"},
)
# Fields for optimzation
lr: Union[float, List[float]] = field(
default=0.001, metadata={"help": "Learning rate for each optimizer. Defaults to 0.001"}
)
optimizer: Union[str, List[str]] = field(default=None, metadata={"help": "Optimizer(s) to use. Defaults to None"})
optimizer_params: Union[Dict, List[Dict]] = field(
default_factory=dict, metadata={"help": "Optimizer(s) arguments. Defaults to {}"}
)
lr_scheduler: Union[str, List[str]] = field(
default=None, metadata={"help": "Learning rate scheduler(s) to use. Defaults to None"}
)
lr_scheduler_params: Dict = field(
default_factory=dict, metadata={"help": "Learning rate scheduler(s) arguments. Defaults to {}"}
)
use_grad_scaler: bool = field(
default=False,
metadata={
"help": "Enable/disable gradient scaler explicitly. It is enabled by default with AMP training. Defaults to False"
},
)
allow_tf32: bool = field(
default=False,
metadata={
"help": "A bool that controls whether TensorFloat-32 tensor cores may be used in matrix multiplications on Ampere or newer GPUs. Default to False."
},
)
cudnn_enable: bool = field(default=True, metadata={"help": "Enable/disable cudnn explicitly. Defaults to True"})
cudnn_deterministic: bool = field(
default=False,
metadata={
"help": "Enable/disable deterministic cudnn operations. Set this True for reproducibility but it slows down training significantly. Defaults to False."
},
)
cudnn_benchmark: bool = field(
default=False,
metadata={
"help": "Enable/disable cudnn benchmark explicitly. Set this False if your input size change constantly. Defaults to False"
},
)
training_seed: int = field(
default=54321,
metadata={"help": "Global seed for torch, random and numpy random number generator. Defaults to 54321"},
)
@dataclass
class TrainerArgs(Coqpit):
"""Trainer arguments that can be accessed from the command line.
Examples::
>>> python train.py --restore_path /path/to/checkpoint.pth
"""
continue_path: str = field(
default="",
metadata={
"help": "Path to a training folder to continue training. Restore the model from the last checkpoint and continue training under the same folder."
},
)
restore_path: str = field(
default="",
metadata={
"help": "Path to a model checkpoit. Restore the model with the given checkpoint and start a new training."
},
)
best_path: str = field(
default="",
metadata={
"help": "Best model file to be used for extracting the best loss. If not specified, the latest best model in continue path is used"
},
)
use_ddp: bool = field(
default=False,
metadata={"help": "Use DDP in distributed training. It is to set in `distribute.py`. Do not set manually."},
)
use_accelerate: bool = field(default=False, metadata={"help": "Use HF Accelerate as the back end for training."})
grad_accum_steps: int = field(
default=1,
metadata={
"help": "Number of gradient accumulation steps. It is used to accumulate gradients over multiple batches."
},
)
overfit_batch: bool = field(default=False, metadata={"help": "Overfit a single batch for debugging."})
skip_train_epoch: bool = field(
default=False,
metadata={"help": "Skip training and only run evaluation and test."},
)
start_with_eval: bool = field(
default=False,
metadata={"help": "Start with evaluation and test."},
)
small_run: int = field(
default=None,
metadata={
"help": "Only use a subset of the samples for debugging. Set the number of samples to use. Defaults to None. "
},
)
gpu: int = field(
default=None, metadata={"help": "GPU ID to use if ```CUDA_VISIBLE_DEVICES``` is not set. Defaults to None."}
)
# only for DDP
rank: int = field(default=0, metadata={"help": "Process rank in a distributed training. Don't set manually."})
group_id: str = field(
default="", metadata={"help": "Process group id in a distributed training. Don't set manually."}
)
class Trainer:
def __init__( # pylint: disable=dangerous-default-value
self,
args: TrainerArgs,
config: Coqpit,
output_path: str,
c_logger: ConsoleLogger = None,
dashboard_logger: "Logger" = None,
model: nn.Module = None,
get_model: Callable = None,
get_data_samples: Callable = None,
train_samples: List = None,
eval_samples: List = None,
test_samples: List = None,
train_loader: DataLoader = None,
eval_loader: DataLoader = None,
training_assets: Dict = {},
parse_command_line_args: bool = True,
callbacks: Dict[str, Callable] = {},
gpu: int = None,
) -> None:
"""Simple yet powerful 🐸💬 TTS trainer for PyTorch. It can train all the available `tts` and `vocoder` models
or easily be customized.
Notes:
Supports Automatic Mixed Precision training. If `Apex` is availabe, it automatically picks that, else
it uses PyTorch's native `amp` module. `Apex` may provide more stable training in some cases.
Args:
args (Union[Coqpit, Namespace]): Training arguments parsed either from console by `argparse` or `TrainerArgs`
config object.
config (Coqpit): Model config object. It includes all the values necessary for initializing, training, evaluating
and testing the model.
output_path (str): Path to the output training folder. All the files are saved under thi path.
c_logger (ConsoleLogger, optional): Console logger for printing training status. If not provided, the default
console logger is used. Defaults to None.
dashboard_logger Union[TensorboardLogger, WandbLogger]: Dashboard logger. If not provided, the tensorboard logger is used.
Defaults to None.
model (nn.Module, optional): Initialized and ready-to-train model. If it is not defined, `Trainer`
initializes a model from the provided config. Defaults to None.
get_model (Callable):
A function that returns a model. It is used to initialize the model when `model` is not provided.
It either takes the config as the only argument or does not take any argument.
Defaults to None
get_data_samples (Callable):
A function that returns a list of training and evaluation samples. Used if `train_samples` and
`eval_samples` are None. Defaults to None.
train_samples (List):
A list of training samples used by the model's `get_train_data_loader` to init the `dataset` and the
`data_loader`. Defaults to None.
eval_samples (List):
A list of evaluation samples used by the model's `get_eval_data_loader` to init the `dataset` and the
`data_loader`. Defaults to None.
train_loader (DataLoader):
A pytorch data loader object for training epochs. Leave as None if you want it to be made during training. Defaults to None.
eval_loader (DataLoader):
A pytorch data loader object for evaluation epochs. Leave as None to be generated during training. Defaults to None.
test_samples (List):
A list of test samples used by the model's `get_test_data_loader` to init the `dataset` and the
`data_loader`. If None, the ```model.test_run()``` is expected to load the data. Defaults to None.
training_assets (Dict):
A dictionary of assets to be used at training and passed to the model's ```train_log(), eval_log(), get_data_loader()```
during training. It can include `AudioProcessor` or/and `Tokenizer`. Defaults to {}.
parse_command_line_args (bool):
If true, parse command-line arguments and update `TrainerArgs` and model `config` values. Set it
to false if you parse the arguments yourself. Defaults to True.
callbacks (Dict[str, Callable]):
A dictionary of callbacks to be used during training. The keys are the callback names and the values
gpu (int):
GPU ID to use for training If "CUDA_VISIBLE_DEVICES" is not set. Defaults to None.
Example::
Running trainer with a model.
>>> args = TrainerArgs(...)
>>> config = ModelConfig(...)
>>> model = Model(config)
>>> trainer = Trainer(args, config, output_path, model=model)
>>> trainer.fit()
TODO:
- Wrap model for not calling .module in DDP.
- Deepspeed integration
- Profiler integration.
- Overfitting to a batch.
- TPU training
"""
if parse_command_line_args:
# parse command-line arguments to override TrainerArgs()
args, coqpit_overrides = self.parse_argv(args)
# get ready for training and parse command-line arguments to override the model config
config, new_fields = self.init_training(args, coqpit_overrides, config)
elif args.continue_path or args.restore_path:
config, new_fields = self.init_training(args, {}, config)
else:
new_fields = {}
# set the output path
if args.continue_path:
# use the same path as the continuing run
output_path = args.continue_path
else:
# override the output path if it is provided
output_path = config.output_path if output_path is None else output_path
# create a new output folder name
output_path = get_experiment_folder_path(config.output_path, config.run_name)
os.makedirs(output_path, exist_ok=True)
# copy training assets to the output folder
copy_model_files(config, output_path, new_fields)
# init class members
self.args = args
self.config = config
self.output_path = output_path
self.training_assets = training_assets
self.grad_accum_steps = args.grad_accum_steps
self.overfit_batch = args.overfit_batch
self.skip_train_epoch = args.skip_train_epoch
self.start_with_eval = args.start_with_eval
assert self.grad_accum_steps > 0, " [!] grad_accum_steps must be greater than 0."
# setup logging
log_file = os.path.join(self.output_path, f"trainer_{args.rank}_log.txt")
self._setup_logger_config(log_file)
# setup training environment
self.use_cuda, self.num_gpus = self.setup_training_environment(args=args, config=config, gpu=gpu)
# init loggers
self.dashboard_logger, self.c_logger = self.init_loggers(self.config, output_path, dashboard_logger, c_logger)
# self.c_logger.logger = logger
if not self.config.log_model_step:
self.config.log_model_step = self.config.save_step
# make sure that start_with_eval is disabled if eval is disabled
if not self.config.run_eval and self.start_with_eval:
self.start_with_eval = False
self.total_steps_done = 0
self.epochs_done = 0
self.restore_step = 0
self.restore_epoch = 0
self.best_loss = {"train_loss": float("inf"), "eval_loss": float("inf") if self.config.run_eval else None}
self.train_loader = None
self.test_loader = None
self.eval_loader = None
self.keep_avg_train = None
self.keep_avg_eval = None
self.use_amp_scaler = (
self.use_cuda
if self.config.mixed_precision and self.config.precision == "fp16"
else self.config.use_grad_scaler
)
if train_samples is not None:
# use the provided samples
self.train_samples = train_samples
self.eval_samples = eval_samples
self.test_samples = test_samples
elif get_data_samples is not None:
# run `get_data_samples` to init the data samples
( # pylint: disable=unbalanced-tuple-unpacking
self.train_samples,
self.eval_samples,
self.test_samples,
) = self.run_get_data_samples(config, get_data_samples)
else:
# expecting to load the samples in `model.get_data_loader()`
self.train_samples = None
self.eval_samples = None
self.test_samples = None
# define custom train and eval loader
self.train_loader = train_loader
self.eval_loader = eval_loader
# only use a subset of the samples if small_run is set
self.setup_small_run(args.small_run)
# init the model
if model is None and get_model is None:
raise ValueError("[!] `model` and `get_model` cannot both be None.")
if model is not None:
self.model = model
else:
self.run_get_model(self.config, get_model)
# init model's training assets
if isimplemented(self.model, "init_for_training"):
self.model.init_for_training()
# setup criterion
self.criterion = self.get_criterion(self.model)
# DISTRUBUTED
if self.use_pt_ddp:
rank_zero_logger_info(" > Using PyTorch DDP", logger)
init_distributed(
args.rank,
self.num_gpus,
args.group_id,
self.config.distributed_backend,
self.config.distributed_url,
)
if self.use_cuda:
self.model.cuda()
if isinstance(self.criterion, list):
for criterion in self.criterion:
if isinstance(criterion, torch.nn.Module):
criterion.cuda()
else:
if isinstance(self.criterion, torch.nn.Module):
self.criterion.cuda()
# setup optimizer
self.optimizer = self.get_optimizer(self.model, self.config)
# If multiple-optimizer setup with grad accumulation and without custom optimize method raise an error
if (
self.grad_accum_steps != 1
and isinstance(self.optimizer, list)
and not isimplemented(self.model, "optimize")
):
raise ValueError(
" [!] Coqui Trainer does not support grad_accum_steps for multiple-optimizer setup, please set grad_accum_steps to 1 or implement in your model a custom method called ´optimize` that need to deal with dangling gradients in multiple-optimizer setup!"
)
# CALLBACK
self.callbacks = TrainerCallback()
self.callbacks.parse_callbacks_dict(callbacks)
self.callbacks.on_init_start(self)
# init AMP
if self.use_amp_scaler:
if self.use_apex:
self.scaler = None
self.model, self.optimizer = amp.initialize(self.model, self.optimizer, opt_level="O1")
self.scaler = torch.cuda.amp.GradScaler()
else:
self.scaler = None
# restore model
if self.args.restore_path:
(self.model, self.optimizer, self.scaler, self.restore_step, self.restore_epoch) = self.restore_model(
self.config, args.restore_path, self.model, self.optimizer, self.scaler
)
self.scaler = torch.cuda.amp.GradScaler()
# setup scheduler
self.scheduler = self.get_scheduler(self.model, self.config, self.optimizer)
self.scheduler = self.restore_scheduler(
self.scheduler, self.args, self.config, self.restore_epoch, self.restore_step
)
# DISTRIBUTED
if self.use_pt_ddp:
self.model = DDP_th(self.model, device_ids=[args.rank], output_device=args.rank)
# setup accelerator
self.setup_accelerate()
# count model size
num_params = count_parameters(self.model)
rank_zero_logger_info(f"\n > Model has {num_params} parameters", logger)
self.callbacks.on_init_end(self)
self.dashboard_logger.add_config(config)
self.save_training_script()
ping_training_run()
@property
def use_apex(self):
"""Return True if using APEX."""
return not self.args.use_accelerate and self._is_apex_available()
@property
def use_pt_ddp(self):
"""Return True if using PyTorch DDP."""
return self.num_gpus > 1 and not self.use_accelerate
@property
def use_accelerate(self):
"""Return True if using HF Accelerate."""
return self.args.use_accelerate
def setup_accelerate(self):
if self.use_accelerate:
self.model, self.optimizer, self.train_loader, self.scheduler, self.accelerator = self.init_accelerate(
model=self.model,
optimizer=self.optimizer,
training_dataloader=self.train_loader,
scheduler=self.scheduler,
grad_accum_steps=self.grad_accum_steps,
mixed_precision=self.config.mixed_precision,
precision=self.config.precision,
)
def prepare_accelerate_loader(self, data_loader):
"""Prepare the accelerator for the training."""
if self.use_accelerate:
return self.accelerator.prepare_data_loader(data_loader)
return data_loader
@staticmethod
def init_accelerate(model, optimizer, training_dataloader, scheduler, grad_accum_steps, mixed_precision, precision):
"""Setup HF Accelerate for the training."""
# check if accelerate is installed
try:
from accelerate import Accelerator # pylint:disable=import-outside-toplevel
except ImportError as e:
raise ImportError("Please install accelerate to use this feature.") from e
_precision = precision if precision is not None else "f16" if mixed_precision else None
if _precision == "float16":
_precision = "f16"
elif _precision == "float8":
_precision = "f8"
elif _precision == "bfloat16":
_precision = "bf16"
accelerator = Accelerator(gradient_accumulation_steps=grad_accum_steps, mixed_precision=_precision)
if isinstance(model, torch.nn.Module):
model = accelerator.prepare_model(model)
if isinstance(optimizer, dict):
for key, optim in optimizer.items():
optimizer[key] = accelerator.prepare_optimizer(optim)
elif isinstance(optimizer, list):
for i, optim in enumerate(optimizer):
optimizer[i] = accelerator.prepare_optimizer(optim)
elif optimizer is not None:
optimizer = accelerator.prepare_optimizer(optimizer)
if isinstance(training_dataloader, torch.utils.data.DataLoader):
training_dataloader = accelerator.prepare_data_loader(training_dataloader)
if isinstance(scheduler, dict):
for key, sched in scheduler.items():
scheduler[key] = accelerator.prepare_scheduler(sched)
elif isinstance(scheduler, list):
for i, sched in enumerate(scheduler):
scheduler[i] = accelerator.prepare_scheduler(sched)
elif scheduler is not None:
scheduler = accelerator.prepare_scheduler(scheduler)
return model, optimizer, training_dataloader, scheduler, accelerator
def save_training_script(self):
"""Save the training script to tracking dashboard and output path."""
file_path = sys.argv[0]
if os.path.isfile(file_path):
file_name = os.path.basename(file_path)
self.dashboard_logger.add_artifact(file_or_dir=file_path, name=file_name, artifact_type="file")
with open(file_path, "r", encoding="utf8") as f:
self.dashboard_logger.add_text("training-script", f"{f.read()}", 0)
shutil.copyfile(file_path, os.path.join(self.output_path, file_name))
@staticmethod
def parse_argv(args: Union[Coqpit, List]):
"""Parse command line arguments to init or override `TrainerArgs()`."""
if isinstance(args, Coqpit):
parser = args.init_argparse(arg_prefix="")
else:
train_config = TrainerArgs()
parser = train_config.init_argparse(arg_prefix="")
training_args, coqpit_overrides = parser.parse_known_args()
args.parse_args(training_args)
return args, coqpit_overrides
@staticmethod
def init_loggers(config: "Coqpit", output_path: str, dashboard_logger=None, c_logger=None):
"""Init console and dashboard loggers.
Use the given logger if passed externally else use config values to pick the right logger.
Return a dashboard logger only for the rank 0 process in DDP
Define a console logger for each process in DDP
Args:
config (Coqpit): Model config.
output_path (str): Output path to save the training artifacts.
dashboard_logger (DashboardLogger): Object passed to the trainer from outside.
c_logger (ConsoleLogger): Object passed to the trained from outside.
Returns:
Initialized dashboard_logger and console_logger objects.
"""
c_logger = ConsoleLogger() if c_logger is None else c_logger
# only allow dashboard logging for the main process in DDP mode
if get_rank() > 0:
return DummyLogger(), c_logger
if dashboard_logger is None:
dashboard_logger = logger_factory(config, output_path)
return dashboard_logger, c_logger
def setup_small_run(self, small_run: int = None):
"""Use a subset of samples for training, evaluation and testing."""
if small_run is not None:
logger.info("[!] Small Run, only using %i samples.", small_run)
self.train_samples = None if self.train_samples is None else self.train_samples[:small_run]
self.eval_samples = None if self.eval_samples is None else self.eval_samples[:small_run]
self.test_samples = None if self.test_samples is None else self.test_samples[:small_run]
@staticmethod
def init_training(args: TrainerArgs, coqpit_overrides: Dict, config: Coqpit = None):
"""Initialize training and update model configs from command line arguments.
Args:
args (argparse.Namespace or dict like): Parsed trainer arguments.
config_overrides (argparse.Namespace or dict like): Parsed config overriding arguments.
config (Coqpit): Model config. If none, it is generated from `args`. Defaults to None.
Returns:
config (Coqpit): Config paramaters.
"""
# set arguments for continuing training
if args.continue_path:
args.config_path = os.path.join(args.continue_path, "config.json")
args.restore_path, best_model = get_last_checkpoint(args.continue_path)
if not args.best_path:
args.best_path = best_model
# use the same config
if config:
config.load_json(args.config_path)
else:
coqpit = Coqpit()
coqpit.load_json(args.config_path)
# override config values from command-line args
# TODO: Maybe it is better to do it outside
if len(coqpit_overrides) > 0:
config.parse_known_args(coqpit_overrides, relaxed_parser=True)
# update the config.json fields and copy it to the output folder
new_fields = {}
if args.rank == 0:
if args.restore_path:
new_fields["restore_path"] = args.restore_path
new_fields["github_branch"] = get_git_branch()
return config, new_fields
@staticmethod
def setup_training_environment(args, config, gpu):
if platform.system() != "Windows":
# https://github.com/pytorch/pytorch/issues/973
import resource # pylint: disable=import-outside-toplevel
rlimit = resource.getrlimit(resource.RLIMIT_NOFILE)
resource.setrlimit(resource.RLIMIT_NOFILE, (4096, rlimit[1]))
# set and initialize Pytorch runtime
use_cuda, num_gpus = setup_torch_training_env(
args=args,
cudnn_enable=config.cudnn_enable,
cudnn_deterministic=config.cudnn_deterministic,
cudnn_benchmark=config.cudnn_benchmark,
use_ddp=args.use_ddp,
training_seed=config.training_seed,
allow_tf32=config.allow_tf32,
gpu=gpu if args.gpu is None else args.gpu,
)
print_training_env(args, config)
return use_cuda, num_gpus
@staticmethod
def run_get_model(config: Coqpit, get_model: Callable) -> nn.Module:
"""Run the `get_model` function and return the model.
Args:
config (Coqpit): Model config.
Returns:
nn.Module: initialized model.
"""
if len(signature(get_model).sig.parameters) == 1:
model = get_model(config)
else:
model = get_model()
return model
@staticmethod
def run_get_data_samples(config: Coqpit, get_data_samples: Callable) -> nn.Module:
if callable(get_data_samples):
if len(signature(get_data_samples).sig.parameters) == 1:
train_samples, eval_samples = get_data_samples(config)
else:
train_samples, eval_samples = get_data_samples()
return train_samples, eval_samples
return None, None
def restore_model(
self,
config: Coqpit,
restore_path: str,
model: nn.Module,
optimizer: torch.optim.Optimizer,
scaler: torch.cuda.amp.GradScaler = None,
) -> Tuple[nn.Module, torch.optim.Optimizer, torch.cuda.amp.GradScaler, int]:
"""Restore training from an old run. It restores model, optimizer, AMP scaler and training stats.
Args:
config (Coqpit): Model config.
restore_path (str): Path to the restored training run.
model (nn.Module): Model to restored.
optimizer (torch.optim.Optimizer): Optimizer to restore.
scaler (torch.cuda.amp.GradScaler, optional): AMP scaler to restore. Defaults to None.
Returns:
Tuple[nn.Module, torch.optim.Optimizer, torch.cuda.amp.GradScaler, int]: [description]
"""
def _restore_list_objs(states, obj):
if isinstance(obj, list):
for idx, state in enumerate(states):
obj[idx].load_state_dict(state)
elif isinstance(obj, dict):
for key, state in states.items():
obj[key].load_state_dict(state)
else:
obj.load_state_dict(states)
return obj
logger.info(" > Restoring from %s ...", os.path.basename(restore_path))
checkpoint = load_fsspec(restore_path, map_location="cpu")
try:
logger.info(" > Restoring Model...")
model.load_state_dict(checkpoint["model"])
logger.info(" > Restoring Optimizer...")
try:
optimizer = _restore_list_objs(checkpoint["optimizer"], optimizer)
except (KeyError, TypeError, RuntimeError):
logger.info(" > Optimizer is not compatible with the restored model.")
if "scaler" in checkpoint and self.use_amp_scaler and checkpoint["scaler"]:
logger.info(" > Restoring Scaler...")
scaler = _restore_list_objs(checkpoint["scaler"], scaler)
except (KeyError, RuntimeError, ValueError):
logger.info(" > Partial model initialization...")
model_dict = model.state_dict()
model_dict = set_partial_state_dict(model_dict, checkpoint["model"], config)
model.load_state_dict(model_dict)
del model_dict
optimizer = self.restore_lr(config, self.args, model, optimizer)
logger.info(" > Model restored from step %i", checkpoint["step"])
restore_step = checkpoint["step"] + 1 # +1 not to immediately checkpoint if the model is restored
restore_epoch = checkpoint["epoch"]
torch.cuda.empty_cache()
return model, optimizer, scaler, restore_step, restore_epoch
def restore_lr(self, config, args, model, optimizer):
# use the same lr if continue training
if not args.continue_path:
if isinstance(optimizer, list):
for idx, optim in enumerate(optimizer):
for group in optim.param_groups:
group["lr"] = self.get_lr(model, config)[idx]
elif isinstance(optimizer, dict):
for optim_name, optim in optimizer.items():
for group in optim.param_groups:
group["lr"] = self.get_lr(model, config)[optim_name]
else:
for group in optimizer.param_groups:
group["lr"] = self.get_lr(model, config)
return optimizer
#########################
# DATA LOADING FUNCTIONS
#########################
def _get_loader(
self,
model: nn.Module,
config: Coqpit,
assets: Dict,
is_eval: str,
samples: List,
verbose: bool,
num_gpus: int,
) -> DataLoader:
if num_gpus > 1:
if isimplemented(model.module, "get_data_loader"):
loader = model.module.get_data_loader(
config,
assets,
is_eval,
samples,
verbose,
num_gpus,
self.args.rank,
)
else:
if isimplemented(model, "get_data_loader"):
loader = model.get_data_loader(
config=config, assets=assets, is_eval=is_eval, samples=samples, verbose=verbose, num_gpus=num_gpus
)
assert (
len(loader) > 0
), " ❗ len(DataLoader) returns 0. Make sure your dataset is not empty or len(dataset) > 0. "
return loader
def get_train_dataloader(self, training_assets: Dict, samples: List, verbose: bool) -> DataLoader:
"""Initialize and return a training data loader.
Call ```model.get_train_data_loader``` if it is implemented, else call ```model.get_data_loader```
and set ```is_eval=False```.
Args:
ap (AudioProcessor): Audio processor.
samples (List): Data samples used for training.
verbose (bool): enable/disable printing loader stats at initialization.
Returns:
DataLoader: Initialized training data loader.
"""
if self.num_gpus > 1:
if isimplemented(self.model.module, "get_train_data_loader"):
loader = self.model.module.get_train_data_loader(
self.config,
self.training_assets,
samples,
verbose,
self.num_gpus,
self.args.rank,
)
return loader
else:
if isimplemented(self.model, "get_train_data_loader"):
loader = self.model.get_train_data_loader(
self.config, self.training_assets, samples, verbose, self.num_gpus
)
return loader
return self._get_loader(
self.model,
self.config,
training_assets,
False,
samples,
verbose,
self.num_gpus,
)
def get_eval_dataloader(self, training_assets: Dict, samples: List, verbose: bool) -> DataLoader:
"""Initialize and return a evaluation data loader.
Call ```model.get_eval_data_loader``` if it is implemented, else call ```model.get_data_loader```
and set ```is_eval=True```.
Args:
ap (AudioProcessor): Audio processor.
samples (List): Data samples used for training.
verbose (bool): enable/disable printing loader stats at initialization.
Returns:
DataLoader: Initialized training data loader.
"""
if self.num_gpus > 1:
if isimplemented(self.model.module, "get_eval_data_loader"):
loader = self.model.module.get_eval_data_loader(
self.config,
self.training_assets,
samples,
verbose,
self.num_gpus,
self.args.rank,
)
return loader
else:
if isimplemented(self.model, "get_eval_data_loader"):
loader = self.model.get_eval_data_loader(
self.config, self.training_assets, samples, verbose, self.num_gpus
)
return loader
return self._get_loader(
self.model,
self.config,
training_assets,
True,
samples,
verbose,
self.num_gpus,
)
def get_test_dataloader(self, training_assets: Dict, samples: List, verbose: bool) -> DataLoader:
"""Initialize and return a evaluation data loader.
Call ```model.get_test_data_loader``` if it is implemented, else call ```model.get_data_loader```
and set ```is_eval=True```.
Args:
ap (AudioProcessor): Audio processor.
samples (List): Data samples used for training.
verbose (bool): enable/disable printing loader stats at initialization.
Returns:
DataLoader: Initialized training data loader.
"""
if self.num_gpus > 1:
if isimplemented(self.model.module, "get_test_data_loader"):
loader = self.model.module.get_test_data_loader(
self.config,
self.training_assets,
samples,
verbose,
self.num_gpus,
self.args.rank,
)
return loader
else:
if isimplemented(self.model, "get_test_data_loader"):
loader = self.model.get_test_data_loader(
self.config, self.training_assets, samples, verbose, self.num_gpus
)
return loader
return self._get_loader(
self.model,
self.config,
training_assets,
True,
samples,
verbose,
self.num_gpus,
)
def format_batch(self, batch: List) -> Dict:
"""Format the dataloader output and return a batch.
1. Call ```model.format_batch```.
2. Pass the batch to the Device.
3. Call ```model.format_batch_on_device```.
Args:
batch (List): Batch returned by the dataloader.
Returns:
Dict: Formatted batch.
"""
try:
if self.num_gpus > 1:
batch = self.model.module.format_batch(batch)
else:
batch = self.model.format_batch(batch)
except NotImplementedError:
pass
if isinstance(batch, dict):
for k, v in batch.items():
batch[k] = to_cuda(v)
elif isinstance(batch, list):
batch = [to_cuda(v) for v in batch]
try:
if self.num_gpus > 1:
batch = self.model.module.format_batch_on_device(batch)
else:
batch = self.model.format_batch_on_device(batch)
except NotImplementedError:
pass
return batch
######################
# TRAIN FUNCTIONS
######################
@staticmethod
def master_params(optimizer: torch.optim.Optimizer):
"""Generator over parameters owned by the optimizer.
Used to select parameters used by the optimizer for gradient clipping.
Args:
optimizer: Target optimizer.
"""
for group in optimizer.param_groups:
for p in group["params"]:
yield p
@staticmethod
def _model_train_step(
batch: Dict, model: nn.Module, criterion: nn.Module, optimizer_idx: int = None
) -> Tuple[Dict, Dict]:
"""
Perform a trainig forward step. Compute model outputs and losses.
Args:
batch (Dict): [description]
model (nn.Module): [description]
criterion (nn.Module): [description]
optimizer_idx (int, optional): [description]. Defaults to None.
Returns:
Tuple[Dict, Dict]: [description]
"""
input_args = [batch, criterion]
if optimizer_idx is not None:
input_args.append(optimizer_idx)
# unwrap model in DDP training
if hasattr(model, "module"):
return model.module.train_step(*input_args)
return model.train_step(*input_args)
def _get_autocast_args(self, mixed_precision: bool, precision: str):
device = "cpu"
dtype = torch.get_autocast_cpu_dtype()
if self.use_cuda:
device = "cuda"
dtype = torch.float32
if mixed_precision:
if precision == "fp16":
dtype = torch.float16
elif precision == "bf16":
dtype = torch.bfloat16
else:
raise ValueError(f" ❗ Unknown precision {precision}")
elif mixed_precision:
dtype = torch.bfloat16
return device, dtype
def detach_loss_dict(
self, loss_dict: Dict, step_optimizer: bool, optimizer_idx: int = None, grad_norm: float = None
):
# detach losses for logging
loss_dict_detached = self._detach_loss_dict(loss_dict)
# loss_dict_detached["loss"] = loss_di`ct_detached["loss"] * float(self.grad_accum_steps)
if optimizer_idx is not None:
loss_dict_detached[f"loss_{optimizer_idx}"] = loss_dict_detached.pop("loss")
if step_optimizer and grad_norm is not None:
loss_dict_detached[f"grad_norm_{optimizer_idx}"] = grad_norm
else:
if step_optimizer and grad_norm is not None:
loss_dict_detached["grad_norm"] = grad_norm
return loss_dict_detached
def _compute_loss(self, batch: Dict, model: nn.Module, criterion: nn.Module, config: Coqpit, optimizer_idx: int):
device, dtype = self._get_autocast_args(config.mixed_precision, config.precision)
with torch.autocast(device_type=device, dtype=dtype, enabled=config.mixed_precision):
if optimizer_idx is not None:
outputs, loss_dict = self._model_train_step(batch, model, criterion, optimizer_idx=optimizer_idx)
else:
outputs, loss_dict = self._model_train_step(batch, model, criterion)
return outputs, loss_dict
@staticmethod
def _set_grad_clip_per_optimizer(config: Coqpit, optimizer_idx: int):
# set gradient clipping threshold
grad_clip = 0.0 # meaning no gradient clipping
if "grad_clip" in config and config.grad_clip is not None:
if optimizer_idx is not None:
try:
grad_clip = config.grad_clip[optimizer_idx]
except TypeError:
logger.info(" [!] You are using multiple optimizers but `grad_clip` is not a list.")
else:
grad_clip = config.grad_clip
return grad_clip
def _compute_grad_norm(self, optimizer: torch.optim.Optimizer):
return torch.norm(torch.cat([param.grad.view(-1) for param in self.master_params(optimizer)], dim=0), p=2)
def _grad_clipping(self, grad_clip: float, optimizer: torch.optim.Optimizer, scaler: "AMPScaler"):
"""Perform gradient clipping"""
if grad_clip is not None and grad_clip > 0:
if scaler:
scaler.unscale_(optimizer)
self.callbacks.before_gradient_clipping(self)
grad_norm = torch.nn.utils.clip_grad_norm_(self.master_params(optimizer), grad_clip)
else:
grad_norm = self._compute_grad_norm(optimizer)
return grad_norm
def optimize(
self,
batch: Dict,
model: nn.Module,
optimizer: torch.optim.Optimizer,
scaler: "AMPScaler",
criterion: nn.Module,
scheduler: Union[torch.optim.lr_scheduler._LRScheduler, List, Dict], # pylint: disable=protected-access
config: Coqpit,
optimizer_idx: int = None,
step_optimizer: bool = True,
num_optimizers: int = 1,
) -> Tuple[Dict, Dict, int]:
"""Perform a forward - backward pass and run the optimizer.
Args:
batch (Dict): Input batch. If
model (nn.Module): Model for training. Defaults to None.
optimizer (Union[nn.optim.Optimizer, List]): Model's optimizer. If it is a list then, `optimizer_idx` must be defined to indicate the optimizer in use.
scaler (AMPScaler): AMP scaler.
criterion (nn.Module): Model's criterion.
scheduler (torch.optim.lr_scheduler._LRScheduler): LR scheduler used by the optimizer.
config (Coqpit): Model config.
optimizer_idx (int, optional): Target optimizer being used. Defaults to None.
step_optimizer (bool, optional): Whether step the optimizer. If False, gradients are accumulated and
model parameters are not updated. Defaults to True.
num_optimizers (int, optional): Number of optimizers. Defaults to 1.
Raises:
RuntimeError: When the loss is NaN.
Returns:
Tuple[Dict, Dict, int, torch.Tensor]: model outputs, losses, step time and gradient norm.
"""
step_start_time = time.time()
# forward pass and loss computation
outputs, loss_dict = self._compute_loss(
batch=batch, model=model, criterion=criterion, config=config, optimizer_idx=optimizer_idx
)
# skip the rest if not outputs from the model
if not loss_dict:
step_time = time.time() - step_start_time
return outputs, {}, step_time
grad_clip = self._set_grad_clip_per_optimizer(config=config, optimizer_idx=optimizer_idx)
# optimizer step
grad_norm = 0
update_lr_scheduler = True
# callback
self.callbacks.before_backward_pass(self, loss_dict)
# accumulated gradients adjustment
loss_dict["loss"] = loss_dict["loss"] / float(self.grad_accum_steps)
if self.use_accelerate:
with self.accelerator.accumulate(model):
ctx_mgr = self.accelerator.autocast if config.mixed_precision else nullcontext
with ctx_mgr():
self.accelerator.backward(loss_dict["loss"])
grad_norm = self._compute_grad_norm(optimizer)
if self.accelerator.sync_gradients and grad_clip is not None and grad_clip > 0:
self.accelerator.clip_grad_norm_(model.parameters(), grad_clip)
optimizer.step()
if not self.config.scheduler_after_epoch and not self.accelerator.optimizer_step_was_skipped:
scheduler.step()
optimizer.zero_grad(set_to_none=True)
else:
if self.use_amp_scaler:
if self.use_apex:
# TODO: verify AMP use for GAN training in TTS
# https://nvidia.github.io/apex/advanced.html?highlight=accumulate#backward-passes-with-multiple-optimizers
with amp.scale_loss(loss_dict["loss"], optimizer) as scaled_loss:
scaled_loss.backward()
if step_optimizer:
grad_norm = self._grad_clipping(grad_clip=grad_clip, optimizer=optimizer, scaler=None)
else:
# model optimizer step in mixed precision mode
scaler.scale(loss_dict["loss"]).backward()
# gradient accumulation
if step_optimizer:
grad_norm = self._grad_clipping(grad_clip=grad_clip, optimizer=optimizer, scaler=scaler)
scale_prev = scaler.get_scale()
scaler.step(optimizer)
# update the scaler at the end of all the optimizer steps
if optimizer_idx is None or (optimizer_idx + 1 == num_optimizers):
scaler.update()
loss_dict["amp_scaler"] = scaler.get_scale() # for logging
update_lr_scheduler = scale_prev <= scaler.get_scale()
else:
# main model optimizer step
loss_dict["loss"].backward()
# gradient accumulation
if step_optimizer:
self.callbacks.before_gradient_clipping(self)
if grad_clip > 0:
grad_norm = torch.nn.utils.clip_grad_norm_(self.master_params(optimizer), grad_clip)
optimizer.step()
# setup lr
if (
scheduler is not None
and update_lr_scheduler
and not self.config.scheduler_after_epoch
and step_optimizer
):
scheduler.step()
# zero-out optimizer
if step_optimizer:
optimizer.zero_grad(set_to_none=True)
# pytorch skips the step when the norm is 0. So ignore the norm value when it is NaN
if isinstance(grad_norm, torch.Tensor) and (torch.isnan(grad_norm) or torch.isinf(grad_norm)):
grad_norm = 0
step_time = time.time() - step_start_time
# detach loss dict
loss_dict_detached = self.detach_loss_dict(loss_dict, step_optimizer, optimizer_idx, grad_norm)
return outputs, loss_dict_detached, step_time
def train_step(self, batch: Dict, batch_n_steps: int, step: int, loader_start_time: float) -> Tuple[Dict, Dict]:
"""Perform a training step on a batch of inputs and log the process.
Args:
batch (Dict): Input batch.
batch_n_steps (int): Number of steps needed to complete an epoch. Needed for logging.
step (int): Current step number in this epoch.
loader_start_time (float): The time when the data loading is started. Needed for logging.
Returns:
Tuple[Dict, Dict]: Model outputs and losses.
"""
self.callbacks.on_train_step_start(self)
# format data
batch = self.format_batch(batch)
loader_time = time.time() - loader_start_time
# conteainers to hold model outputs and losses for each optimizer.
outputs_per_optimizer = None
loss_dict = {}
# OPTIMIZATION
if isimplemented(self.model, "optimize"): # pylint: disable=too-many-nested-blocks
# custom optimize for the model
step_time = time.time()
device, dtype = self._get_autocast_args(self.config.mixed_precision, self.config.precision)
with torch.autocast(device_type=device, dtype=dtype, enabled=self.config.mixed_precision):
outputs, loss_dict_new = self.model.optimize(
batch,
self,
)
step_time = time.time() - step_time
# If None, skip the step
if outputs is None:
return None, None
# TODO: find a way to log grad_norm for custom optimize
loss_dict_new = self.detach_loss_dict(loss_dict_new, True, None, None)
loss_dict.update(loss_dict_new)
else:
# gradient accumulation
# TODO: grad accumulation for each optimizer
step_optimizer = True
if ((step + 1) % self.grad_accum_steps != 0) and (step + 1 != batch_n_steps):
step_optimizer = False
if not isinstance(self.optimizer, list):
# auto training with a single optimizer
outputs, loss_dict_new, step_time = self.optimize(
batch,
self.model,
self.optimizer,
self.scaler,
self.criterion,
self.scheduler,
self.config,
step_optimizer=step_optimizer,
num_optimizers=1,
)
loss_dict.update(loss_dict_new)
else:
# auto training with multiple optimizers (e.g. GAN)
outputs_per_optimizer = [None] * len(self.optimizer)
total_step_time = 0
for idx, optimizer in enumerate(self.optimizer):
criterion = self.criterion
# scaler = self.scaler[idx] if self.use_amp_scaler else None
scaler = self.scaler
scheduler = None
if self.scheduler is not None:
scheduler = self.scheduler[idx]
outputs, loss_dict_new, step_time = self.optimize(
batch,
self.model,
optimizer,
scaler,
criterion,
scheduler,
self.config,
idx,
step_optimizer=step_optimizer,
num_optimizers=len(self.optimizer),
)
# skip the rest if the model returns None
total_step_time += step_time
outputs_per_optimizer[idx] = outputs
# merge loss_dicts from each optimizer
# rename duplicates with the optimizer idx
# if None, model skipped this optimizer
if loss_dict_new is not None:
for k, v in loss_dict_new.items():
if k in loss_dict:
loss_dict[f"{k}-{idx}"] = v
else:
loss_dict[k] = v
step_time = total_step_time
outputs = outputs_per_optimizer
# clear any pesky gradients after gradient accumulation
if step_optimizer:
self.model.zero_grad(set_to_none=True)
# update avg runtime stats
keep_avg_update = {}
keep_avg_update["avg_loader_time"] = loader_time
keep_avg_update["avg_step_time"] = step_time
self.keep_avg_train.update_values(keep_avg_update)
# update avg loss stats
update_eval_values = {}
for key, value in loss_dict.items():
update_eval_values["avg_" + key] = value
self.keep_avg_train.update_values(update_eval_values)
# print training progress
if self.total_steps_done % self.config.print_step == 0:
# log learning rates
lrs = {}
if isinstance(self.optimizer, list):
for idx, optimizer in enumerate(self.optimizer):
current_lr = self.optimizer[idx].param_groups[0]["lr"]
lrs.update({f"current_lr_{idx}": current_lr})
elif isinstance(self.optimizer, dict):
for key, optimizer in self.optimizer.items():
current_lr = self.optimizer[key].param_groups[0]["lr"]
lrs.update({f"current_lr_{key}": current_lr})
else:
current_lr = self.optimizer.param_groups[0]["lr"]
lrs = {"current_lr": current_lr}
# log run-time stats
loss_dict.update(lrs)
loss_dict.update(
{
"step_time": round(step_time, 4),
"loader_time": round(loader_time, 4),
}
)
self.c_logger.print_train_step(
batch_n_steps,
step,
self.total_steps_done,
loss_dict,
self.keep_avg_train.avg_values,
)
if self.args.rank == 0:
# Plot Training Iter Stats
# reduce TB load and don't log every step
if self.total_steps_done % self.config.plot_step == 0:
self.dashboard_logger.train_step_stats(self.total_steps_done, loss_dict)
if self.total_steps_done % self.config.save_step == 0 and self.total_steps_done != 0:
if self.config.save_checkpoints:
# checkpoint the model
self.save_checkpoint()
if self.total_steps_done % self.config.log_model_step == 0:
# log checkpoint as artifact
self.update_training_dashboard_logger(batch=batch, outputs=outputs)
self.dashboard_logger.flush()
self.total_steps_done += 1
self.callbacks.on_train_step_end(self)
return outputs, loss_dict
def train_epoch(self) -> None:
"""Main entry point for the training loop. Run training on the all training samples."""
# initialize the data loader
if self.train_loader is None:
self.train_loader = self.get_train_dataloader(
self.training_assets,
self.train_samples,
verbose=True,
)
self.train_loader = self.prepare_accelerate_loader(self.train_loader)
# set model to training mode
torch.set_grad_enabled(True)
if self.num_gpus > 1:
self.model.module.train()
else:
self.model.train()
epoch_start_time = time.time()
self.callbacks.on_train_epoch_start(self)
self.c_logger.print_train_start()
loader_start_time = time.time()
# TRAINING EPOCH -> iterate over the training samples
batch_num_steps = len(self.train_loader)
for cur_step, batch in enumerate(self.train_loader):
outputs, _ = self.train_step(batch, batch_num_steps, cur_step, loader_start_time)
if outputs is None:
logger.info(" [!] `train_step()` retuned `None` outputs. Skipping training step.")
continue
del outputs
loader_start_time = time.time()
# RUN EVAL -> run evaluation epoch in the middle of training. Useful for big datasets.
if self.config.run_eval_steps is not None and (self.total_steps_done % self.config.run_eval_steps == 0):
self.eval_epoch()
if self.num_gpus > 1:
self.model.module.train()
else:
self.model.train()
torch.set_grad_enabled(True)
epoch_time = time.time() - epoch_start_time
self.callbacks.on_train_epoch_end(self)
# scheduler step
if self.scheduler is not None and self.config.scheduler_after_epoch:
if isinstance(self.scheduler, list):
for scheduler in self.scheduler:
if scheduler is not None:
scheduler.step()
elif isinstance(self.scheduler, dict): # only with `model.optimize()``
for scheduler in self.scheduler.values():
if scheduler is not None:
scheduler.step()
else:
self.scheduler.step()
# plot self.epochs_done Stats
if self.args.rank == 0:
epoch_stats = {"epoch_time": epoch_time}
epoch_stats.update(self.keep_avg_train.avg_values)
self.dashboard_logger.train_epoch_stats(self.total_steps_done, epoch_stats)
if self.config.model_param_stats:
self.dashboard_logger.model_weights(self.model, self.total_steps_done)
torch.cuda.empty_cache()
#######################
# EVAL FUNCTIONS
#######################
def _model_eval_step(
self, batch: Dict, model: nn.Module, criterion: nn.Module, optimizer_idx: int = None
) -> Tuple[Dict, Dict]:
"""
Perform a evaluation forward pass. Compute model outputs and losses with no gradients.
Args:
batch (Dict): IBatch of inputs.
model (nn.Module): Model to call evaluation.
criterion (nn.Module): Model criterion.
optimizer_idx (int, optional): Optimizer ID to define the closure in multi-optimizer training. Defaults to None.
Returns:
Tuple[Dict, Dict]: model outputs and losses.
"""
input_args = [batch, criterion]
if isimplemented(model, "optimize"):
if hasattr(model, "module"):
return model.module.eval_step(batch, self)
return model.eval_step(batch, self)
if optimizer_idx is not None:
input_args.append(optimizer_idx)
if hasattr(model, "module"):
return model.module.eval_step(*input_args)
return model.eval_step(*input_args)
def eval_step(self, batch: Dict, step: int) -> Tuple[Dict, Dict]:
"""Perform a evaluation step on a batch of inputs and log the process.
Args:
batch (Dict): Input batch.
step (int): Current step number in this epoch.
Returns:
Tuple[Dict, Dict]: Model outputs and losses.
"""
with torch.no_grad():
outputs = []
loss_dict = {}
if not isinstance(self.optimizer, list) or isimplemented(self.model, "optimize"):
outputs, loss_dict = self._model_eval_step(batch, self.model, self.criterion)
if outputs is None:
return None, None
else:
outputs = [None] * len(self.optimizer)
for idx, _ in enumerate(self.optimizer):
criterion = self.criterion
outputs_, loss_dict_new = self._model_eval_step(batch, self.model, criterion, idx)
if outputs_ is None:
return None, None
outputs[idx] = outputs_
if loss_dict_new:
loss_dict_new[f"loss_{idx}"] = loss_dict_new.pop("loss")
loss_dict.update(loss_dict_new)
loss_dict = self._detach_loss_dict(loss_dict)
# update avg stats
update_eval_values = {}
for key, value in loss_dict.items():
update_eval_values["avg_" + key] = value
self.keep_avg_eval.update_values(update_eval_values)
if self.config.print_eval:
self.c_logger.print_eval_step(step, loss_dict, self.keep_avg_eval.avg_values)
return outputs, loss_dict
def eval_epoch(self) -> None:
"""Main entry point for the evaluation loop. Run evaluation on the all validation samples."""
# initialize it when eval_epoch is called alone.
self.keep_avg_eval = KeepAverage() if self.keep_avg_eval is None else self.keep_avg_eval
if self.eval_loader is None:
self.eval_loader = (
self.get_eval_dataloader(
self.training_assets,
self.eval_samples,
verbose=True,
)
if self.config.run_eval
else None
)
torch.set_grad_enabled(False)
self.model.eval()
self.c_logger.print_eval_start()
loader_start_time = time.time()
batch = None
outputs = None
for cur_step, batch in enumerate(self.eval_loader):
# format data
batch = self.format_batch(batch)
loader_time = time.time() - loader_start_time
self.keep_avg_eval.update_values({"avg_loader_time": loader_time})
outputs_, _ = self.eval_step(batch, cur_step)
if outputs_ is None:
logger.info(" [!] `eval_step()` retuned `None` outputs. Skipping evaluation step.")
continue
outputs = outputs_
loader_start_time = time.time()
# plot epoch stats, artifacts and figures
if self.args.rank == 0 and outputs is not None:
if hasattr(self.model, "module") and isimplemented(self.model.module, "eval_log"):
self.model.module.eval_log(
batch,
outputs,
self.dashboard_logger,
self.training_assets,
self.total_steps_done,
)
elif isimplemented(self.model, "eval_log"):
self.model.eval_log(
batch,
outputs,
self.dashboard_logger,
self.training_assets,
self.total_steps_done,
)
self.dashboard_logger.eval_stats(self.total_steps_done, self.keep_avg_eval.avg_values)
torch.cuda.empty_cache()
##################################
# TESTING
##################################
def test_run(self) -> None:
"""Run model test.
Test run is expected to pass over test samples and produce logging artifacts.
If ```model.test_run()``` is defined, it will be called and it is expected to set and execute everything
in the model.
Else if ```mode.test()``` is defined, it will be called and it takes an test data loader as an argument
and iterate over it.
"""
self.model.eval()
test_outputs = None
if isimplemented(self.model, "test_run") or (
self.num_gpus > 1 and isimplemented(self.model.module, "test_run")
):
# handle everything in ```model.test_run()`
if self.num_gpus > 1:
test_outputs = self.model.module.test_run(self.training_assets)
else:
test_outputs = self.model.test_run(self.training_assets)
elif isimplemented(self.model, "test") or (self.num_gpus > 1 and isimplemented(self.model.module, "test")):
self.test_loader = self.get_test_dataloader(
self.training_assets,
self.test_samples if self.test_samples else self.eval_samples,
verbose=True,
)
# use test_loader to load test samples
if self.num_gpus > 1:
test_outputs = self.model.module.test(self.training_assets, self.test_loader, None)
else:
test_outputs = self.model.test(self.training_assets, self.test_loader, None)
if isimplemented(self.model, "test_log") or (
self.num_gpus > 1 and isimplemented(self.model.module, "test_log")
):
if self.num_gpus > 1:
self.model.module.test_log(
test_outputs, self.dashboard_logger, self.training_assets, self.total_steps_done
)
else:
self.model.test_log(test_outputs, self.dashboard_logger, self.training_assets, self.total_steps_done)
def _restore_best_loss(self):
"""Restore the best loss from the args.best_path if provided else
from the model (`args.continue_path`) used for resuming the training"""
if self.args.continue_path and (self.restore_step != 0 or self.args.best_path):
logger.info(" > Restoring best loss from %s ...", os.path.basename(self.args.best_path))
ch = load_fsspec(self.args.restore_path, map_location="cpu")
if "model_loss" in ch:
if isinstance(ch["model_loss"], dict):
self.best_loss = ch["model_loss"]
# For backwards-compatibility:
elif isinstance(ch["model_loss"], float):
if self.config.run_eval:
self.best_loss = {"train_loss": None, "eval_loss": ch["model_loss"]}
else:
self.best_loss = {"train_loss": ch["model_loss"], "eval_loss": None}
logger.info(" > Starting with loaded last best loss %s", self.best_loss)
def test(self, model=None, test_samples=None) -> None:
"""Run evaluation steps on the test data split. You can either provide the model and the test samples
explicitly or the trainer use values from the initialization.
Args:
model (nn.Module, optional): Model to use for testing. If None, use the model given in the initialization.
Defaults to None.
test_samples (List[str], optional): List of test samples to use for testing. If None, use the test samples
given in the initialization. Defaults to None.
"""
logger.info(" > USING TEST SET...")
self.keep_avg_eval = KeepAverage()
if model is not None:
self.model = model
eval_samples_cache = self.eval_samples
if test_samples is not None:
self.eval_samples = test_samples
else:
self.eval_samples = self.test_samples
self.eval_epoch()
self.c_logger.print_epoch_end(self.epochs_done, self.keep_avg_eval.avg_values)
self.eval_samples = eval_samples_cache
###################################
# FIT FUNCTIONS
###################################
def _fit(self) -> None:
"""🏃 train -> evaluate -> test for the number of epochs."""
self._restore_best_loss()
self.total_steps_done = self.restore_step
for epoch in range(0, self.config.epochs):
if self.num_gpus > 1:
# let all processes sync up before starting with a new epoch of training
dist.barrier()
self.callbacks.on_epoch_start(self)
self.keep_avg_train = KeepAverage()
self.keep_avg_eval = KeepAverage() if self.config.run_eval else None
self.epochs_done = epoch
self.c_logger.print_epoch_start(epoch, self.config.epochs, self.output_path)
if not self.skip_train_epoch and not self.start_with_eval:
self.train_epoch()
if self.config.run_eval:
self.eval_epoch()
if epoch >= self.config.test_delay_epochs and self.args.rank <= 0:
self.test_run()
self.c_logger.print_epoch_end(
epoch,
self.keep_avg_eval.avg_values if self.config.run_eval else self.keep_avg_train.avg_values,
)
if self.args.rank in [None, 0]:
self.save_best_model()
self.callbacks.on_epoch_end(self)
self.start_with_eval = False
def fit_with_largest_batch_size(self, starting_batch_size=2048) -> None:
cuda_meminfo()
bs = starting_batch_size
while True:
gc.collect()
torch.cuda.empty_cache()
try:
gc.collect()
torch.cuda.empty_cache()
self.config.batch_size = bs
logger.info(" > current batch size: %i", self.config.batch_size)
self._fit()
except RuntimeError as exception:
if bs > 1 and should_reduce_batch_size(exception):
bs //= 2
gc.collect()
torch.cuda.empty_cache()
else:
raise
except Exception as exception: # pylint: disable=broad-except
# catches the torch.cuda.OutOfMemoryError
if bs > 1 and should_reduce_batch_size(exception):
bs //= 2
gc.collect()
torch.cuda.empty_cache()
else:
raise
else:
break
def fit(self) -> None:
"""Where the ✨magic✨ happens..."""
try:
self._fit()
if self.args.rank == 0:
self.dashboard_logger.finish()
except KeyboardInterrupt:
logger.info(" > Keyboard interrupt detected.")
if self.config.save_on_interrupt:
logger.info(" > Saving model before exiting...")
# save the model on keyboard interrupt
self.save_checkpoint()
# update the training dashboard logger
self.update_training_dashboard_logger()
# call the keyboard interrupt callback
self.callbacks.on_keyboard_interrupt(self)
# if the output folder is empty remove the run.
remove_experiment_folder(self.output_path)
# clear the DDP processes
if self.num_gpus > 1:
dist.destroy_process_group()
# finish the wandb run and sync data
if self.args.rank == 0:
self.dashboard_logger.finish()
# stop without error signal
try:
sys.exit(1)
except SystemExit:
os._exit(1) # pylint: disable=protected-access
except BaseException: # pylint: disable=broad-except
remove_experiment_folder(self.output_path)
traceback.print_exc()
sys.exit(1)
def profile_fit(self, torch_profiler, epochs=None, small_run=None):
"""Run training under the torch profiler.
Example::
Run torch profiler to profile CPU, GPU and memory usage with Tensorboard logging.
>>> import torch
>>> profiler = torch.profiler.profile(
>>> activities=[
>>> torch.profiler.ProfilerActivity.CPU,
>>> torch.profiler.ProfilerActivity.CUDA,
>>> ],
>>> schedule=torch.profiler.schedule(wait=1, warmup=1, active=3, repeat=2),
>>> on_trace_ready=torch.profiler.tensorboard_trace_handler("./profiler/"),
>>> record_shapes=True,
>>> profile_memory=True,
>>> with_stack=True,
>>> )
>>> prof = trainer.profile_fit(profiler, epochs=1, small_run=64)
"""
self.dashboard_logger = DummyLogger()
# train the model for a custom number of epochs
if epochs:
self.config.epocshs = epochs
# use a smaller set of training samples for profiling
if small_run:
self.setup_small_run(small_run)
# run profiler
self.config.run_eval = False
self.config.test_delay_epochs = 9999999
self.config.epochs = epochs
# set a callback to progress the profiler
self.callbacks_on_train_step_end = [ # pylint: disable=attribute-defined-outside-init
lambda trainer: trainer.torch_profiler.step()
]
# set the profiler to access in the Trainer
self.torch_profiler = torch_profiler # pylint: disable=attribute-defined-outside-init
# set logger output for Tensorboard
# self.torch_profiler.on_trace_ready = torch.profiler.tensorboard_trace_handler(self.output_path)
self.torch_profiler.start()
self.fit()
self.torch_profiler.stop()
return self.torch_profiler
@rank_zero_only
def save_best_model(self) -> None:
"""Save the best model. It only saves if the current target loss is smaller then the previous."""
eval_loss = self._pick_target_avg_loss(self.keep_avg_eval)
train_loss = self._pick_target_avg_loss(self.keep_avg_train)
# save the model and update the best_loss
self.best_loss = save_best_model(
{"train_loss": train_loss, "eval_loss": eval_loss},
self.best_loss,
self.config,
self.model,
self.optimizer,
self.scaler if self.use_amp_scaler else None,
self.total_steps_done,
self.epochs_done,
self.output_path,
keep_all_best=self.config.save_all_best,
keep_after=self.config.save_best_after,
save_func=self.dashboard_logger.save_model,
)
@rank_zero_only
def save_checkpoint(self) -> None:
"""Save the current model checkpoint."""
eval_loss = self._pick_target_avg_loss(self.keep_avg_eval)
train_loss = self._pick_target_avg_loss(self.keep_avg_train)
save_checkpoint(
self.config,
self.model,
self.optimizer,
self.scaler if self.use_amp_scaler else None,
self.total_steps_done,
self.epochs_done,
self.output_path,
model_loss={"train_loss": train_loss, "eval_loss": eval_loss},
save_n_checkpoints=self.config.save_n_checkpoints,
save_func=self.dashboard_logger.save_model,
)
@rank_zero_only
def update_training_dashboard_logger(self, batch=None, outputs=None):
aliases = [
f"epoch-{self.epochs_done}",
f"step-{self.total_steps_done}",
]
self.dashboard_logger.add_artifact(
file_or_dir=self.output_path, name="checkpoint", artifact_type="model", aliases=aliases
)
# training visualizations
if batch is not None and outputs is not None:
if hasattr(self.model, "module") and isimplemented(self.model.module, "train_log"):
self.model.module.train_log(
batch,
outputs,
self.dashboard_logger,
self.training_assets,
self.total_steps_done,
)
elif isimplemented(self.model, "train_log"):
self.model.train_log(
batch,
outputs,
self.dashboard_logger,
self.training_assets,
self.total_steps_done,
)
#####################
# GET FUNCTIONS
#####################
@staticmethod
def get_optimizer(model: nn.Module, config: Coqpit) -> Union[torch.optim.Optimizer, List]:
"""Receive the optimizer from the model if model implements `get_optimizer()` else
check the optimizer parameters in the config and try initiating the optimizer.
Args:
model (nn.Module): Training model.
config (Coqpit): Training configuration.
Returns:
Union[torch.optim.Optimizer, List]: A optimizer or a list of optimizers. GAN models define a list.
"""
optimizer = None
if isimplemented(model, "get_optimizer"):
try:
optimizer = model.get_optimizer()
except NotImplementedError:
optimizer = None
if optimizer is None:
optimizer_name = config.optimizer
optimizer_params = {} if config.optimizer_params is None else config.optimizer_params
return get_optimizer(optimizer_name, optimizer_params, config.lr, model)
return optimizer
@staticmethod
def get_lr(model: nn.Module, config: Coqpit) -> Union[float, List[float]]:
"""Set the initial learning rate by the model if model implements `get_lr()` else try setting the learning rate
fromthe config.
Args:
model (nn.Module): Training model.
config (Coqpit): Training configuration.
Returns:
Union[float, List[float]]: A single learning rate or a list of learning rates, one for each optimzier.
"""
lr = None
if isimplemented(model, "get_lr"):
try:
lr = model.get_lr()
except NotImplementedError:
lr = None
if lr is None:
lr = config.lr
return lr
@staticmethod
def get_scheduler(
model: nn.Module, config: Coqpit, optimizer: Union[torch.optim.Optimizer, List, Dict]
) -> Union[torch.optim.lr_scheduler._LRScheduler, List]: # pylint: disable=protected-access
"""Receive the scheduler from the model if model implements `get_scheduler()` else
check the config and try initiating the scheduler.
Args:
model (nn.Module): Training model.
config (Coqpit): Training configuration.
Returns:
Union[torch.optim.Optimizer, List, Dict]: A scheduler or a list of schedulers, one for each optimizer.
"""
scheduler = None
if isimplemented(model, "get_scheduler"):
try:
scheduler = model.get_scheduler(optimizer)
except NotImplementedError:
scheduler = None
if isinstance(scheduler, dict) and not isimplemented(model, "optimize"):
raise ValueError(
" [!] Dictionary of schedulers are only supported with the manual optimization `model.optimize()`."
)
if scheduler is None:
lr_scheduler = config.lr_scheduler
lr_scheduler_params = config.lr_scheduler_params
return get_scheduler(lr_scheduler, lr_scheduler_params, optimizer)
return scheduler
@staticmethod
def restore_scheduler(
scheduler: Union["Scheduler", List, Dict], args: Coqpit, config: Coqpit, restore_epoch: int, restore_step: int
) -> Union["Scheduler", List]:
"""Restore scheduler wrt restored model."""
if scheduler is not None: # pylint: disable=too-many-nested-blocks
if args.continue_path:
if isinstance(scheduler, list):
for s in scheduler:
if s is not None:
if config.scheduler_after_epoch:
s.last_epoch = restore_epoch
else:
s.last_epoch = restore_step
elif isinstance(scheduler, dict):
for s in scheduler.values():
if s is not None:
if config.scheduler_after_epoch:
s.last_epoch = restore_epoch
else:
s.last_epoch = restore_step
else:
if config.scheduler_after_epoch:
scheduler.last_epoch = restore_epoch
else:
scheduler.last_epoch = restore_step
return scheduler
@staticmethod
def get_criterion(model: nn.Module) -> nn.Module:
"""Receive the criterion from the model. Model must implement `get_criterion()`.
Args:
model (nn.Module): Training model.
Returns:
nn.Module: Criterion layer.
"""
criterion = None
criterion = model.get_criterion()
return criterion
####################
# HELPER FUNCTIONS
####################
@staticmethod
def _detach_loss_dict(loss_dict: Dict) -> Dict:
"""Detach loss values from autograp.
Args:
loss_dict (Dict): losses.
Returns:
Dict: losses detached from autograph.
"""
loss_dict_detached = {}
for key, value in loss_dict.items():
if isinstance(value, (int, float)):
loss_dict_detached[key] = value
else:
loss_dict_detached[key] = value.detach().cpu().item()
return loss_dict_detached
def _pick_target_avg_loss(self, keep_avg_target: KeepAverage) -> Dict:
"""Pick the target loss to compare models"""
# if the keep_avg_target is None or empty return None
if keep_avg_target is None or len(list(keep_avg_target.avg_values.keys())) == 0:
return None
target_avg_loss = None
# return if target loss defined in the model config
# if not available in Dict use loss_1 as by default loss
if "target_loss" in self.config and self.config.target_loss:
if f"avg_{self.config.target_loss}" in keep_avg_target.avg_values.keys():
return keep_avg_target[f"avg_{self.config.target_loss}"]
raise ValueError(
" [!] Target loss not found in the keep_avg_target. You might be exiting the training loop before it is computed or set the target_loss in the model config incorrectly."
)
# take the average of loss_{optimizer_idx} as the target loss when there are multiple optimizers
if isinstance(self.optimizer, list):
target_avg_loss = 0
for idx in range(len(self.optimizer)):
if f"avg_loss_{idx}" in keep_avg_target.avg_values:
target_avg_loss += keep_avg_target[f"avg_loss_{idx}"]
target_avg_loss /= len(self.optimizer)
else:
target_avg_loss = keep_avg_target.avg_values.get("avg_loss", 0)
return target_avg_loss
def _setup_logger_config(self, log_file: str) -> None:
"""Set up the logger based on the process rank in DDP."""
logger_new = logging.getLogger("trainer")
handler = logging.FileHandler(log_file, mode="a")
fmt = logging.Formatter("")
handler.setFormatter(fmt)
logger_new.addHandler(handler)
# only log to a file if rank > 0 in DDP
if self.args.rank > 0:
logger_new.handlers = [h for h in logger_new.handlers if not isinstance(h, logging.StreamHandler)]
@staticmethod
def _is_apex_available() -> bool:
"""Check if Nvidia's APEX is available."""
return importlib.util.find_spec("apex") is not None