2021-10-28 16:21:23 +00:00
|
|
|
#!/usr/bin/env python
|
|
|
|
# -*- encoding: utf-8 -*-
|
|
|
|
|
2021-12-27 07:04:32 +00:00
|
|
|
from abc import ABC, abstractmethod
|
|
|
|
from typing import Callable
|
|
|
|
|
|
|
|
import torch
|
|
|
|
import torch.distributed as dist
|
2023-08-31 05:51:28 +00:00
|
|
|
|
2021-12-27 07:04:32 +00:00
|
|
|
from colossalai.communication import all_reduce
|
2021-10-28 16:21:23 +00:00
|
|
|
from colossalai.context import ParallelMode
|
2021-12-27 07:04:32 +00:00
|
|
|
from colossalai.core import global_context as gpc
|
2023-09-04 11:56:42 +00:00
|
|
|
from colossalai.legacy.registry import HOOKS
|
2021-12-27 07:04:32 +00:00
|
|
|
from colossalai.utils import get_current_device, is_no_pp_or_last_stage
|
|
|
|
|
2021-10-28 16:21:23 +00:00
|
|
|
from ._base_hook import BaseHook
|
2022-04-19 02:13:08 +00:00
|
|
|
from ._commons_ import _format_number
|
2021-12-27 07:04:32 +00:00
|
|
|
|
|
|
|
|
|
|
|
class Metric(ABC):
|
|
|
|
"""A basic class of metric collectors. It collects a specific
|
2022-03-25 05:02:39 +00:00
|
|
|
metric during training or evaluation and would always be used with
|
2023-08-31 05:51:28 +00:00
|
|
|
:class:`MetricHook` to help it update its states and show the
|
|
|
|
metric. So please use corresponding hook class to make the metric
|
2021-12-27 07:04:32 +00:00
|
|
|
collector works.
|
|
|
|
|
2022-03-25 05:02:39 +00:00
|
|
|
Args:
|
|
|
|
epoch_only (bool): Whether the metric only read for the full epoch.
|
2021-12-27 07:04:32 +00:00
|
|
|
"""
|
2022-01-04 12:52:31 +00:00
|
|
|
|
2021-12-27 07:04:32 +00:00
|
|
|
def __init__(self, epoch_only: bool):
|
|
|
|
# is the metric only read for the full epoch
|
|
|
|
self._epoch_only = epoch_only
|
|
|
|
|
|
|
|
@property
|
|
|
|
def epoch_only(self):
|
|
|
|
"""Returns :attr:`epoch_only`.
|
|
|
|
"""
|
|
|
|
return self._epoch_only
|
|
|
|
|
|
|
|
@abstractmethod
|
|
|
|
def reset(self) -> None:
|
|
|
|
"""Resets the metric to it's initial state.
|
|
|
|
By default, this is called at the start of each epoch.
|
|
|
|
"""
|
|
|
|
pass
|
|
|
|
|
|
|
|
@abstractmethod
|
|
|
|
def update(self, *args, **kwargs) -> None:
|
|
|
|
"""Updates the metric's state using the passed batch output.
|
|
|
|
By default, this is called once for each batch.
|
|
|
|
"""
|
|
|
|
pass
|
|
|
|
|
|
|
|
@abstractmethod
|
2022-06-07 09:21:11 +00:00
|
|
|
def get_last_step_value(self) -> float:
|
2021-12-27 07:04:32 +00:00
|
|
|
"""Returns the metric value in the last iteration.
|
|
|
|
"""
|
|
|
|
pass
|
|
|
|
|
|
|
|
@abstractmethod
|
|
|
|
def get_accumulated_value(self):
|
|
|
|
"""Computes the metric based on it's accumulated state.
|
|
|
|
By default, this is called at the end of each epoch.
|
|
|
|
|
|
|
|
:return: the actual quantity of interest
|
|
|
|
:rtype: Any
|
|
|
|
"""
|
|
|
|
pass
|
|
|
|
|
|
|
|
@staticmethod
|
|
|
|
@abstractmethod
|
|
|
|
def is_better(a, b) -> bool:
|
|
|
|
"""Compares a and b, and returns whether a is better than b
|
|
|
|
|
|
|
|
:return: The result of comparison
|
|
|
|
:rtype: bool
|
|
|
|
"""
|
|
|
|
pass
|
|
|
|
|
|
|
|
|
|
|
|
class LossMetric(Metric):
|
|
|
|
"""A metric collector for loss.
|
|
|
|
|
2022-03-25 05:02:39 +00:00
|
|
|
Args:
|
|
|
|
epoch_only (bool): Whether the metric only read for the full epoch.
|
2021-12-27 07:04:32 +00:00
|
|
|
"""
|
2022-01-04 12:52:31 +00:00
|
|
|
|
2021-12-27 07:04:32 +00:00
|
|
|
def __init__(self, epoch_only):
|
|
|
|
super().__init__(epoch_only=epoch_only)
|
|
|
|
self.last_step_loss = torch.zeros(1, device=get_current_device())
|
|
|
|
self.accum_loss = torch.zeros(1, device=get_current_device())
|
|
|
|
self.count = 0
|
|
|
|
|
|
|
|
def reset(self) -> None:
|
|
|
|
"""Sets :attr:`last_step_loss` and :attr:`accum_loss` to zero.
|
|
|
|
"""
|
|
|
|
self.last_step_loss.zero_()
|
|
|
|
self.accum_loss.zero_()
|
|
|
|
self.count = 0
|
|
|
|
|
|
|
|
def update(self, loss) -> None:
|
|
|
|
"""Updates :attr:`last_step_loss` and :attr:`accum_loss` with current loss.
|
|
|
|
It expects the output has loss.
|
|
|
|
|
2022-03-25 05:02:39 +00:00
|
|
|
Args:
|
|
|
|
loss (:class:`torch.tensor`): Current loss of the output.
|
2021-12-27 07:04:32 +00:00
|
|
|
"""
|
|
|
|
# expect output to be logits, label and loss
|
|
|
|
loss_ = loss.detach()
|
|
|
|
self.last_step_loss.copy_(loss_)
|
|
|
|
self.accum_loss.add_(loss_)
|
|
|
|
self.count += 1
|
|
|
|
|
|
|
|
def get_accumulated_value(self):
|
|
|
|
"""Returns accumulated loss.
|
|
|
|
"""
|
|
|
|
if gpc.is_initialized(ParallelMode.DATA):
|
|
|
|
dist.all_reduce(self.accum_loss, op=dist.ReduceOp.SUM, group=gpc.get_group(ParallelMode.DATA))
|
|
|
|
self.accum_loss.div_(gpc.get_world_size(ParallelMode.DATA))
|
|
|
|
|
|
|
|
self.accum_loss.div_(self.count)
|
|
|
|
return self.accum_loss.item()
|
|
|
|
|
2022-06-07 09:21:11 +00:00
|
|
|
def get_last_step_value(self) -> float:
|
2021-12-27 07:04:32 +00:00
|
|
|
"""Returns :attr:`last_step_loss`.
|
|
|
|
"""
|
2022-06-07 09:21:11 +00:00
|
|
|
return self.last_step_loss.cpu().item()
|
2021-12-27 07:04:32 +00:00
|
|
|
|
2022-01-21 02:44:30 +00:00
|
|
|
@staticmethod
|
2021-12-27 07:04:32 +00:00
|
|
|
def is_better(a, b):
|
|
|
|
return a < b
|
|
|
|
|
|
|
|
|
|
|
|
class LearningRateMetric(Metric):
|
|
|
|
"""A metric collector for learning rate.
|
|
|
|
|
2022-03-25 05:02:39 +00:00
|
|
|
Args:
|
|
|
|
epoch_only (bool): Whether the metric only read for the full epoch.
|
|
|
|
initial_lr (float, optional): Initial learning rate, defaults to 0.0.
|
2021-12-27 07:04:32 +00:00
|
|
|
"""
|
2022-01-04 12:52:31 +00:00
|
|
|
|
2021-12-27 07:04:32 +00:00
|
|
|
def __init__(self, epoch_only: bool, initial_lr: float = 0.):
|
|
|
|
super().__init__(epoch_only=epoch_only)
|
|
|
|
self.lr = initial_lr
|
|
|
|
|
|
|
|
def reset(self) -> None:
|
|
|
|
pass
|
|
|
|
|
|
|
|
def update(self, lr) -> None:
|
|
|
|
self.lr = lr
|
|
|
|
|
2022-06-07 09:21:11 +00:00
|
|
|
def get_last_step_value(self) -> float:
|
|
|
|
return self.lr
|
2021-12-27 07:04:32 +00:00
|
|
|
|
|
|
|
def get_accumulated_value(self):
|
|
|
|
return self.lr
|
|
|
|
|
2022-01-21 02:44:30 +00:00
|
|
|
@staticmethod
|
2021-12-27 07:04:32 +00:00
|
|
|
def is_better(a, b) -> bool:
|
|
|
|
pass
|
|
|
|
|
|
|
|
|
|
|
|
class AccuracyMetric(Metric):
|
|
|
|
"""A metric collector for accuracy. It only works for classification
|
|
|
|
tasks.
|
|
|
|
|
2022-03-25 05:02:39 +00:00
|
|
|
Args:
|
|
|
|
epoch_only (bool): Whether the metric only read for the full epoch.
|
|
|
|
accuracy_func (:class:`typing.Callable`): Accuracy function for the classification task.
|
2021-12-27 07:04:32 +00:00
|
|
|
"""
|
2022-01-04 12:52:31 +00:00
|
|
|
|
2021-12-27 07:04:32 +00:00
|
|
|
def __init__(self, epoch_only: bool, accuracy_func: Callable):
|
|
|
|
super().__init__(epoch_only=epoch_only)
|
|
|
|
self.acc = accuracy_func
|
|
|
|
self.last_step_sum = torch.zeros(1, device=get_current_device())
|
|
|
|
self.last_step_correct = torch.zeros(1, device=get_current_device())
|
|
|
|
self.accumulated_sum = torch.zeros(1, device=get_current_device())
|
|
|
|
self.accumulated_correct = torch.zeros(1, device=get_current_device())
|
|
|
|
|
|
|
|
def reset(self) -> None:
|
|
|
|
self.last_step_sum.zero_()
|
|
|
|
self.last_step_correct.zero_()
|
|
|
|
self.accumulated_sum.zero_()
|
|
|
|
self.accumulated_correct.zero_()
|
|
|
|
|
2021-12-29 15:32:10 +00:00
|
|
|
def update(self, logits, targets, batch_size) -> None:
|
2021-12-27 07:04:32 +00:00
|
|
|
"""Updates last step accuracy and accumulated accuracy with current logits
|
|
|
|
and labels. It expects the output has logits and labels.
|
|
|
|
|
2022-03-25 05:02:39 +00:00
|
|
|
Args:
|
|
|
|
logits (:class:`torch.tensor`): The logits output of the model.
|
|
|
|
targets (:class:`torch.tensor`): Real labels of the dataset.
|
|
|
|
batch_size (int): Batch size of the task.
|
2021-12-27 07:04:32 +00:00
|
|
|
"""
|
|
|
|
if isinstance(logits, (list, tuple)):
|
|
|
|
logits = logits[0]
|
|
|
|
if isinstance(targets, (list, tuple)):
|
|
|
|
targets = targets[0]
|
|
|
|
# update
|
|
|
|
correct = self.acc(logits, targets)
|
|
|
|
|
2021-12-29 15:32:10 +00:00
|
|
|
self.last_step_sum.fill_(batch_size)
|
2021-12-27 07:04:32 +00:00
|
|
|
self.last_step_correct.fill_(correct)
|
|
|
|
self.accumulated_sum += self.last_step_sum
|
|
|
|
self.accumulated_correct += self.last_step_correct
|
|
|
|
|
2022-06-07 09:21:11 +00:00
|
|
|
def get_last_step_value(self) -> float:
|
2021-12-27 07:04:32 +00:00
|
|
|
self.last_step_sum = all_reduce(self.last_step_sum, ParallelMode.DATA)
|
|
|
|
self.last_step_correct = all_reduce(self.last_step_correct, ParallelMode.DATA)
|
2022-06-07 09:21:11 +00:00
|
|
|
return _format_number((self.last_step_correct / self.last_step_sum).cpu().item())
|
2021-12-27 07:04:32 +00:00
|
|
|
|
|
|
|
def get_accumulated_value(self):
|
|
|
|
self.accumulated_sum = all_reduce(self.accumulated_sum, ParallelMode.DATA)
|
|
|
|
self.accumulated_correct = all_reduce(self.accumulated_correct, ParallelMode.DATA)
|
|
|
|
return (self.accumulated_correct / self.accumulated_sum).item()
|
|
|
|
|
2022-01-21 02:44:30 +00:00
|
|
|
@staticmethod
|
2021-12-27 07:04:32 +00:00
|
|
|
def is_better(a, b) -> bool:
|
|
|
|
return a > b
|
2021-10-28 16:21:23 +00:00
|
|
|
|
|
|
|
|
|
|
|
class MetricHook(BaseHook):
|
2023-08-31 05:51:28 +00:00
|
|
|
"""Specialized hook classes for :class:`Metric`.
|
|
|
|
Some help metric collectors initialize, reset and
|
|
|
|
update their states. Others are used to display and
|
2021-10-28 16:21:23 +00:00
|
|
|
record the metric.
|
|
|
|
|
2022-03-25 05:02:39 +00:00
|
|
|
Args:
|
|
|
|
priority (int): Priority in the printing, hooks with small priority will be printed in front
|
|
|
|
defaults to 1. If different hooks share same priority, the order of printing would
|
|
|
|
depend on the hooks order in the hook list.
|
2021-10-28 16:21:23 +00:00
|
|
|
"""
|
2022-01-04 12:52:31 +00:00
|
|
|
|
2021-12-27 07:04:32 +00:00
|
|
|
def __init__(
|
|
|
|
self,
|
|
|
|
priority: int,
|
|
|
|
):
|
Develop/experiments (#59)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
* Split conv2d, class token, positional embedding in 2d, Fix random number in ddp
Fix convergence in cifar10, Imagenet1000
* Integrate 1d tensor parallel in Colossal-AI (#39)
* fixed 1D and 2D convergence (#38)
* optimized 2D operations
* fixed 1D ViT convergence problem
* Feature/ddp (#49)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* support torch ddp
* fix loss accumulation
* add log for ddp
* change seed
* modify timing hook
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* Feature/pipeline (#40)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* optimize communication of pipeline parallel
* fix grad clip for pipeline
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* optimized 3d layer to fix slow computation ; tested imagenet performance with 3d; reworked lr_scheduler config definition; fixed launch args; fixed some printing issues; simplified apis of 3d layers (#51)
* Update 2.5d layer code to get a similar accuracy on imagenet-1k dataset
* update api for better usability (#58)
update api for better usability
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
Co-authored-by: puck_WCR <46049915+WANG-CR@users.noreply.github.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
Co-authored-by: アマデウス <kurisusnowdeng@users.noreply.github.com>
Co-authored-by: BoxiangW <45734921+BoxiangW@users.noreply.github.com>
2021-12-09 07:08:29 +00:00
|
|
|
super().__init__(priority)
|
2021-11-18 11:45:06 +00:00
|
|
|
self._is_stage_to_compute = is_no_pp_or_last_stage()
|
2021-10-28 16:21:23 +00:00
|
|
|
|
Develop/experiments (#59)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
* Split conv2d, class token, positional embedding in 2d, Fix random number in ddp
Fix convergence in cifar10, Imagenet1000
* Integrate 1d tensor parallel in Colossal-AI (#39)
* fixed 1D and 2D convergence (#38)
* optimized 2D operations
* fixed 1D ViT convergence problem
* Feature/ddp (#49)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* support torch ddp
* fix loss accumulation
* add log for ddp
* change seed
* modify timing hook
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* Feature/pipeline (#40)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* optimize communication of pipeline parallel
* fix grad clip for pipeline
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* optimized 3d layer to fix slow computation ; tested imagenet performance with 3d; reworked lr_scheduler config definition; fixed launch args; fixed some printing issues; simplified apis of 3d layers (#51)
* Update 2.5d layer code to get a similar accuracy on imagenet-1k dataset
* update api for better usability (#58)
update api for better usability
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
Co-authored-by: puck_WCR <46049915+WANG-CR@users.noreply.github.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
Co-authored-by: アマデウス <kurisusnowdeng@users.noreply.github.com>
Co-authored-by: BoxiangW <45734921+BoxiangW@users.noreply.github.com>
2021-12-09 07:08:29 +00:00
|
|
|
def _check_metric_states_initialization(self, trainer):
|
|
|
|
if 'metrics' not in trainer.states:
|
|
|
|
self.init_runner_states(trainer, 'metrics', dict(train={}, test={}))
|
2021-10-28 16:21:23 +00:00
|
|
|
|
|
|
|
|
|
|
|
@HOOKS.register_module
|
|
|
|
class LossHook(MetricHook):
|
|
|
|
"""Specialized hook class for :class:`Loss`.
|
|
|
|
|
2022-03-25 05:02:39 +00:00
|
|
|
Args:
|
|
|
|
priority (int, optional): Priority in the printing, hooks with small priority will be printed in front
|
|
|
|
defaults to 0. If different hooks share same priority, the order of printing would
|
|
|
|
depend on the hooks order in the hook list.
|
2021-10-28 16:21:23 +00:00
|
|
|
"""
|
2022-01-04 12:52:31 +00:00
|
|
|
|
Develop/experiments (#59)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
* Split conv2d, class token, positional embedding in 2d, Fix random number in ddp
Fix convergence in cifar10, Imagenet1000
* Integrate 1d tensor parallel in Colossal-AI (#39)
* fixed 1D and 2D convergence (#38)
* optimized 2D operations
* fixed 1D ViT convergence problem
* Feature/ddp (#49)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* support torch ddp
* fix loss accumulation
* add log for ddp
* change seed
* modify timing hook
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* Feature/pipeline (#40)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* optimize communication of pipeline parallel
* fix grad clip for pipeline
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* optimized 3d layer to fix slow computation ; tested imagenet performance with 3d; reworked lr_scheduler config definition; fixed launch args; fixed some printing issues; simplified apis of 3d layers (#51)
* Update 2.5d layer code to get a similar accuracy on imagenet-1k dataset
* update api for better usability (#58)
update api for better usability
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
Co-authored-by: puck_WCR <46049915+WANG-CR@users.noreply.github.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
Co-authored-by: アマデウス <kurisusnowdeng@users.noreply.github.com>
Co-authored-by: BoxiangW <45734921+BoxiangW@users.noreply.github.com>
2021-12-09 07:08:29 +00:00
|
|
|
def __init__(self, priority: int = 0):
|
|
|
|
super().__init__(priority)
|
|
|
|
|
|
|
|
def after_hook_is_attached(self, trainer):
|
|
|
|
self._check_metric_states_initialization(trainer)
|
2021-10-28 16:21:23 +00:00
|
|
|
|
2021-11-18 11:45:06 +00:00
|
|
|
if self._is_stage_to_compute:
|
2021-12-27 07:04:32 +00:00
|
|
|
self.train_loss = LossMetric(epoch_only=False)
|
|
|
|
self.test_loss = LossMetric(epoch_only=True)
|
2021-10-28 16:21:23 +00:00
|
|
|
|
|
|
|
# register the metric calculator
|
2021-12-27 07:04:32 +00:00
|
|
|
trainer.states['metrics']['train']['Loss'] = self.train_loss
|
|
|
|
trainer.states['metrics']['test']['Loss'] = self.test_loss
|
2021-10-28 16:21:23 +00:00
|
|
|
|
Develop/experiments (#59)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
* Split conv2d, class token, positional embedding in 2d, Fix random number in ddp
Fix convergence in cifar10, Imagenet1000
* Integrate 1d tensor parallel in Colossal-AI (#39)
* fixed 1D and 2D convergence (#38)
* optimized 2D operations
* fixed 1D ViT convergence problem
* Feature/ddp (#49)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* support torch ddp
* fix loss accumulation
* add log for ddp
* change seed
* modify timing hook
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* Feature/pipeline (#40)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* optimize communication of pipeline parallel
* fix grad clip for pipeline
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* optimized 3d layer to fix slow computation ; tested imagenet performance with 3d; reworked lr_scheduler config definition; fixed launch args; fixed some printing issues; simplified apis of 3d layers (#51)
* Update 2.5d layer code to get a similar accuracy on imagenet-1k dataset
* update api for better usability (#58)
update api for better usability
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
Co-authored-by: puck_WCR <46049915+WANG-CR@users.noreply.github.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
Co-authored-by: アマデウス <kurisusnowdeng@users.noreply.github.com>
Co-authored-by: BoxiangW <45734921+BoxiangW@users.noreply.github.com>
2021-12-09 07:08:29 +00:00
|
|
|
def before_train_epoch(self, trainer):
|
2021-11-18 11:45:06 +00:00
|
|
|
if self._is_stage_to_compute:
|
|
|
|
self.train_loss.reset()
|
2021-10-28 16:21:23 +00:00
|
|
|
|
Develop/experiments (#59)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
* Split conv2d, class token, positional embedding in 2d, Fix random number in ddp
Fix convergence in cifar10, Imagenet1000
* Integrate 1d tensor parallel in Colossal-AI (#39)
* fixed 1D and 2D convergence (#38)
* optimized 2D operations
* fixed 1D ViT convergence problem
* Feature/ddp (#49)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* support torch ddp
* fix loss accumulation
* add log for ddp
* change seed
* modify timing hook
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* Feature/pipeline (#40)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* optimize communication of pipeline parallel
* fix grad clip for pipeline
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* optimized 3d layer to fix slow computation ; tested imagenet performance with 3d; reworked lr_scheduler config definition; fixed launch args; fixed some printing issues; simplified apis of 3d layers (#51)
* Update 2.5d layer code to get a similar accuracy on imagenet-1k dataset
* update api for better usability (#58)
update api for better usability
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
Co-authored-by: puck_WCR <46049915+WANG-CR@users.noreply.github.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
Co-authored-by: アマデウス <kurisusnowdeng@users.noreply.github.com>
Co-authored-by: BoxiangW <45734921+BoxiangW@users.noreply.github.com>
2021-12-09 07:08:29 +00:00
|
|
|
def after_train_iter(self, trainer, logits, label, loss):
|
2021-11-18 11:45:06 +00:00
|
|
|
if self._is_stage_to_compute:
|
|
|
|
self.train_loss.update(loss)
|
2021-10-28 16:21:23 +00:00
|
|
|
|
Develop/experiments (#59)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
* Split conv2d, class token, positional embedding in 2d, Fix random number in ddp
Fix convergence in cifar10, Imagenet1000
* Integrate 1d tensor parallel in Colossal-AI (#39)
* fixed 1D and 2D convergence (#38)
* optimized 2D operations
* fixed 1D ViT convergence problem
* Feature/ddp (#49)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* support torch ddp
* fix loss accumulation
* add log for ddp
* change seed
* modify timing hook
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* Feature/pipeline (#40)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* optimize communication of pipeline parallel
* fix grad clip for pipeline
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* optimized 3d layer to fix slow computation ; tested imagenet performance with 3d; reworked lr_scheduler config definition; fixed launch args; fixed some printing issues; simplified apis of 3d layers (#51)
* Update 2.5d layer code to get a similar accuracy on imagenet-1k dataset
* update api for better usability (#58)
update api for better usability
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
Co-authored-by: puck_WCR <46049915+WANG-CR@users.noreply.github.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
Co-authored-by: アマデウス <kurisusnowdeng@users.noreply.github.com>
Co-authored-by: BoxiangW <45734921+BoxiangW@users.noreply.github.com>
2021-12-09 07:08:29 +00:00
|
|
|
def before_test_epoch(self, trainer):
|
2021-11-18 11:45:06 +00:00
|
|
|
if self._is_stage_to_compute:
|
|
|
|
self.test_loss.reset()
|
2021-10-28 16:21:23 +00:00
|
|
|
|
Develop/experiments (#59)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
* Split conv2d, class token, positional embedding in 2d, Fix random number in ddp
Fix convergence in cifar10, Imagenet1000
* Integrate 1d tensor parallel in Colossal-AI (#39)
* fixed 1D and 2D convergence (#38)
* optimized 2D operations
* fixed 1D ViT convergence problem
* Feature/ddp (#49)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* support torch ddp
* fix loss accumulation
* add log for ddp
* change seed
* modify timing hook
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* Feature/pipeline (#40)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* optimize communication of pipeline parallel
* fix grad clip for pipeline
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* optimized 3d layer to fix slow computation ; tested imagenet performance with 3d; reworked lr_scheduler config definition; fixed launch args; fixed some printing issues; simplified apis of 3d layers (#51)
* Update 2.5d layer code to get a similar accuracy on imagenet-1k dataset
* update api for better usability (#58)
update api for better usability
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
Co-authored-by: puck_WCR <46049915+WANG-CR@users.noreply.github.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
Co-authored-by: アマデウス <kurisusnowdeng@users.noreply.github.com>
Co-authored-by: BoxiangW <45734921+BoxiangW@users.noreply.github.com>
2021-12-09 07:08:29 +00:00
|
|
|
def after_test_iter(self, trainer, logits, label, loss):
|
2021-11-18 11:45:06 +00:00
|
|
|
if self._is_stage_to_compute:
|
|
|
|
self.test_loss.update(loss)
|
2021-10-28 16:21:23 +00:00
|
|
|
|
|
|
|
|
Develop/experiments (#59)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
* Split conv2d, class token, positional embedding in 2d, Fix random number in ddp
Fix convergence in cifar10, Imagenet1000
* Integrate 1d tensor parallel in Colossal-AI (#39)
* fixed 1D and 2D convergence (#38)
* optimized 2D operations
* fixed 1D ViT convergence problem
* Feature/ddp (#49)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* support torch ddp
* fix loss accumulation
* add log for ddp
* change seed
* modify timing hook
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* Feature/pipeline (#40)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* optimize communication of pipeline parallel
* fix grad clip for pipeline
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* optimized 3d layer to fix slow computation ; tested imagenet performance with 3d; reworked lr_scheduler config definition; fixed launch args; fixed some printing issues; simplified apis of 3d layers (#51)
* Update 2.5d layer code to get a similar accuracy on imagenet-1k dataset
* update api for better usability (#58)
update api for better usability
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
Co-authored-by: puck_WCR <46049915+WANG-CR@users.noreply.github.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
Co-authored-by: アマデウス <kurisusnowdeng@users.noreply.github.com>
Co-authored-by: BoxiangW <45734921+BoxiangW@users.noreply.github.com>
2021-12-09 07:08:29 +00:00
|
|
|
@HOOKS.register_module
|
2021-12-27 07:04:32 +00:00
|
|
|
class AccuracyHook(MetricHook):
|
|
|
|
"""Specialized hook class for :class:`Accuracy`.
|
2021-10-28 16:21:23 +00:00
|
|
|
|
2022-03-25 05:02:39 +00:00
|
|
|
Args:
|
|
|
|
accuracy_func (:class:`typing.Callable`): Accuracy function for the classification task.
|
|
|
|
priority (int, optional): Priority in the printing, hooks with small priority will be printed in front
|
|
|
|
defaults to 0. If different hooks share same priority, the order of printing would
|
|
|
|
depend on the hooks order in the hook list.
|
2021-10-28 16:21:23 +00:00
|
|
|
"""
|
2022-01-04 12:52:31 +00:00
|
|
|
|
2021-12-27 07:04:32 +00:00
|
|
|
def __init__(self, accuracy_func: Callable, priority: int = 0):
|
Develop/experiments (#59)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
* Split conv2d, class token, positional embedding in 2d, Fix random number in ddp
Fix convergence in cifar10, Imagenet1000
* Integrate 1d tensor parallel in Colossal-AI (#39)
* fixed 1D and 2D convergence (#38)
* optimized 2D operations
* fixed 1D ViT convergence problem
* Feature/ddp (#49)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* support torch ddp
* fix loss accumulation
* add log for ddp
* change seed
* modify timing hook
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* Feature/pipeline (#40)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* optimize communication of pipeline parallel
* fix grad clip for pipeline
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* optimized 3d layer to fix slow computation ; tested imagenet performance with 3d; reworked lr_scheduler config definition; fixed launch args; fixed some printing issues; simplified apis of 3d layers (#51)
* Update 2.5d layer code to get a similar accuracy on imagenet-1k dataset
* update api for better usability (#58)
update api for better usability
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
Co-authored-by: puck_WCR <46049915+WANG-CR@users.noreply.github.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
Co-authored-by: アマデウス <kurisusnowdeng@users.noreply.github.com>
Co-authored-by: BoxiangW <45734921+BoxiangW@users.noreply.github.com>
2021-12-09 07:08:29 +00:00
|
|
|
super().__init__(priority)
|
2021-12-27 07:04:32 +00:00
|
|
|
self.accuracy_func = accuracy_func
|
2021-10-28 16:21:23 +00:00
|
|
|
|
Develop/experiments (#59)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
* Split conv2d, class token, positional embedding in 2d, Fix random number in ddp
Fix convergence in cifar10, Imagenet1000
* Integrate 1d tensor parallel in Colossal-AI (#39)
* fixed 1D and 2D convergence (#38)
* optimized 2D operations
* fixed 1D ViT convergence problem
* Feature/ddp (#49)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* support torch ddp
* fix loss accumulation
* add log for ddp
* change seed
* modify timing hook
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* Feature/pipeline (#40)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* optimize communication of pipeline parallel
* fix grad clip for pipeline
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* optimized 3d layer to fix slow computation ; tested imagenet performance with 3d; reworked lr_scheduler config definition; fixed launch args; fixed some printing issues; simplified apis of 3d layers (#51)
* Update 2.5d layer code to get a similar accuracy on imagenet-1k dataset
* update api for better usability (#58)
update api for better usability
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
Co-authored-by: puck_WCR <46049915+WANG-CR@users.noreply.github.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
Co-authored-by: アマデウス <kurisusnowdeng@users.noreply.github.com>
Co-authored-by: BoxiangW <45734921+BoxiangW@users.noreply.github.com>
2021-12-09 07:08:29 +00:00
|
|
|
def after_hook_is_attached(self, trainer):
|
|
|
|
self._check_metric_states_initialization(trainer)
|
2021-11-18 11:45:06 +00:00
|
|
|
if self._is_stage_to_compute:
|
2021-12-27 07:04:32 +00:00
|
|
|
self.metric = AccuracyMetric(epoch_only=True, accuracy_func=self.accuracy_func)
|
2021-10-28 16:21:23 +00:00
|
|
|
|
|
|
|
# register the metric
|
2021-12-27 07:04:32 +00:00
|
|
|
trainer.states['metrics']['test']['Accuracy'] = self.metric
|
2021-10-28 16:21:23 +00:00
|
|
|
|
Develop/experiments (#59)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
* Split conv2d, class token, positional embedding in 2d, Fix random number in ddp
Fix convergence in cifar10, Imagenet1000
* Integrate 1d tensor parallel in Colossal-AI (#39)
* fixed 1D and 2D convergence (#38)
* optimized 2D operations
* fixed 1D ViT convergence problem
* Feature/ddp (#49)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* support torch ddp
* fix loss accumulation
* add log for ddp
* change seed
* modify timing hook
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* Feature/pipeline (#40)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* optimize communication of pipeline parallel
* fix grad clip for pipeline
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* optimized 3d layer to fix slow computation ; tested imagenet performance with 3d; reworked lr_scheduler config definition; fixed launch args; fixed some printing issues; simplified apis of 3d layers (#51)
* Update 2.5d layer code to get a similar accuracy on imagenet-1k dataset
* update api for better usability (#58)
update api for better usability
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
Co-authored-by: puck_WCR <46049915+WANG-CR@users.noreply.github.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
Co-authored-by: アマデウス <kurisusnowdeng@users.noreply.github.com>
Co-authored-by: BoxiangW <45734921+BoxiangW@users.noreply.github.com>
2021-12-09 07:08:29 +00:00
|
|
|
def before_test(self, trainer):
|
2021-11-18 11:45:06 +00:00
|
|
|
if self._is_stage_to_compute:
|
2021-10-28 16:21:23 +00:00
|
|
|
self.metric.reset()
|
|
|
|
|
2021-12-27 07:04:32 +00:00
|
|
|
def after_test_iter(self, trainer, logits, targets, *args):
|
2021-11-18 11:45:06 +00:00
|
|
|
if self._is_stage_to_compute:
|
2022-04-07 07:54:14 +00:00
|
|
|
batch_size = trainer.engine.schedule.batch_size
|
2021-12-29 15:32:10 +00:00
|
|
|
self.metric.update(logits, targets, batch_size)
|
2021-12-27 07:04:32 +00:00
|
|
|
|
|
|
|
|
|
|
|
class ThroughputMetric(Metric):
|
2022-01-10 10:05:58 +00:00
|
|
|
"""Metric for :class:`Throughput`.
|
|
|
|
|
2022-03-25 05:02:39 +00:00
|
|
|
Args:
|
|
|
|
epoch_only (bool): Whether the metric only read for the full epoch.
|
2022-01-10 10:05:58 +00:00
|
|
|
"""
|
2022-04-19 02:13:08 +00:00
|
|
|
|
2022-04-20 02:05:39 +00:00
|
|
|
def __init__(self, epoch_only: bool, ignored_steps: int = 0, tflop_per_step: int = 0, use_local: bool = False):
|
2021-12-27 07:04:32 +00:00
|
|
|
super().__init__(epoch_only=epoch_only)
|
2022-02-14 03:15:02 +00:00
|
|
|
self.ignored_steps = ignored_steps
|
|
|
|
self.cur_steps = 0
|
2021-12-27 07:04:32 +00:00
|
|
|
self.accumulated_num_samples = torch.zeros(1, device=get_current_device())
|
|
|
|
self.accumulated_used_time = torch.zeros(1, device=get_current_device())
|
|
|
|
self.last_step_num_samples = torch.zeros(1, device=get_current_device())
|
|
|
|
self.last_step_used_time = torch.zeros(1, device=get_current_device())
|
2022-04-19 02:13:08 +00:00
|
|
|
self._tflop_per_step = tflop_per_step
|
2022-04-20 02:05:39 +00:00
|
|
|
self._use_local = use_local
|
2021-12-27 07:04:32 +00:00
|
|
|
|
|
|
|
def reset(self) -> None:
|
2022-02-14 03:15:02 +00:00
|
|
|
# self.cur_steps = 0
|
2021-12-27 07:04:32 +00:00
|
|
|
self.accumulated_num_samples.zero_()
|
|
|
|
self.accumulated_used_time.zero_()
|
|
|
|
self.last_step_num_samples.zero_()
|
|
|
|
self.last_step_used_time.zero_()
|
|
|
|
|
2021-12-29 15:32:10 +00:00
|
|
|
def update(self, num_samples, time) -> None:
|
2022-02-14 03:15:02 +00:00
|
|
|
self.cur_steps += 1
|
2021-12-29 15:32:10 +00:00
|
|
|
self.last_step_num_samples.fill_(num_samples)
|
2021-12-27 07:04:32 +00:00
|
|
|
self.last_step_used_time.fill_(time)
|
2022-02-14 03:15:02 +00:00
|
|
|
if self.cur_steps >= self.ignored_steps:
|
|
|
|
self.accumulated_num_samples += self.last_step_num_samples
|
|
|
|
self.accumulated_used_time += self.last_step_used_time
|
2021-12-27 07:04:32 +00:00
|
|
|
|
2022-06-07 09:21:11 +00:00
|
|
|
def get_last_step_value(self) -> float:
|
|
|
|
if self._use_local:
|
|
|
|
self.last_step_num_samples *= gpc.get_world_size(ParallelMode.DATA)
|
|
|
|
else:
|
|
|
|
self.last_step_used_time = all_reduce(self.last_step_used_time, ParallelMode.DATA) / \
|
2023-09-04 11:56:42 +00:00
|
|
|
gpc.get_world_size(ParallelMode.DATA)
|
2022-06-07 09:21:11 +00:00
|
|
|
self.last_step_num_samples = all_reduce(self.last_step_num_samples, ParallelMode.DATA)
|
|
|
|
|
|
|
|
sample_per_sec = _format_number(self.last_step_num_samples / (self.last_step_used_time + 1e-12).item())
|
|
|
|
return sample_per_sec
|
|
|
|
|
|
|
|
def get_last_step_info(self) -> str:
|
2022-04-20 02:05:39 +00:00
|
|
|
if self._use_local:
|
|
|
|
self.last_step_num_samples *= gpc.get_world_size(ParallelMode.DATA)
|
|
|
|
else:
|
|
|
|
self.last_step_used_time = all_reduce(self.last_step_used_time, ParallelMode.DATA) / \
|
2023-09-04 11:56:42 +00:00
|
|
|
gpc.get_world_size(ParallelMode.DATA)
|
2022-04-20 02:05:39 +00:00
|
|
|
self.last_step_num_samples = all_reduce(self.last_step_num_samples, ParallelMode.DATA)
|
|
|
|
|
2022-04-19 02:13:08 +00:00
|
|
|
sample_per_sec = _format_number(self.last_step_num_samples / (self.last_step_used_time + 1e-12).item())
|
|
|
|
if self._tflop_per_step > 0:
|
|
|
|
tflops = _format_number(self._tflop_per_step / (self.last_step_used_time.item() + 1e-12))
|
|
|
|
return f"{sample_per_sec} sample_per_sec, {tflops} Tflops"
|
|
|
|
else:
|
|
|
|
return f"{sample_per_sec} sample_per_sec"
|
|
|
|
|
|
|
|
def get_accumulated_value(self) -> float:
|
2021-12-27 07:04:32 +00:00
|
|
|
self.accumulated_used_time = all_reduce(self.accumulated_used_time, ParallelMode.DATA) / \
|
|
|
|
gpc.get_world_size(ParallelMode.DATA)
|
|
|
|
self.accumulated_num_samples = all_reduce(self.accumulated_num_samples, ParallelMode.DATA)
|
|
|
|
return (self.accumulated_num_samples / (self.accumulated_used_time + 1e-12)).item()
|
|
|
|
|
2022-01-21 02:44:30 +00:00
|
|
|
@staticmethod
|
2021-12-27 07:04:32 +00:00
|
|
|
def is_better(a, b) -> bool:
|
|
|
|
pass
|
2021-10-28 16:21:23 +00:00
|
|
|
|
|
|
|
|
|
|
|
@HOOKS.register_module
|
2021-12-27 07:04:32 +00:00
|
|
|
class ThroughputHook(MetricHook):
|
2022-03-25 05:02:39 +00:00
|
|
|
"""Specialized hook class for :class:`Throughput`. Hook to measure execution throughput (samples/sec).
|
2022-01-10 10:05:58 +00:00
|
|
|
|
2022-03-25 05:02:39 +00:00
|
|
|
Args:
|
|
|
|
ignored_steps (int, optional): the number of initial training steps to ignore.
|
|
|
|
priority (int, optional): Priority in the printing, hooks with small priority will be printed in front
|
|
|
|
defaults to 10. If different hooks share same priority, the order of printing would
|
|
|
|
depend on the hooks order in the hook list.
|
2022-04-20 02:05:39 +00:00
|
|
|
tflop_per_step(int, optional): tera floating point operations per step.
|
|
|
|
use_local (bool, optional): Whether to use local time for throughput calculation.
|
2022-01-10 10:05:58 +00:00
|
|
|
"""
|
2022-04-19 02:13:08 +00:00
|
|
|
|
2022-04-20 02:05:39 +00:00
|
|
|
def __init__(self, ignored_steps: int = 0, priority: int = 10, tflop_per_step: int = 0, use_local=False):
|
Develop/experiments (#59)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
* Split conv2d, class token, positional embedding in 2d, Fix random number in ddp
Fix convergence in cifar10, Imagenet1000
* Integrate 1d tensor parallel in Colossal-AI (#39)
* fixed 1D and 2D convergence (#38)
* optimized 2D operations
* fixed 1D ViT convergence problem
* Feature/ddp (#49)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* support torch ddp
* fix loss accumulation
* add log for ddp
* change seed
* modify timing hook
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* Feature/pipeline (#40)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* optimize communication of pipeline parallel
* fix grad clip for pipeline
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* optimized 3d layer to fix slow computation ; tested imagenet performance with 3d; reworked lr_scheduler config definition; fixed launch args; fixed some printing issues; simplified apis of 3d layers (#51)
* Update 2.5d layer code to get a similar accuracy on imagenet-1k dataset
* update api for better usability (#58)
update api for better usability
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
Co-authored-by: puck_WCR <46049915+WANG-CR@users.noreply.github.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
Co-authored-by: アマデウス <kurisusnowdeng@users.noreply.github.com>
Co-authored-by: BoxiangW <45734921+BoxiangW@users.noreply.github.com>
2021-12-09 07:08:29 +00:00
|
|
|
super().__init__(priority)
|
2022-02-14 03:15:02 +00:00
|
|
|
self.ignored_steps = ignored_steps
|
2022-04-19 02:13:08 +00:00
|
|
|
self._tflop_per_step = tflop_per_step
|
2022-04-20 02:05:39 +00:00
|
|
|
self._use_local = use_local
|
2021-10-28 16:21:23 +00:00
|
|
|
|
Develop/experiments (#59)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
* Split conv2d, class token, positional embedding in 2d, Fix random number in ddp
Fix convergence in cifar10, Imagenet1000
* Integrate 1d tensor parallel in Colossal-AI (#39)
* fixed 1D and 2D convergence (#38)
* optimized 2D operations
* fixed 1D ViT convergence problem
* Feature/ddp (#49)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* support torch ddp
* fix loss accumulation
* add log for ddp
* change seed
* modify timing hook
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* Feature/pipeline (#40)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* optimize communication of pipeline parallel
* fix grad clip for pipeline
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* optimized 3d layer to fix slow computation ; tested imagenet performance with 3d; reworked lr_scheduler config definition; fixed launch args; fixed some printing issues; simplified apis of 3d layers (#51)
* Update 2.5d layer code to get a similar accuracy on imagenet-1k dataset
* update api for better usability (#58)
update api for better usability
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
Co-authored-by: puck_WCR <46049915+WANG-CR@users.noreply.github.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
Co-authored-by: アマデウス <kurisusnowdeng@users.noreply.github.com>
Co-authored-by: BoxiangW <45734921+BoxiangW@users.noreply.github.com>
2021-12-09 07:08:29 +00:00
|
|
|
def after_hook_is_attached(self, trainer):
|
|
|
|
self._check_metric_states_initialization(trainer)
|
2021-11-18 11:45:06 +00:00
|
|
|
if self._is_stage_to_compute:
|
2022-04-19 02:13:08 +00:00
|
|
|
self.metric = ThroughputMetric(epoch_only=True,
|
|
|
|
ignored_steps=self.ignored_steps,
|
2022-04-20 02:05:39 +00:00
|
|
|
tflop_per_step=self._tflop_per_step,
|
|
|
|
use_local=self._use_local)
|
2021-10-28 16:21:23 +00:00
|
|
|
|
|
|
|
# register the metric
|
2021-12-27 07:04:32 +00:00
|
|
|
trainer.states['metrics']['train']['Throughput'] = self.metric
|
|
|
|
trainer.states['metrics']['test']['Throughput'] = self.metric
|
2021-10-28 16:21:23 +00:00
|
|
|
|
2021-12-27 07:04:32 +00:00
|
|
|
def before_train_epoch(self, trainer):
|
2022-01-04 12:52:31 +00:00
|
|
|
if self._is_stage_to_compute:
|
|
|
|
self.metric.reset()
|
2021-10-28 16:21:23 +00:00
|
|
|
|
2021-12-29 15:32:10 +00:00
|
|
|
def after_train_iter(self, trainer, *args):
|
2022-01-04 12:52:31 +00:00
|
|
|
if self._is_stage_to_compute:
|
2022-04-19 02:13:08 +00:00
|
|
|
self.metric.update(trainer.engine.schedule.batch_size,
|
|
|
|
trainer._timer.get_timer('Train-step').get_elapsed_time())
|
2021-10-28 16:21:23 +00:00
|
|
|
|
Develop/experiments (#59)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
* Split conv2d, class token, positional embedding in 2d, Fix random number in ddp
Fix convergence in cifar10, Imagenet1000
* Integrate 1d tensor parallel in Colossal-AI (#39)
* fixed 1D and 2D convergence (#38)
* optimized 2D operations
* fixed 1D ViT convergence problem
* Feature/ddp (#49)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* support torch ddp
* fix loss accumulation
* add log for ddp
* change seed
* modify timing hook
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* Feature/pipeline (#40)
* remove redundancy func in setup (#19) (#20)
* use env to control the language of doc (#24) (#25)
* Support TP-compatible Torch AMP and Update trainer API (#27)
* Add gradient accumulation, fix lr scheduler
* fix FP16 optimizer and adapted torch amp with tensor parallel (#18)
* fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes
* fixed trainer
* Revert "fixed trainer"
This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097.
* improved consistency between trainer, engine and schedule (#23)
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
* add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29)
* add explanation for ViT example (#35) (#36)
* optimize communication of pipeline parallel
* fix grad clip for pipeline
Co-authored-by: Frank Lee <somerlee.9@gmail.com>
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
* optimized 3d layer to fix slow computation ; tested imagenet performance with 3d; reworked lr_scheduler config definition; fixed launch args; fixed some printing issues; simplified apis of 3d layers (#51)
* Update 2.5d layer code to get a similar accuracy on imagenet-1k dataset
* update api for better usability (#58)
update api for better usability
Co-authored-by: 1SAA <c2h214748@gmail.com>
Co-authored-by: ver217 <lhx0217@gmail.com>
Co-authored-by: puck_WCR <46049915+WANG-CR@users.noreply.github.com>
Co-authored-by: binmakeswell <binmakeswell@gmail.com>
Co-authored-by: アマデウス <kurisusnowdeng@users.noreply.github.com>
Co-authored-by: BoxiangW <45734921+BoxiangW@users.noreply.github.com>
2021-12-09 07:08:29 +00:00
|
|
|
def before_test(self, trainer):
|
2022-01-04 12:52:31 +00:00
|
|
|
if self._is_stage_to_compute:
|
|
|
|
self.metric.reset()
|
2021-10-28 16:21:23 +00:00
|
|
|
|
2021-12-29 15:32:10 +00:00
|
|
|
def after_test_iter(self, trainer, *args):
|
2022-01-04 12:52:31 +00:00
|
|
|
if self._is_stage_to_compute:
|
2022-04-19 02:13:08 +00:00
|
|
|
self.metric.update(trainer.engine.schedule.batch_size,
|
|
|
|
trainer._timer.get_timer('Test-step').get_elapsed_time())
|