# Copyright 2022 Tomoki Hayashi
# 2022 Dongji Gao
# Apache 2.0 (http://www.apache.org/licenses/LICENSE-2.0)
"""Trainer module for GAN-based UASR training."""
import argparse
import dataclasses
import logging
import math
import time
from contextlib import contextmanager
from typing import Dict, Iterable, List, Optional, Sequence, Tuple
import torch
from packaging.version import parse as V
from typeguard import typechecked
from espnet2.schedulers.abs_scheduler import AbsBatchStepScheduler, AbsScheduler
from espnet2.torch_utils.device_funcs import to_device
from espnet2.torch_utils.recursive_op import recursive_average
from espnet2.train.distributed_utils import DistributedOption
from espnet2.train.reporter import SubReporter
from espnet2.train.trainer import Trainer, TrainerOptions
from espnet2.utils.build_dataclass import build_dataclass
from espnet2.utils.types import str2bool
if torch.distributed.is_available():
from torch.distributed import ReduceOp
if V(torch.__version__) >= V("1.6.0"):
from torch.cuda.amp import GradScaler, autocast
else:
# Nothing to do if torch<1.6.0
@contextmanager
def autocast(enabled=True): # NOQA
yield
GradScaler = None
try:
import fairscale
except ImportError:
fairscale = None
[docs]@dataclasses.dataclass
class UASRTrainerOptions(TrainerOptions):
"""Trainer option dataclass for UASRTrainer."""
generator_first: bool
max_num_warning: int
[docs]class UASRTrainer(Trainer):
"""Trainer for GAN-based UASR training.
If you'd like to use this trainer, the model must inherit
espnet.train.abs_gan_espnet_model.AbsGANESPnetModel.
"""
[docs] @classmethod
@typechecked
def build_options(cls, args: argparse.Namespace) -> TrainerOptions:
"""Build options consumed by train(), eval(), and plot_attention()."""
return build_dataclass(UASRTrainerOptions, args)
[docs] @classmethod
def add_arguments(cls, parser: argparse.ArgumentParser):
"""Add additional arguments for GAN-trainer."""
parser.add_argument(
"--generator_first",
type=str2bool,
default=False,
help="Whether to update generator first.",
)
parser.add_argument(
"--max_num_warning",
type=int,
default=10,
help="Maximum number of warning shown",
)
[docs] @classmethod
@typechecked
def train_one_epoch(
cls,
model: torch.nn.Module,
iterator: Iterable[Tuple[List[str], Dict[str, torch.Tensor]]],
optimizers: Sequence[torch.optim.Optimizer],
schedulers: Sequence[Optional[AbsScheduler]],
scaler: Optional[GradScaler],
reporter: SubReporter,
summary_writer,
options: UASRTrainerOptions,
distributed_option: DistributedOption,
) -> bool:
"""Train one epoch for UASR."""
grad_noise = options.grad_noise
accum_grad = options.accum_grad
grad_clip = options.grad_clip
grad_clip_type = options.grad_clip_type
log_interval = options.log_interval
no_forward_run = options.no_forward_run
ngpu = options.ngpu
use_wandb = options.use_wandb
distributed = distributed_option.distributed
max_num_warning = options.max_num_warning
cur_num_warning = 0
hide_warning = False
# Check unavailable options
# TODO(kan-bayashi): Support the use of these options
if accum_grad > 1:
raise NotImplementedError(
"accum_grad > 1 is not supported in GAN-based training."
)
if grad_noise:
raise NotImplementedError(
"grad_noise is not supported in GAN-based training."
)
if log_interval is None:
try:
log_interval = max(len(iterator) // 20, 10)
except TypeError:
log_interval = 100
model.train()
all_steps_are_invalid = True
model.number_epochs = reporter.epoch
# [For distributed] Because iteration counts are not always equals between
# processes, send stop-flag to the other processes if iterator is finished
iterator_stop = torch.tensor(0).to("cuda" if ngpu > 0 else "cpu")
start_time = time.perf_counter()
for iiter, (_, batch) in enumerate(
reporter.measure_iter_time(iterator, "iter_time"), 1
):
assert isinstance(batch, dict), type(batch)
if distributed:
torch.distributed.all_reduce(iterator_stop, ReduceOp.SUM)
if iterator_stop > 0:
break
batch = to_device(batch, "cuda" if ngpu > 0 else "cpu")
if no_forward_run:
all_steps_are_invalid = False
continue
turn_start_time = time.perf_counter()
model.number_updates = iiter - 1
is_discriminative_step = model.is_discriminative_step()
optim_idx = model.get_optim_index()
if is_discriminative_step:
turns = ["discriminator"]
else:
turns = ["generator"]
for turn in turns:
with autocast(scaler is not None):
with reporter.measure_time(f"{turn}_forward_time"):
retval = model(**batch)
# Note(jiatong):
# Supporting only one patterns
# for the returned value from the model
# must be tuple or list type
if not (isinstance(retval, list) or isinstance(retval, tuple)):
raise RuntimeError("model output must be tuple or list.")
else:
loss, stats, weight, _ = retval
stats = {k: v for k, v in stats.items() if v is not None}
if ngpu > 1 or distributed:
# Apply weighted averaging for loss and stats
loss = (loss * weight.type(loss.dtype)).sum()
# if distributed, this method can also apply all_reduce()
stats, weight = recursive_average(stats, weight, distributed)
# Now weight is summation over all workers
loss /= weight
if distributed:
# NOTE(kamo): Multiply world_size since DistributedDataParallel
# automatically normalizes the gradient by world_size.
loss *= torch.distributed.get_world_size()
reporter.register(stats, weight)
with reporter.measure_time(f"{turn}_backward_time"):
if scaler is not None:
# Scales loss. Calls backward() on scaled loss
# to create scaled gradients.
# Backward passes under autocast are not recommended.
# Backward ops run in the same dtype autocast chose
# for corresponding forward ops.
scaler.scale(loss).backward()
else:
loss.backward()
if scaler is not None:
# Unscales the gradients of optimizer's assigned params in-place
for iopt, optimizer in enumerate(optimizers):
if optim_idx is not None and iopt != optim_idx:
continue
scaler.unscale_(optimizer)
# TODO(kan-bayashi): Compute grad norm without clipping
grad_norm = None
if grad_clip > 0.0:
# compute the gradient norm to check if it is normal or not
grad_norm = torch.nn.utils.clip_grad_norm_(
model.parameters(),
max_norm=grad_clip,
norm_type=grad_clip_type,
)
# PyTorch<=1.4, clip_grad_norm_ returns float value
if not isinstance(grad_norm, torch.Tensor):
grad_norm = torch.tensor(grad_norm)
if grad_norm is None or torch.isfinite(grad_norm):
all_steps_are_invalid = False
with reporter.measure_time(f"{turn}_optim_step_time"):
for iopt, (optimizer, scheduler) in enumerate(
zip(optimizers, schedulers)
):
if optim_idx is not None and iopt != optim_idx:
continue
if scaler is not None:
# scaler.step() first unscales the gradients of
# the optimizer's assigned params.
scaler.step(optimizer)
# Updates the scale for next iteration.
scaler.update()
else:
optimizer.step()
if isinstance(scheduler, AbsBatchStepScheduler):
scheduler.step()
else:
cur_num_warning += 1
if cur_num_warning >= max_num_warning:
if not hide_warning:
logging.info("Warning info folded...")
hide_warning = True
if not hide_warning:
logging.warning(
f"The grad norm is {grad_norm}. "
"Skipping updating the model."
)
# Must invoke scaler.update() if unscale_() is used in the
# iteration to avoid the following error:
# RuntimeError: unscale_() has already been called
# on this optimizer since the last update().
# Note that if the gradient has inf/nan values,
# scaler.step skips optimizer.step().
if scaler is not None:
for iopt, optimizer in enumerate(optimizers):
if optim_idx is not None and iopt != optim_idx:
continue
scaler.step(optimizer)
scaler.update()
for iopt, optimizer in enumerate(optimizers):
# NOTE(kan-bayashi): In the case of GAN, we need to clear
# the gradient of both optimizers after every update.
optimizer.zero_grad()
# Register lr and train/load time[sec/step],
# where step refers to accum_grad * mini-batch
reporter.register(
{
f"optim{optim_idx}_lr{i}": pg["lr"]
for i, pg in enumerate(optimizers[optim_idx].param_groups)
if "lr" in pg
},
)
reporter.register(
{f"{turn}_train_time": time.perf_counter() - turn_start_time}
)
turn_start_time = time.perf_counter()
reporter.register({"train_time": time.perf_counter() - start_time})
start_time = time.perf_counter()
# NOTE(kamo): Call log_message() after next()
reporter.next()
if iiter % log_interval == 0:
logging.info(reporter.log_message(-log_interval))
if summary_writer is not None:
reporter.tensorboard_add_scalar(summary_writer, -log_interval)
if use_wandb:
reporter.wandb_log()
else:
if distributed:
iterator_stop.fill_(1)
torch.distributed.all_reduce(iterator_stop, ReduceOp.SUM)
if hide_warning:
logging.warning(
f"{cur_num_warning}/{iiter} iterations skipped due to inf/nan grad norm"
)
return all_steps_are_invalid
[docs] @classmethod
@torch.no_grad()
@typechecked
def validate_one_epoch(
cls,
model: torch.nn.Module,
iterator: Iterable[Dict[str, torch.Tensor]],
reporter: SubReporter,
options: UASRTrainerOptions,
distributed_option: DistributedOption,
) -> None:
"""Validate one epoch."""
ngpu = options.ngpu
no_forward_run = options.no_forward_run
distributed = distributed_option.distributed
vocab_seen_list = []
model.eval()
logging.info("Doing validation")
# [For distributed] Because iteration counts are not always equals between
# processes, send stop-flag to the other processes if iterator is finished
iterator_stop = torch.tensor(0).to("cuda" if ngpu > 0 else "cpu")
print_hyp = True
for _, batch in iterator:
assert isinstance(batch, dict), type(batch)
if distributed:
torch.distributed.all_reduce(iterator_stop, ReduceOp.SUM)
if iterator_stop > 0:
break
batch = to_device(batch, "cuda" if ngpu > 0 else "cpu")
if no_forward_run:
continue
retval = model(**batch, do_validation=True, print_hyp=print_hyp)
print_hyp = False
if not (isinstance(retval, list) or isinstance(retval, tuple)):
raise RuntimeError("model output must be tuple or list.")
else:
loss, stats, weight, vocab_seen = retval
vocab_seen_list.append(vocab_seen)
stats = {k: v for k, v in stats.items() if v is not None}
if ngpu > 1 or distributed:
# Apply weighted averaging for stats.
# if distributed, this method can also apply all_reduce()
stats, weight = recursive_average(stats, weight, distributed)
reporter.register(stats)
reporter.next()
else:
if distributed:
iterator_stop.fill_(1)
torch.distributed.all_reduce(iterator_stop, ReduceOp.SUM)
# compute phone error rate
total_num_errors = 0
total_num_ref_tokens = 0
assert (
"batch_num_errors" in reporter.stats
and "batch_num_ref_tokens" in reporter.stats
)
for batch_num_errors, batch_num_ref_tokens in zip(
reporter.stats["batch_num_errors"], reporter.stats["batch_num_ref_tokens"]
):
total_num_errors += batch_num_errors.value
total_num_ref_tokens += batch_num_ref_tokens.value
phone_error_rate = total_num_errors / total_num_ref_tokens
reporter.register({"phone_error_rate": phone_error_rate})
# compute lm_ppl
if model.kenlm:
assert (
"batch_lm_log_prob" in reporter.stats
and "batch_num_hyp_tokens" in reporter.stats
and "batch_size" in reporter.stats
)
assert (
len(reporter.stats["batch_lm_log_prob"])
== len(reporter.stats["batch_num_hyp_tokens"])
== len(reporter.stats["batch_size"])
)
total_lm_log_prob = 0
total_num_tokens = 0
total_num_sentences = 0
for log_prob, num_tokens, batch_size in zip(
reporter.stats["batch_lm_log_prob"],
reporter.stats["batch_num_hyp_tokens"],
reporter.stats["batch_size"],
):
total_lm_log_prob += log_prob.value
total_num_tokens += num_tokens.value
total_num_sentences += batch_size.value
lm_ppl = math.pow(
10, -total_lm_log_prob / (total_num_tokens + total_num_sentences)
)
vocab_seen = torch.stack(vocab_seen_list).sum(dim=0).bool().sum()
vocab_seen_rate = vocab_seen / model.vocab_size
assert vocab_seen_rate <= 1.0
weighted_lm_ppl = lm_ppl / vocab_seen_rate**2
reporter.register({"lm_ppl": lm_ppl})
reporter.register({"weighted_lm_ppl": weighted_lm_ppl})
reporter.next()