# Copyright 2022 Cerebras Systems.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Contains the Logging class that handles setting up and logging to the standard Python logger."""
from __future__ import annotations
from typing import TYPE_CHECKING, List, Optional
import torch
import cerebras.pytorch as cstorch
from cerebras.modelzoo.common.half_dtype import cb16_to_fp32
from cerebras.modelzoo.trainer.callbacks import RateProfiler
from cerebras.modelzoo.trainer.loggers import Logger
from cerebras.pytorch.utils.data.utils import infer_batch_size
if TYPE_CHECKING:
from ..trainer import Trainer
[docs]class ProgressLogger(Logger):
"""
Callback that handles setting up and logging to the standard Python logger.
"""
def __init__(self):
"""Sets up the rate tracker and total samples tracker."""
self.rate_profiler = None
# Keep track of the total samples processed
# across all stages
self.total_samples = cstorch.utils.tracker.RateTracker()
self.accum_loss = None
def setup(self, trainer):
self.rate_profiler = trainer.get_callback(RateProfiler)
@property
def postfix(self) -> List[str]:
"""Returns the postfix to append to the progress message."""
if self.rate_profiler is not None:
rate = self.rate_profiler.rate
global_rate = self.rate_profiler.global_rate
return [
f"Rate={self.format_rate(rate)}",
f"GlobalRate={self.format_rate(global_rate)}",
]
return []
def on_fit_start(self, trainer, train_dataloader, val_dataloader, loop):
self.total_samples.reset()
def on_train_batch_end(self, trainer, model, outputs, batch, batch_idx):
batch_size = infer_batch_size(batch)
if batch_size:
self.total_samples.add(batch_size)
loss = cb16_to_fp32(outputs["loss"])
self.print_training_progress(trainer, loss, batch_size)
[docs] @cstorch.step_closure
def print_training_progress(
self, trainer: Trainer, loss: torch.Tensor, batch_size: Optional[int]
):
"""Print training progress and log metrics."""
if trainer.should_run_optimizer_step:
if self.accum_loss is not None:
loss = self.accum_loss + loss.item()
self.accum_loss = None
else:
loss = loss.item()
if trainer.is_log_step:
progress_msg = [
f"| Train Device={trainer.backend.device}",
f"Step={trainer.global_step}",
f"Loss={loss:.5f}",
*self.postfix,
]
trainer.logger.info(", ".join(progress_msg))
self._log_loss_rate_metrics(trainer, loss, batch_size)
else:
# accumulate loss for gradient accumulation
if self.accum_loss is None:
self.accum_loss = loss.item()
else:
self.accum_loss += loss.item()
def on_train_end(self, trainer, model, loop, loop_idx):
trainer.logger.info("Training completed successfully!")
def on_validate_start(self, trainer, model, val_dataloader, loop):
# pylint: disable=attribute-defined-outside-init
self.total_eval_loss = 0
self.total_eval_steps = 0
def on_validate_batch_end(self, trainer, model, outputs, batch, batch_idx):
loss = cb16_to_fp32(outputs["loss"])
self.print_validation_progress(trainer, loss, infer_batch_size(batch))
[docs] @cstorch.step_closure
def print_validation_progress(
self,
trainer: Trainer,
loss: torch.Tensor,
batch_size: Optional[int],
):
"""Print validation progress and log metrics."""
self.total_eval_loss += loss.item()
self.total_eval_steps += 1
if trainer.is_log_step:
progress_msg = [
f"| Eval Device={trainer.backend.device}",
f"GlobalStep={trainer.global_step}",
f"Batch={trainer.executor.user_iteration}",
f"Loss={loss.item():.5f}",
*self.postfix,
]
trainer.logger.info(", ".join(progress_msg))
if trainer.is_final_iteration:
avg_eval_loss = self.total_eval_loss / self.total_eval_steps
trainer.logger.info(f"Avg Eval Loss: {avg_eval_loss}")
self._log_loss_rate_metrics(trainer, avg_eval_loss, batch_size)
def on_validate_end(self, trainer, model, loop):
trainer.logger.info("Evaluation completed successfully!")
def on_fit_end(self, trainer, loop):
# pylint: disable=protected-access
if trainer.backend._impl.is_e2e_execution:
trainer.logger.info(
f"Processed {int(self.total_samples.total_count)} training sample(s) "
f"in {self.total_samples.elapsed_seconds()} seconds."
)
def on_fit_exception(self, trainer, exception):
self.on_fit_end(trainer, None)
def log_metrics(self, metrics, step):
pass
def _log_loss_rate_metrics(
self, trainer: Trainer, loss: float, batch_size: Optional[int]
):
trainer.log_metrics(loss=loss)
if self.rate_profiler:
metrics = {
"local_samples_per_sec": self.rate_profiler.rate,
"avg_samples_per_sec": self.rate_profiler.global_rate,
}
if batch_size:
metrics["avg_steps_per_sec"] = (
self.rate_profiler.global_rate / batch_size
)
trainer.log_metrics(**metrics)