Merge branch 'dev'

.pre-commit-config.yaml

@@ -29,9 +29,9 @@ repos:
     rev: 0.7.9
     hooks:
       - id: mdformat
-        args: ["--number", "--table-width", "200"]
+        args: ["--number", "--table-width", "200", '--disable-escape', 'backslash', '--disable-escape', 'link-enclosure']
         additional_dependencies:
-          - mdformat-openmmlab
+          - "mdformat-openmmlab>=0.0.4"
           - mdformat_frontmatter
           - linkify-it-py
   - repo: https://github.com/codespell-project/codespell

README.md

@@ -64,6 +64,12 @@ The MMClassification 1.0 has released! It's still unstable and in release candid
 to [the 1.x branch](https://github.com/open-mmlab/mmclassification/tree/1.x) and discuss it with us in
 [the discussion](https://github.com/open-mmlab/mmclassification/discussions).
 
+v0.25.0 was released in 06/12/2022.
+Highlights of the new version:
+
+- Support MLU backend.
+- Add `dist_train_arm.sh` for ARM device.
+
 v0.24.1 was released in 31/10/2022.
 Highlights of the new version:
 
@@ -75,13 +81,6 @@ Highlights of the new version:
 - Support **HorNet**, **EfficientFormerm**, **SwinTransformer V2** and **MViT** backbones.
 - Support Standford Cars dataset.
 
-v0.23.0 was released in 1/5/2022.
-Highlights of the new version:
-
-- Support **DenseNet**, **VAN** and **PoolFormer**, and provide pre-trained models.
-- Support training on IPU.
-- New style API docs, welcome [view it](https://mmclassification.readthedocs.io/en/master/api/models.html).
-
 Please refer to [changelog.md](docs/en/changelog.md) for more details and other release history.
 
 ## Installation

README_zh-CN.md

@@ -63,6 +63,11 @@ MMClassification 是一款基于 PyTorch 的开源图像分类工具箱，是 [O
 
 MMClassification 1.0 已经发布！目前仍在公测中，如果希望试用，请切换到 [1.x 分支](https://github.com/open-mmlab/mmclassification/tree/1.x)，并在[讨论版](https://github.com/open-mmlab/mmclassification/discussions) 参加开发讨论！
 
+2022/12/06 发布了 v0.25.0 版本
+
+- 支持 MLU 设备
+- 添加了用于 ARM 设备训练的 `dist_train_arm.sh`
+
 2022/10/31 发布了 v0.24.1 版本
 
 - 支持了华为昇腾 NPU 设备。

configs/t2t_vit/README.md

@@ -6,7 +6,7 @@
 
 ## Abstract
 
-Transformers, which are popular for language modeling, have been explored for solving vision tasks recently, \\eg, the Vision Transformer (ViT) for image classification. The ViT model splits each image into a sequence of tokens with fixed length and then applies multiple Transformer layers to model their global relation for classification. However, ViT achieves inferior performance to CNNs when trained from scratch on a midsize dataset like ImageNet. We find it is because: 1) the simple tokenization of input images fails to model the important local structure such as edges and lines among neighboring pixels, leading to low training sample efficiency; 2) the redundant attention backbone design of ViT leads to limited feature richness for fixed computation budgets and limited training samples. To overcome such limitations, we propose a new Tokens-To-Token Vision Transformer (T2T-ViT), which incorporates 1) a layer-wise Tokens-to-Token (T2T) transformation to progressively structurize the image to tokens by recursively aggregating neighboring Tokens into one Token (Tokens-to-Token), such that local structure represented by surrounding tokens can be modeled and tokens length can be reduced; 2) an efficient backbone with a deep-narrow structure for vision transformer motivated by CNN architecture design after empirical study. Notably, T2T-ViT reduces the parameter count and MACs of vanilla ViT by half, while achieving more than 3.0% improvement when trained from scratch on ImageNet. It also outperforms ResNets and achieves comparable performance with MobileNets by directly training on ImageNet. For example, T2T-ViT with comparable size to ResNet50 (21.5M parameters) can achieve 83.3% top1 accuracy in image resolution 384×384 on ImageNet.
+Transformers, which are popular for language modeling, have been explored for solving vision tasks recently, e.g., the Vision Transformer (ViT) for image classification. The ViT model splits each image into a sequence of tokens with fixed length and then applies multiple Transformer layers to model their global relation for classification. However, ViT achieves inferior performance to CNNs when trained from scratch on a midsize dataset like ImageNet. We find it is because: 1) the simple tokenization of input images fails to model the important local structure such as edges and lines among neighboring pixels, leading to low training sample efficiency; 2) the redundant attention backbone design of ViT leads to limited feature richness for fixed computation budgets and limited training samples. To overcome such limitations, we propose a new Tokens-To-Token Vision Transformer (T2T-ViT), which incorporates 1) a layer-wise Tokens-to-Token (T2T) transformation to progressively structurize the image to tokens by recursively aggregating neighboring Tokens into one Token (Tokens-to-Token), such that local structure represented by surrounding tokens can be modeled and tokens length can be reduced; 2) an efficient backbone with a deep-narrow structure for vision transformer motivated by CNN architecture design after empirical study. Notably, T2T-ViT reduces the parameter count and MACs of vanilla ViT by half, while achieving more than 3.0% improvement when trained from scratch on ImageNet. It also outperforms ResNets and achieves comparable performance with MobileNets by directly training on ImageNet. For example, T2T-ViT with comparable size to ResNet50 (21.5M parameters) can achieve 83.3% top1 accuracy in image resolution 384×384 on ImageNet.
 
 <div align=center>
 <img src="https://user-images.githubusercontent.com/26739999/142578381-e9040610-05d9-457c-8bf5-01c2fa94add2.png" width="60%"/>

docker/serve/Dockerfile

@@ -4,7 +4,7 @@ ARG CUDNN="7"
 FROM pytorch/pytorch:${PYTORCH}-cuda${CUDA}-cudnn${CUDNN}-devel
 
 ARG MMCV="1.7.0"
-ARG MMCLS="0.24.1"
+ARG MMCLS="0.25.0"
 
 ENV PYTHONUNBUFFERED TRUE
 

docs/en/changelog.md

@@ -1,5 +1,33 @@
 # Changelog
 
+## v0.25.0(06/12/2022)
+
+### Highlights
+
+- Support MLU backend.
+
+### New Features
+
+- Support MLU backend. ([#1159](https://github.com/open-mmlab/mmclassification/pull/1159))
+- Support Activation Checkpointing for ConvNeXt. ([#1152](https://github.com/open-mmlab/mmclassification/pull/1152))
+
+### Improvements
+
+- Add `dist_train_arm.sh` for ARM device and update NPU results. ([#1218](https://github.com/open-mmlab/mmclassification/pull/1218))
+
+### Bug Fixes
+
+- Fix a bug caused `MMClsWandbHook` stuck. ([#1242](https://github.com/open-mmlab/mmclassification/pull/1242))
+- Fix the redundant `device_ids` in `tools/test.py`. ([#1215](https://github.com/open-mmlab/mmclassification/pull/1215))
+
+### Docs Update
+
+- Add version banner and version warning in master docs. ([#1216](https://github.com/open-mmlab/mmclassification/pull/1216))
+- Update NPU support doc. ([#1198](https://github.com/open-mmlab/mmclassification/pull/1198))
+- Fixed typo in `pytorch2torchscript.md`. ([#1173](https://github.com/open-mmlab/mmclassification/pull/1173))
+- Fix typo in `miscellaneous.md`. ([#1137](https://github.com/open-mmlab/mmclassification/pull/1137))
+- further detail for the doc for `ClassBalancedDataset`. ([#901](https://github.com/open-mmlab/mmclassification/pull/901))
+
 ## v0.24.1(31/10/2022)
 
 ### New Features
@@ -28,14 +56,14 @@
 
 ### Improvements
 
-- \[Improve\] replace loop of progressbar in api/test. ([#878](https://github.com/open-mmlab/mmclassification/pull/878))
+- [Improve] replace loop of progressbar in api/test. ([#878](https://github.com/open-mmlab/mmclassification/pull/878))
-- \[Enhance\] RepVGG for YOLOX-PAI. ([#1025](https://github.com/open-mmlab/mmclassification/pull/1025))
+- [Enhance] RepVGG for YOLOX-PAI. ([#1025](https://github.com/open-mmlab/mmclassification/pull/1025))
-- \[Enhancement\] Update VAN. ([#1017](https://github.com/open-mmlab/mmclassification/pull/1017))
+- [Enhancement] Update VAN. ([#1017](https://github.com/open-mmlab/mmclassification/pull/1017))
-- \[Refactor\] Re-write `get_sinusoid_encoding` from third-party implementation. ([#965](https://github.com/open-mmlab/mmclassification/pull/965))
+- [Refactor] Re-write `get_sinusoid_encoding` from third-party implementation. ([#965](https://github.com/open-mmlab/mmclassification/pull/965))
-- \[Improve\] Upgrade onnxsim to v0.4.0. ([#915](https://github.com/open-mmlab/mmclassification/pull/915))
+- [Improve] Upgrade onnxsim to v0.4.0. ([#915](https://github.com/open-mmlab/mmclassification/pull/915))
-- \[Improve\] Fixed typo in `RepVGG`. ([#985](https://github.com/open-mmlab/mmclassification/pull/985))
+- [Improve] Fixed typo in `RepVGG`. ([#985](https://github.com/open-mmlab/mmclassification/pull/985))
-- \[Improve\] Using `train_step` instead of `forward` in PreciseBNHook ([#964](https://github.com/open-mmlab/mmclassification/pull/964))
+- [Improve] Using `train_step` instead of `forward` in PreciseBNHook ([#964](https://github.com/open-mmlab/mmclassification/pull/964))
-- \[Improve\] Use `forward_dummy` to calculate FLOPS. ([#953](https://github.com/open-mmlab/mmclassification/pull/953))
+- [Improve] Use `forward_dummy` to calculate FLOPS. ([#953](https://github.com/open-mmlab/mmclassification/pull/953))
 
 ### Bug Fixes
 
@@ -102,13 +130,13 @@
 
 ### New Features
 
-- \[Feature\] Support resize relative position embedding in `SwinTransformer`. ([#749](https://github.com/open-mmlab/mmclassification/pull/749))
+- [Feature] Support resize relative position embedding in `SwinTransformer`. ([#749](https://github.com/open-mmlab/mmclassification/pull/749))
-- \[Feature\] Add PoolFormer backbone and checkpoints. ([#746](https://github.com/open-mmlab/mmclassification/pull/746))
+- [Feature] Add PoolFormer backbone and checkpoints. ([#746](https://github.com/open-mmlab/mmclassification/pull/746))
 
 ### Improvements
 
-- \[Enhance\] Improve CPE performance by reduce memory copy. ([#762](https://github.com/open-mmlab/mmclassification/pull/762))
+- [Enhance] Improve CPE performance by reduce memory copy. ([#762](https://github.com/open-mmlab/mmclassification/pull/762))
-- \[Enhance\] Add extra dataloader settings in configs. ([#752](https://github.com/open-mmlab/mmclassification/pull/752))
+- [Enhance] Add extra dataloader settings in configs. ([#752](https://github.com/open-mmlab/mmclassification/pull/752))
 
 ## v0.22.0(30/3/2022)
 
@@ -120,29 +148,29 @@
 
 ### New Features
 
-- \[Feature\] Add CSPNet and backbone and checkpoints ([#735](https://github.com/open-mmlab/mmclassification/pull/735))
+- [Feature] Add CSPNet and backbone and checkpoints ([#735](https://github.com/open-mmlab/mmclassification/pull/735))
-- \[Feature\] Add `CustomDataset`. ([#738](https://github.com/open-mmlab/mmclassification/pull/738))
+- [Feature] Add `CustomDataset`. ([#738](https://github.com/open-mmlab/mmclassification/pull/738))
-- \[Feature\] Add diff seeds to diff ranks.  ([#744](https://github.com/open-mmlab/mmclassification/pull/744))
+- [Feature] Add diff seeds to diff ranks.  ([#744](https://github.com/open-mmlab/mmclassification/pull/744))
-- \[Feature\] Support ConvMixer. ([#716](https://github.com/open-mmlab/mmclassification/pull/716))
+- [Feature] Support ConvMixer. ([#716](https://github.com/open-mmlab/mmclassification/pull/716))
-- \[Feature\] Our `dist_train` & `dist_test` tools support distributed training on multiple machines. ([#734](https://github.com/open-mmlab/mmclassification/pull/734))
+- [Feature] Our `dist_train` & `dist_test` tools support distributed training on multiple machines. ([#734](https://github.com/open-mmlab/mmclassification/pull/734))
-- \[Feature\] Add RepMLP backbone and checkpoints. ([#709](https://github.com/open-mmlab/mmclassification/pull/709))
+- [Feature] Add RepMLP backbone and checkpoints. ([#709](https://github.com/open-mmlab/mmclassification/pull/709))
-- \[Feature\] Support CUB dataset. ([#703](https://github.com/open-mmlab/mmclassification/pull/703))
+- [Feature] Support CUB dataset. ([#703](https://github.com/open-mmlab/mmclassification/pull/703))
-- \[Feature\] Support ResizeMix. ([#676](https://github.com/open-mmlab/mmclassification/pull/676))
+- [Feature] Support ResizeMix. ([#676](https://github.com/open-mmlab/mmclassification/pull/676))
 
 ### Improvements
 
-- \[Enhance\] Use `--a-b` instead of `--a_b` in arguments. ([#754](https://github.com/open-mmlab/mmclassification/pull/754))
+- [Enhance] Use `--a-b` instead of `--a_b` in arguments. ([#754](https://github.com/open-mmlab/mmclassification/pull/754))
-- \[Enhance\] Add `get_cat_ids` and `get_gt_labels` to KFoldDataset. ([#721](https://github.com/open-mmlab/mmclassification/pull/721))
+- [Enhance] Add `get_cat_ids` and `get_gt_labels` to KFoldDataset. ([#721](https://github.com/open-mmlab/mmclassification/pull/721))
-- \[Enhance\] Set torch seed in `worker_init_fn`. ([#733](https://github.com/open-mmlab/mmclassification/pull/733))
+- [Enhance] Set torch seed in `worker_init_fn`. ([#733](https://github.com/open-mmlab/mmclassification/pull/733))
 
 ### Bug Fixes
 
-- \[Fix\] Fix the discontiguous output feature map of ConvNeXt. ([#743](https://github.com/open-mmlab/mmclassification/pull/743))
+- [Fix] Fix the discontiguous output feature map of ConvNeXt. ([#743](https://github.com/open-mmlab/mmclassification/pull/743))
 
 ### Docs Update
 
-- \[Docs\] Add brief installation steps in README for copy&paste. ([#755](https://github.com/open-mmlab/mmclassification/pull/755))
+- [Docs] Add brief installation steps in README for copy&paste. ([#755](https://github.com/open-mmlab/mmclassification/pull/755))
-- \[Docs\] fix logo url link from mmocr to mmcls. ([#732](https://github.com/open-mmlab/mmclassification/pull/732))
+- [Docs] fix logo url link from mmocr to mmcls. ([#732](https://github.com/open-mmlab/mmclassification/pull/732))
 
 ## v0.21.0(04/03/2022)
 
@@ -245,18 +273,18 @@
 
 ### Improvements
 
-- \[Reproduction\] Reproduce RegNetX training accuracy. ([#587](https://github.com/open-mmlab/mmclassification/pull/587))
+- [Reproduction] Reproduce RegNetX training accuracy. ([#587](https://github.com/open-mmlab/mmclassification/pull/587))
-- \[Reproduction\] Reproduce training results of T2T-ViT. ([#610](https://github.com/open-mmlab/mmclassification/pull/610))
+- [Reproduction] Reproduce training results of T2T-ViT. ([#610](https://github.com/open-mmlab/mmclassification/pull/610))
-- \[Enhance\] Provide high-acc training settings of ResNet. ([#572](https://github.com/open-mmlab/mmclassification/pull/572))
+- [Enhance] Provide high-acc training settings of ResNet. ([#572](https://github.com/open-mmlab/mmclassification/pull/572))
-- \[Enhance\] Set a random seed when the user does not set a seed. ([#554](https://github.com/open-mmlab/mmclassification/pull/554))
+- [Enhance] Set a random seed when the user does not set a seed. ([#554](https://github.com/open-mmlab/mmclassification/pull/554))
-- \[Enhance\] Added `NumClassCheckHook` and unit tests. ([#559](https://github.com/open-mmlab/mmclassification/pull/559))
+- [Enhance] Added `NumClassCheckHook` and unit tests. ([#559](https://github.com/open-mmlab/mmclassification/pull/559))
-- \[Enhance\] Enhance feature extraction function. ([#593](https://github.com/open-mmlab/mmclassification/pull/593))
+- [Enhance] Enhance feature extraction function. ([#593](https://github.com/open-mmlab/mmclassification/pull/593))
-- \[Enhance\] Improve efficiency of precision, recall, f1_score and support. ([#595](https://github.com/open-mmlab/mmclassification/pull/595))
+- [Enhance] Improve efficiency of precision, recall, f1_score and support. ([#595](https://github.com/open-mmlab/mmclassification/pull/595))
-- \[Enhance\] Improve accuracy calculation performance. ([#592](https://github.com/open-mmlab/mmclassification/pull/592))
+- [Enhance] Improve accuracy calculation performance. ([#592](https://github.com/open-mmlab/mmclassification/pull/592))
-- \[Refactor\] Refactor `analysis_log.py`. ([#529](https://github.com/open-mmlab/mmclassification/pull/529))
+- [Refactor] Refactor `analysis_log.py`. ([#529](https://github.com/open-mmlab/mmclassification/pull/529))
-- \[Refactor\] Use new API of matplotlib to handle blocking input in visualization. ([#568](https://github.com/open-mmlab/mmclassification/pull/568))
+- [Refactor] Use new API of matplotlib to handle blocking input in visualization. ([#568](https://github.com/open-mmlab/mmclassification/pull/568))
-- \[CI\] Cancel previous runs that are not completed. ([#583](https://github.com/open-mmlab/mmclassification/pull/583))
+- [CI] Cancel previous runs that are not completed. ([#583](https://github.com/open-mmlab/mmclassification/pull/583))
-- \[CI\] Skip build CI if only configs or docs modification. ([#575](https://github.com/open-mmlab/mmclassification/pull/575))
+- [CI] Skip build CI if only configs or docs modification. ([#575](https://github.com/open-mmlab/mmclassification/pull/575))
 
 ### Bug Fixes
 

docs/en/faq.md

@@ -18,7 +18,8 @@ and make sure you fill in all required information in the template.
   | MMClassification version |      MMCV version      |
   | :----------------------: | :--------------------: |
   |           dev            |  mmcv>=1.7.0, \<1.9.0  |
-  |     0.24.1 (master)      |  mmcv>=1.4.2, \<1.9.0  |
+  |     0.25.0 (master)      |  mmcv>=1.4.2, \<1.9.0  |
+  |          0.24.1          |  mmcv>=1.4.2, \<1.9.0  |
   |          0.23.2          |  mmcv>=1.4.2, \<1.7.0  |
   |          0.22.1          |  mmcv>=1.4.2, \<1.6.0  |
   |          0.21.0          | mmcv>=1.4.2, \<=1.5.0  |

docs/zh_CN/faq.md

@@ -16,7 +16,8 @@
   | MMClassification version |      MMCV version      |
   | :----------------------: | :--------------------: |
   |           dev            |  mmcv>=1.7.0, \<1.9.0  |
-  |     0.24.1 (master)      |  mmcv>=1.4.2, \<1.9.0  |
+  |     0.25.0 (master)      |  mmcv>=1.4.2, \<1.9.0  |
+  |          0.24.1          |  mmcv>=1.4.2, \<1.9.0  |
   |          0.23.2          |  mmcv>=1.4.2, \<1.7.0  |
   |          0.22.1          |  mmcv>=1.4.2, \<1.6.0  |
   |          0.21.0          | mmcv>=1.4.2, \<=1.5.0  |

mmcls/apis/train.py

@@ -10,11 +10,11 @@ from mmcv.runner import (DistSamplerSeedHook, Fp16OptimizerHook,
 
 from mmcls.core import DistEvalHook, DistOptimizerHook, EvalHook
 from mmcls.datasets import build_dataloader, build_dataset
-from mmcls.utils import (get_root_logger, wrap_distributed_model,
+from mmcls.utils import (auto_select_device, get_root_logger,
-                         wrap_non_distributed_model)
+                         wrap_distributed_model, wrap_non_distributed_model)
 
 
-def init_random_seed(seed=None, device='cuda'):
+def init_random_seed(seed=None, device=None):
     """Initialize random seed.
 
     If the seed is not set, the seed will be automatically randomized,
@@ -30,7 +30,8 @@ def init_random_seed(seed=None, device='cuda'):
     """
     if seed is not None:
         return seed
-
+    if device is None:
+        device = auto_select_device()
     # Make sure all ranks share the same random seed to prevent
     # some potential bugs. Please refer to
     # https://github.com/open-mmlab/mmdetection/issues/6339

mmcls/core/hook/wandblogger_hook.py

@@ -3,7 +3,7 @@ import os.path as osp
 
 import numpy as np
 from mmcv.runner import HOOKS, BaseRunner
-from mmcv.runner.dist_utils import master_only
+from mmcv.runner.dist_utils import get_dist_info, master_only
 from mmcv.runner.hooks.checkpoint import CheckpointHook
 from mmcv.runner.hooks.evaluation import DistEvalHook, EvalHook
 from mmcv.runner.hooks.logger.wandb import WandbLoggerHook
@@ -190,7 +190,6 @@ class MMClsWandbHook(WandbLoggerHook):
             # Log the evaluation table
             self._log_eval_table(runner.epoch + 1)
 
-    @master_only
     def after_train_iter(self, runner):
         if self.get_mode(runner) == 'train':
             # An ugly patch. The iter-based eval hook will call the
@@ -201,6 +200,10 @@ class MMClsWandbHook(WandbLoggerHook):
         else:
             super(MMClsWandbHook, self).after_train_iter(runner)
 
+        rank, _ = get_dist_info()
+        if rank != 0:
+            return
+
         if self.by_epoch:
             return
 

mmcls/core/utils/dist_utils.py

@@ -8,6 +8,8 @@ from mmcv.runner import OptimizerHook, get_dist_info
 from torch._utils import (_flatten_dense_tensors, _take_tensors,
                           _unflatten_dense_tensors)
 
+from mmcls.utils import auto_select_device
+
 
 def _allreduce_coalesced(tensors, world_size, bucket_size_mb=-1):
     if bucket_size_mb > 0:
@@ -59,7 +61,7 @@ class DistOptimizerHook(OptimizerHook):
         runner.optimizer.step()
 
 
-def sync_random_seed(seed=None, device='cuda'):
+def sync_random_seed(seed=None, device=None):
     """Make sure different ranks share the same seed.
 
     All workers must call this function, otherwise it will deadlock.
@@ -81,6 +83,8 @@ def sync_random_seed(seed=None, device='cuda'):
     Returns:
         int: Seed to be used.
     """
+    if device is None:
+        device = auto_select_device()
     if seed is None:
         seed = np.random.randint(2**31)
     assert isinstance(seed, int)

mmcls/datasets/samplers/distributed_sampler.py

@@ -4,7 +4,6 @@ from torch.utils.data import DistributedSampler as _DistributedSampler
 
 from mmcls.core.utils import sync_random_seed
 from mmcls.datasets import SAMPLERS
-from mmcls.utils import auto_select_device
 
 
 @SAMPLERS.register_module()
@@ -31,7 +30,7 @@ class DistributedSampler(_DistributedSampler):
         # in the same order based on the same seed. Then different ranks
         # could use different indices to select non-overlapped data from the
         # same data list.
-        self.seed = sync_random_seed(seed, device=auto_select_device())
+        self.seed = sync_random_seed(seed)
 
     def __iter__(self):
         # deterministically shuffle based on epoch

mmcls/models/backbones/convnext.py

@@ -6,6 +6,7 @@ from typing import Sequence
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
+import torch.utils.checkpoint as cp
 from mmcv.cnn.bricks import (NORM_LAYERS, DropPath, build_activation_layer,
                              build_norm_layer)
 from mmcv.runner import BaseModule
@@ -77,8 +78,11 @@ class ConvNeXtBlock(BaseModule):
                  mlp_ratio=4.,
                  linear_pw_conv=True,
                  drop_path_rate=0.,
-                 layer_scale_init_value=1e-6):
+                 layer_scale_init_value=1e-6,
+                 with_cp=False):
         super().__init__()
+        self.with_cp = with_cp
+
         self.depthwise_conv = nn.Conv2d(
             in_channels,
             in_channels,
@@ -108,24 +112,33 @@ class ConvNeXtBlock(BaseModule):
             drop_path_rate) if drop_path_rate > 0. else nn.Identity()
 
     def forward(self, x):
-        shortcut = x
-        x = self.depthwise_conv(x)
-        x = self.norm(x)
 
-        if self.linear_pw_conv:
+        def _inner_forward(x):
-            x = x.permute(0, 2, 3, 1)  # (N, C, H, W) -> (N, H, W, C)
+            shortcut = x
+            x = self.depthwise_conv(x)
+            x = self.norm(x)
 
-        x = self.pointwise_conv1(x)
+            if self.linear_pw_conv:
-        x = self.act(x)
+                x = x.permute(0, 2, 3, 1)  # (N, C, H, W) -> (N, H, W, C)
-        x = self.pointwise_conv2(x)
 
-        if self.linear_pw_conv:
+            x = self.pointwise_conv1(x)
-            x = x.permute(0, 3, 1, 2)  # permute back
+            x = self.act(x)
+            x = self.pointwise_conv2(x)
 
-        if self.gamma is not None:
+            if self.linear_pw_conv:
-            x = x.mul(self.gamma.view(1, -1, 1, 1))
+                x = x.permute(0, 3, 1, 2)  # permute back
+
+            if self.gamma is not None:
+                x = x.mul(self.gamma.view(1, -1, 1, 1))
+
+            x = shortcut + self.drop_path(x)
+            return x
+
+        if self.with_cp and x.requires_grad:
+            x = cp.checkpoint(_inner_forward, x)
+        else:
+            x = _inner_forward(x)
 
-        x = shortcut + self.drop_path(x)
         return x
 
 
@@ -169,6 +182,8 @@ class ConvNeXt(BaseBackbone):
         gap_before_final_norm (bool): Whether to globally average the feature
             map before the final norm layer. In the official repo, it's only
             used in classification task. Defaults to True.
+        with_cp (bool): Use checkpoint or not. Using checkpoint will save some
+            memory while slowing down the training speed. Defaults to False.
         init_cfg (dict, optional): Initialization config dict
     """  # noqa: E501
     arch_settings = {
@@ -206,6 +221,7 @@ class ConvNeXt(BaseBackbone):
                  out_indices=-1,
                  frozen_stages=0,
                  gap_before_final_norm=True,
+                 with_cp=False,
                  init_cfg=None):
         super().__init__(init_cfg=init_cfg)
 
@@ -288,8 +304,8 @@ class ConvNeXt(BaseBackbone):
                     norm_cfg=norm_cfg,
                     act_cfg=act_cfg,
                     linear_pw_conv=linear_pw_conv,
-                    layer_scale_init_value=layer_scale_init_value)
+                    layer_scale_init_value=layer_scale_init_value,
-                for j in range(depth)
+                    with_cp=with_cp) for j in range(depth)
             ])
             block_idx += depth
 

mmcls/utils/distribution.py

@@ -19,6 +19,9 @@ def wrap_non_distributed_model(model, device='cuda', dim=0, *args, **kwargs):
     if device == 'npu':
         from mmcv.device.npu import NPUDataParallel
         model = NPUDataParallel(model.npu(), dim=dim, *args, **kwargs)
+    elif device == 'mlu':
+        from mmcv.device.mlu import MLUDataParallel
+        model = MLUDataParallel(model.mlu(), dim=dim, *args, **kwargs)
     elif device == 'cuda':
         from mmcv.parallel import MMDataParallel
         model = MMDataParallel(model.cuda(), dim=dim, *args, **kwargs)
@@ -57,6 +60,15 @@ def wrap_distributed_model(model, device='cuda', *args, **kwargs):
         from torch.npu import current_device
         model = NPUDistributedDataParallel(
             model.npu(), *args, device_ids=[current_device()], **kwargs)
+    elif device == 'mlu':
+        import os
+
+        from mmcv.device.mlu import MLUDistributedDataParallel
+        model = MLUDistributedDataParallel(
+            model.mlu(),
+            *args,
+            device_ids=[int(os.environ['LOCAL_RANK'])],
+            **kwargs)
     elif device == 'cuda':
         from mmcv.parallel import MMDistributedDataParallel
         from torch.cuda import current_device

mmcls/version.py

@@ -1,6 +1,6 @@
 # Copyright (c) OpenMMLab. All rights reserved
 
-__version__ = '0.24.1'
+__version__ = '0.25.0'
 
 
 def parse_version_info(version_str):

tests/test_models/test_backbones/test_convnext.py

@@ -84,3 +84,13 @@ def test_convnext():
     for i in range(2, 4):
         assert model.downsample_layers[i].training
         assert model.stages[i].training
+
+    # Test Activation Checkpointing
+    model = ConvNeXt(arch='tiny', out_indices=-1, with_cp=True)
+    model.init_weights()
+    model.train()
+
+    imgs = torch.randn(1, 3, 224, 224)
+    feat = model(imgs)
+    assert len(feat) == 1
+    assert feat[0].shape == torch.Size([1, 768])

tools/test.py

@@ -195,10 +195,7 @@ def main():
                                   **show_kwargs)
     else:
         model = wrap_distributed_model(
-            model,
+            model, device=cfg.device, broadcast_buffers=False)
-            device=cfg.device,
-            device_ids=[int(os.environ['LOCAL_RANK'])],
-            broadcast_buffers=False)
         outputs = multi_gpu_test(model, data_loader, args.tmpdir,
                                  args.gpu_collect)