Merge branch 'dev'

Author: mzr1996

.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))
-- \[Enhance\] 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))
-- \[Refactor\] 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))
-- \[Improve\] 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))
-- \[Improve\] Use `forward_dummy` to calculate FLOPS. ([#953](https://github.com/open-mmlab/mmclassification/pull/953))
+- [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))
+- [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))
+- [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] 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))
 
 ### 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))
-- \[Feature\] Add PoolFormer backbone and checkpoints. ([#746](https://github.com/open-mmlab/mmclassification/pull/746))
+- [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))
 
 ### Improvements
 
-- \[Enhance\] 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] 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))
 
 ## 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 `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))
-- \[Feature\] 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))
-- \[Feature\] 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 ResizeMix. ([#676](https://github.com/open-mmlab/mmclassification/pull/676))
+- [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 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] 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] Support CUB dataset. ([#703](https://github.com/open-mmlab/mmclassification/pull/703))
+- [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))
-- \[Enhance\] 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] 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] 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))
-- \[Docs\] fix logo url link from mmocr to mmcls. ([#732](https://github.com/open-mmlab/mmclassification/pull/732))
+- [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))
 
 ## 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 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))
-- \[Enhance\] 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))
-- \[Enhance\] 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 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))
-- \[Refactor\] 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))
-- \[CI\] Skip build CI if only configs or docs modification. ([#575](https://github.com/open-mmlab/mmclassification/pull/575))
+- [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))
+- [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] 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] 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))
+- [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))
+- [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))
 
 ### 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 mmcls.core import DistEvalHook, DistOptimizerHook, EvalHook
 from mmcls.datasets import build_dataloader, build_dataset
-from mmcls.utils import (get_root_logger, wrap_distributed_model,
-                         wrap_non_distributed_model)
+from mmcls.utils import (auto_select_device, get_root_logger,
+                         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 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 @@ def after_train_epoch(self, runner):
             # 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 @@ def after_train_iter(self, runner):
         else:
             super(MMClsWandbHook, self).after_train_iter(runner)
 
+        rank, _ = get_dist_info()
+        if rank != 0:
+            return
+
         if self.by_epoch:
             return