From a5eb179153794bfa1e349137af048290101d0946 Mon Sep 17 00:00:00 2001
From: Zaida Zhou <58739961+zhouzaida@users.noreply.github.com>
Date: Sat, 8 Oct 2022 11:19:30 +0800
Subject: [PATCH] [Docs] Fix the docstring of model sub-package (#573)
---
docs/zh_cn/tutorials/hook.md | 2 +-
mmengine/model/averaged_model.py | 7 +++---
mmengine/model/base_model/base_model.py | 8 +++----
mmengine/model/wrappers/distributed.py | 23 ++++++++++---------
.../model/wrappers/seperate_distributed.py | 7 +++---
5 files changed, 25 insertions(+), 22 deletions(-)
diff --git a/docs/zh_cn/tutorials/hook.md b/docs/zh_cn/tutorials/hook.md
index d74c426d..14913abe 100644
--- a/docs/zh_cn/tutorials/hook.md
+++ b/docs/zh_cn/tutorials/hook.md
@@ -114,7 +114,7 @@ default_hooks = dict(checkpoint=dict(type='CheckpointHook', save_best='auto'))
也å¯ä»¥ç›´æŽ¥æŒ‡å®š `save_best` çš„å€¼ä¸ºè¯„ä»·æŒ‡æ ‡ï¼Œä¾‹å¦‚åœ¨åˆ†ç±»ä»»åŠ¡ä¸ï¼Œå¯ä»¥æŒ‡å®šä¸º `save_best='top-1'`ï¼Œåˆ™ä¼šæ ¹æ® `'top-1'` 的值判æ–当å‰æƒé‡æ˜¯å¦æœ€ä¼˜ã€‚
-除了 `save_best` å‚数，和ä¿å˜æœ€ä¼˜æƒé‡ç›¸å…³çš„å‚数还有 `rule`，`greater_keys` å’Œ `less_keys`，这三者用æ¥åˆ¤æ– `save_bes` 的值是越大越好还是越å°è¶Šå¥½ã€‚例如指定了 `save_best='top-1'`，å¯ä»¥æŒ‡å®š `rule='greater'`，则表示该值越大表示æƒé‡è¶Šå¥½ã€‚
+除了 `save_best` å‚数，和ä¿å˜æœ€ä¼˜æƒé‡ç›¸å…³çš„å‚数还有 `rule`，`greater_keys` å’Œ `less_keys`，这三者用æ¥åˆ¤æ– `save_best` 的值是越大越好还是越å°è¶Šå¥½ã€‚例如指定了 `save_best='top-1'`，å¯ä»¥æŒ‡å®š `rule='greater'`，则表示该值越大表示æƒé‡è¶Šå¥½ã€‚
- 指定ä¿å˜æƒé‡çš„路径
diff --git a/mmengine/model/averaged_model.py b/mmengine/model/averaged_model.py
index 4326be1d..94da035b 100644
--- a/mmengine/model/averaged_model.py
+++ b/mmengine/model/averaged_model.py
@@ -29,12 +29,13 @@ class BaseAveragedModel(nn.Module):
In mmengine, we provide two ways to use the model averaging:
1. Use the model averaging module in hook:
- We provide an EMAHook to apply the model averaging during training.
- Add ``custom_hooks=[dict(type='EMAHook')]`` to the config or the runner.
- The hook is implemented in mmengine/hooks/ema_hook.py
+ We provide an :class:`mmengine.hooks.EMAHook` to apply the model
+ averaging during training. Add ``custom_hooks=[dict(type='EMAHook')]``
+ to the config or the runner.
2. Use the model averaging module directly in the algorithm. Take the ema
teacher in semi-supervise as an example:
+
>>> from mmengine.model import ExponentialMovingAverage
>>> student = ResNet(depth=50)
>>> # use ema model as teacher
diff --git a/mmengine/model/base_model/base_model.py b/mmengine/model/base_model/base_model.py
index 19fd7845..25ce0719 100644
--- a/mmengine/model/base_model/base_model.py
+++ b/mmengine/model/base_model/base_model.py
@@ -92,12 +92,12 @@ class BaseModel(BaseModule):
:class:`IterBasedTrainLoop` will call this method to update model
parameters. The default parameter update process is as follows:
- 1. Calls ``self.data_processor(data, training=False) to collect
- batch_inputs and corresponding data_samples(labels).
+ 1. Calls ``self.data_processor(data, training=False)`` to collect
+ batch_inputs and corresponding data_samples(labels).
2. Calls ``self(batch_inputs, data_samples, mode='loss')`` to get raw
- loss
+ loss
3. Calls ``self.parse_losses`` to get ``parsed_losses`` tensor used to
- backward and dict of loss tensor used to log messages.
+ backward and dict of loss tensor used to log messages.
4. Calls ``optim_wrapper.update_params(loss)`` to update model.
Args:
diff --git a/mmengine/model/wrappers/distributed.py b/mmengine/model/wrappers/distributed.py
index 889dd4c4..36310900 100644
--- a/mmengine/model/wrappers/distributed.py
+++ b/mmengine/model/wrappers/distributed.py
@@ -26,9 +26,10 @@ class MMDistributedDataParallel(DistributedDataParallel):
default model forward, gradient back propagation, parameter updating
logic. To take advantage of DistributedDataParallel's automatic gradient
synchronization, ``train_step`` calls ``DistributedDataParallel.forward``
- to calculate the losses, and call other methods of :obj:`BaseModel` to
+ to calculate the losses, and call other methods of :class:`BaseModel` to
pre-process data and parse losses. Finally, update model parameters by
- :obj:``OptimWrapper`` and return the loss dictionary used for logging.
+ :class:`OptimWrapper` and return the loss dictionary used
+ for logging.
- ``val_step``: Called by ``runner.val_loop`` and get the inference
results. Since there is no gradient synchronization requirement,
@@ -43,11 +44,10 @@ class MMDistributedDataParallel(DistributedDataParallel):
the computational graph with `loss` as the root.
There are two cases
- - Parameters were not used during
- forward pass.
- - Parameters were not used to produce
- loss.
- Default: False.
+ - Parameters were not used during forward pass.
+ - Parameters were not used to produce loss.
+
+ Defaults to False.
**kwargs: keyword arguments passed to ``DistributedDataParallel``.
@@ -57,8 +57,8 @@ class MMDistributedDataParallel(DistributedDataParallel):
output for single-device CUDA modules.
- dim (int): Defaults to 0.
- broadcast_buffers (bool): Flag that enables syncing (
- broadcasting) buffers of the module at beginning of the
- ``forward`` function. Defaults to True
+ broadcasting) buffers of the module at beginning of the
+ ``forward`` function. Defaults to True
- find_unused_parameters (bool): Whether to find parameters of
module, which are not in the forward graph. Defaults to False.
- process_group (ProcessGroup, optional): The process group to be
@@ -70,7 +70,8 @@ class MMDistributedDataParallel(DistributedDataParallel):
- gradient_as_bucket_view (bool): Defaults to False.
- static_graph (bool): Defaults to False.
- See more information about arguments in `https://pytorch.org/docs/stable/generated/torch.nn.parallel.DistributedDataParallel.html#torch.nn.parallel.DistributedDataParallel`_ # noqa E501
+ See more information about arguments in
+ :class:`torch.nn.parallel.DistributedDataParallel`.
Note:
If model has multiple submodules and each module has
@@ -100,7 +101,7 @@ class MMDistributedDataParallel(DistributedDataParallel):
:meth:`train_step` will perform the following steps in order:
- If :attr:`module` defines the preprocess method,
- call ``module.preprocess`` to pre-processing data.
+ call ``module.preprocess`` to pre-processing data.
- Call ``module.forward(**data)`` and get losses.
- Parse losses.
- Call ``optim_wrapper.optimizer_step`` to update parameters.
diff --git a/mmengine/model/wrappers/seperate_distributed.py b/mmengine/model/wrappers/seperate_distributed.py
index b7306326..ac9c2383 100644
--- a/mmengine/model/wrappers/seperate_distributed.py
+++ b/mmengine/model/wrappers/seperate_distributed.py
@@ -41,7 +41,7 @@ class MMSeparateDistributedDataParallel(DistributedDataParallel):
find_unused_parameters (bool): Same as that in
``torch.nn.parallel.distributed.DistributedDataParallel``.
Traverse the autograd graph of all tensors contained in returned
- value of the wrapped module’s forward function. Defaults to False.
+ value of the wrapped module's forward function. Defaults to False.
**kwargs: Keyword arguments passed to ``MMDistributedDataParallel``.
- device_ids (List[int] or torch.device, optional): CUDA devices
@@ -58,7 +58,8 @@ class MMSeparateDistributedDataParallel(DistributedDataParallel):
- gradient_as_bucket_view (bool): Defaults to False.
- static_graph (bool): Defaults to False.
- See more information about arguments in `https://pytorch.org/docs/stable/generated/torch.nn.parallel.DistributedDataParallel.html#torch.nn.parallel.DistributedDataParallel`_ # noqa E501
+ See more information about arguments in
+ :class:`torch.nn.parallel.DistributedDataParallel`.
"""
def __init__(self,
@@ -144,7 +145,7 @@ class MMSeparateDistributedDataParallel(DistributedDataParallel):
Args:
mode (bool): whether to set training mode (``True``) or evaluation
- mode (``False``). Default: ``True``.
+ mode (``False``). Defaults to ``True``.
Returns:
Module: self.
--
GitLab