first commit

dff2c686 · renzhc · 8f9dd0ed · dff2c686 · dff2c686 · dff2c686
Commit dff2c686 authored Sep 03, 2024 by renzhc
20 changed files
--- a/configs/repmlp/repmlp-base_delopy_8xb64_in1k.py
+++ b/configs/repmlp/repmlp-base_delopy_8xb64_in1k.py
+_base_ = ['./repmlp-base_8xb64_in1k.py']
+
+model = dict(backbone=dict(deploy=True))
--- a/configs/repmlp/repmlp-base_deploy_8xb64_in1k-256px.py
+++ b/configs/repmlp/repmlp-base_deploy_8xb64_in1k-256px.py
+_base_ = ['./repmlp-base_8xb64_in1k-256px.py']
+
+model = dict(backbone=dict(deploy=True))
--- a/configs/repvgg/README.md
+++ b/configs/repvgg/README.md
+# RepVGG
+
+> [RepVGG: Making VGG-style ConvNets Great Again](https://arxiv.org/abs/2101.03697)
+
+<!-- [ALGORITHM] -->
+
+## Introduction
+
+RepVGG is a VGG-style convolutional architecture. It has the following advantages:
+
+1. The model has a VGG-like plain (a.k.a. feed-forward) topology 1 without any branches. I.e., every layer takes the output of its only preceding layer as input and feeds the output into its only following layer.
+2. The model’s body uses only 3 × 3 conv and ReLU.
+3. The concrete architecture (including the specific depth and layer widths) is instantiated with no automatic search, manual refinement, compound scaling, nor other heavy designs.
+
+<div align=center>
+<img src="https://user-images.githubusercontent.com/26739999/142573223-f7f14d32-ea08-43a1-81ad-5a6a83ee0122.png" width="60%"/>
+</div>
+
+## Abstract
+
+<details>
+
+<summary>Show the paper's abstract</summary>
+
+<br>
+We present a simple but powerful architecture of convolutional neural network, which has a VGG-like inference-time body composed of nothing but a stack of 3x3 convolution and ReLU, while the training-time model has a multi-branch topology. Such decoupling of the training-time and inference-time architecture is realized by a structural re-parameterization technique so that the model is named RepVGG. On ImageNet, RepVGG reaches over 80% top-1 accuracy, which is the first time for a plain model, to the best of our knowledge. On NVIDIA 1080Ti GPU, RepVGG models run 83% faster than ResNet-50 or 101% faster than ResNet-101 with higher accuracy and show favorable accuracy-speed trade-off compared to the state-of-the-art models like EfficientNet and RegNet.
+</br>
+
+</details>
+
+## How to use it?
+
+<!-- [TABS-BEGIN] -->
+
+**Predict image**
+
+```python
+from mmpretrain import inference_model, get_model
+
+model = get_model('repvgg-A0_8xb32_in1k', pretrained=True)
+model.backbone.switch_to_deploy()
+predict = inference_model(model, 'demo/bird.JPEG')
+print(predict['pred_class'])
+print(predict['pred_score'])
+```
+
+**Use the model**
+
+```python
+import torch
+from mmpretrain import get_model
+
+model = get_model('repvgg-A0_8xb32_in1k', pretrained=True)
+inputs = torch.rand(1, 3, 224, 224)
+out = model(inputs)
+print(type(out))
+# To extract features.
+feats = model.extract_feat(inputs)
+print(type(feats))
+```
+
+**Train/Test Command**
+
+Prepare your dataset according to the [docs](https://mmpretrain.readthedocs.io/en/latest/user_guides/dataset_prepare.html#prepare-dataset).
+
+Train:
+
+```shell
+python tools/train.py configs/repvgg/repvgg-A0_8xb32_in1k.py
+```
+
+Test:
+
+```shell
+python tools/test.py configs/repvgg/repvgg-A0_8xb32_in1k.py https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A0_8xb32_in1k_20221213-60ae8e23.pth
+```
+
+Test with reparameterized model:
+
+```shell
+python tools/test.py configs/repvgg/repvgg-A0_8xb32_in1k.py repvgg_A0_deploy.pth --cfg-options model.backbone.deploy=True
+```
+
+**Reparameterization**
+
+The checkpoints provided are all `training-time` models. Use the reparameterize tool to switch them to more efficient `inference-time` architecture, which not only has fewer parameters but also less calculations.
+
+```bash
+python tools/convert_models/reparameterize_model.py ${CFG_PATH} ${SRC_CKPT_PATH} ${TARGET_CKPT_PATH}
+```
+
+`${CFG_PATH}` is the config file, `${SRC_CKPT_PATH}` is the source chenpoint file, `${TARGET_CKPT_PATH}` is the target deploy weight file path.
+
+To use reparameterized weights, the config file must switch to the deploy config files.
+
+```bash
+python tools/test.py ${deploy_cfg} ${deploy_checkpoint} --metrics accuracy
+```
+
+You can also use `backbone.switch_to_deploy()` to switch to the deploy mode in Python code. For example:
+
+```python
+from mmpretrain.models import RepVGG
+
+backbone = RepVGG(arch='A0')
+backbone.switch_to_deploy()
+```
+
+<!-- [TABS-END] -->
+
+## Models and results
+
+### Image Classification on ImageNet-1k
+
+| Model                         |   Pretrain   | Params (M) | Flops (G) | Top-1 (%) | Top-5 (%) |               Config                |                                        Download                                         |
+| :---------------------------- | :----------: | :--------: | :-------: | :-------: | :-------: | :---------------------------------: | :-------------------------------------------------------------------------------------: |
+| `repvgg-A0_8xb32_in1k`        | From scratch |    8.31    |   1.36    |   72.37   |   90.56   |  [config](repvgg-A0_8xb32_in1k.py)  | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A0_8xb32_in1k_20221213-60ae8e23.pth) \| [log](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A0_8xb32_in1k_20221213-60ae8e23.log) |
+| `repvgg-A1_8xb32_in1k`        | From scratch |   12.79    |   2.36    |   74.23   |   91.80   |  [config](repvgg-A1_8xb32_in1k.py)  | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A1_8xb32_in1k_20221213-f81bf3df.pth) \| [log](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A1_8xb32_in1k_20221213-f81bf3df.log) |
+| `repvgg-A2_8xb32_in1k`        | From scratch |   25.50    |   5.12    |   76.49   |   93.09   |  [config](repvgg-A2_8xb32_in1k.py)  | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A2_8xb32_in1k_20221213-a8767caf.pth) \| [log](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A2_8xb32_in1k_20221213-a8767caf.log) |
+| `repvgg-B0_8xb32_in1k`        | From scratch |    3.42    |   15.82   |   75.27   |   92.21   |  [config](repvgg-B0_8xb32_in1k.py)  | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B0_8xb32_in1k_20221213-5091ecc7.pth) \| [log](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B0_8xb32_in1k_20221213-5091ecc7.log) |
+| `repvgg-B1_8xb32_in1k`        | From scratch |   51.83    |   11.81   |   78.19   |   94.04   |  [config](repvgg-B1_8xb32_in1k.py)  | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B1_8xb32_in1k_20221213-d17c45e7.pth) \| [log](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B1_8xb32_in1k_20221213-d17c45e7.log) |
+| `repvgg-B1g2_8xb32_in1k`      | From scratch |   41.36    |   8.81    |   77.87   |   93.99   | [config](repvgg-B1g2_8xb32_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B1g2_8xb32_in1k_20221213-ae6428fd.pth) \| [log](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B1g2_8xb32_in1k_20221213-ae6428fd.log) |
+| `repvgg-B1g4_8xb32_in1k`      | From scratch |   36.13    |   7.30    |   77.81   |   93.77   | [config](repvgg-B1g4_8xb32_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B1g4_8xb32_in1k_20221213-a7a4aaea.pth) \| [log](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B1g4_8xb32_in1k_20221213-a7a4aaea.log) |
+| `repvgg-B2_8xb32_in1k`        | From scratch |   80.32    |   18.37   |   78.58   |   94.23   |  [config](repvgg-B2_8xb32_in1k.py)  | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B2_8xb32_in1k_20221213-d8b420ef.pth) \| [log](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B2_8xb32_in1k_20221213-d8b420ef.log) |
+| `repvgg-B2g4_8xb32_in1k`      | From scratch |   55.78    |   11.33   |   79.44   |   94.72   | [config](repvgg-B2g4_8xb32_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B2g4_8xb32_in1k_20221213-0c1990eb.pth) \| [log](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B2g4_8xb32_in1k_20221213-0c1990eb.log) |
+| `repvgg-B3_8xb32_in1k`        | From scratch |   110.96   |   26.21   |   80.58   |   95.33   |  [config](repvgg-B3_8xb32_in1k.py)  | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B3_8xb32_in1k_20221213-927a329a.pth) \| [log](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B3_8xb32_in1k_20221213-927a329a.log) |
+| `repvgg-B3g4_8xb32_in1k`      | From scratch |   75.63    |   16.06   |   80.26   |   95.15   | [config](repvgg-B3g4_8xb32_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B3g4_8xb32_in1k_20221213-e01cb280.pth) \| [log](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B3g4_8xb32_in1k_20221213-e01cb280.log) |
+| `repvgg-D2se_3rdparty_in1k`\* | From scratch |   120.39   |   32.84   |   81.81   |   95.94   | [config](repvgg-D2se_8xb32_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-D2se_3rdparty_4xb64-autoaug-lbs-mixup-coslr-200e_in1k_20210909-cf3139b7.pth) |
+
+*Models with * are converted from the [official repo](https://github.com/DingXiaoH/RepVGG/blob/9f272318abfc47a2b702cd0e916fca8d25d683e7/repvgg.py#L250). The config files of these models are only for inference. We haven't reproduce the training results.*
+
+## Citation
+
+```bibtex
+@inproceedings{ding2021repvgg,
+  title={Repvgg: Making vgg-style convnets great again},
+  author={Ding, Xiaohan and Zhang, Xiangyu and Ma, Ningning and Han, Jungong and Ding, Guiguang and Sun, Jian},
+  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
+  pages={13733--13742},
+  year={2021}
+}
+```
--- a/configs/repvgg/metafile.yml
+++ b/configs/repvgg/metafile.yml
+Collections:
+  - Name: RepVGG
+    Metadata:
+      Training Data: ImageNet-1k
+      Architecture:
+        - re-parameterization Convolution
+        - VGG-style Neural Network
+    Paper:
+      URL: https://arxiv.org/abs/2101.03697
+      Title: 'RepVGG: Making VGG-style ConvNets Great Again'
+    README: configs/repvgg/README.md
+    Code:
+      URL: https://github.com/open-mmlab/mmpretrain/blob/v0.16.0/mmcls/models/backbones/repvgg.py#L257
+      Version: v0.16.0
+
+Models:
+  - Name: repvgg-A0_8xb32_in1k
+    In Collection: RepVGG
+    Config: configs/repvgg/repvgg-A0_8xb32_in1k.py
+    Metadata:
+      FLOPs: 1360233728
+      Parameters: 8309384
+    Results:
+    - Dataset: ImageNet-1k
+      Task: Image Classification
+      Metrics:
+        Top 1 Accuracy: 72.37
+        Top 5 Accuracy: 90.56
+    Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A0_8xb32_in1k_20221213-60ae8e23.pth
+  - Name: repvgg-A1_8xb32_in1k
+    In Collection: RepVGG
+    Config: configs/repvgg/repvgg-A1_8xb32_in1k.py
+    Metadata:
+      FLOPs: 2362750208
+      Parameters: 12789864
+    Results:
+    - Dataset: ImageNet-1k
+      Task: Image Classification
+      Metrics:
+        Top 1 Accuracy: 74.23
+        Top 5 Accuracy: 91.80
+    Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A1_8xb32_in1k_20221213-f81bf3df.pth
+  - Name: repvgg-A2_8xb32_in1k
+    In Collection: RepVGG
+    Config: configs/repvgg/repvgg-A2_8xb32_in1k.py
+    Metadata:
+      FLOPs: 5115612544
+      Parameters: 25499944
+    Results:
+    - Dataset: ImageNet-1k
+      Task: Image Classification
+      Metrics:
+        Top 1 Accuracy: 76.49
+        Top 5 Accuracy: 93.09
+    Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A2_8xb32_in1k_20221213-a8767caf.pth
+  - Name: repvgg-B0_8xb32_in1k
+    In Collection: RepVGG
+    Config: configs/repvgg/repvgg-B0_8xb32_in1k.py
+    Metadata:
+      FLOPs: 15820000000
+      Parameters: 3420000
+    Results:
+    - Dataset: ImageNet-1k
+      Task: Image Classification
+      Metrics:
+        Top 1 Accuracy: 75.27
+        Top 5 Accuracy: 92.21
+    Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B0_8xb32_in1k_20221213-5091ecc7.pth
+  - Name: repvgg-B1_8xb32_in1k
+    In Collection: RepVGG
+    Config: configs/repvgg/repvgg-B1_8xb32_in1k.py
+    Metadata:
+      FLOPs: 11813537792
+      Parameters: 51829480
+    Results:
+    - Dataset: ImageNet-1k
+      Task: Image Classification
+      Metrics:
+        Top 1 Accuracy: 78.19
+        Top 5 Accuracy: 94.04
+    Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B1_8xb32_in1k_20221213-d17c45e7.pth
+  - Name: repvgg-B1g2_8xb32_in1k
+    In Collection: RepVGG
+    Config: configs/repvgg/repvgg-B1g2_8xb32_in1k.py
+    Metadata:
+      FLOPs: 8807794688
+      Parameters: 41360104
+    Results:
+    - Dataset: ImageNet-1k
+      Task: Image Classification
+      Metrics:
+        Top 1 Accuracy: 77.87
+        Top 5 Accuracy: 93.99
+    Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B1g2_8xb32_in1k_20221213-ae6428fd.pth
+  - Name: repvgg-B1g4_8xb32_in1k
+    In Collection: RepVGG
+    Config: configs/repvgg/repvgg-B1g4_8xb32_in1k.py
+    Metadata:
+      FLOPs: 7304923136
+      Parameters: 36125416
+    Results:
+    - Dataset: ImageNet-1k
+      Task: Image Classification
+      Metrics:
+        Top 1 Accuracy: 77.81
+        Top 5 Accuracy: 93.77
+    Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B1g4_8xb32_in1k_20221213-a7a4aaea.pth
+  - Name: repvgg-B2_8xb32_in1k
+    In Collection: RepVGG
+    Config: configs/repvgg/repvgg-B2_8xb32_in1k.py
+    Metadata:
+      FLOPs: 18374175232
+      Parameters: 80315112
+    Results:
+    - Dataset: ImageNet-1k
+      Task: Image Classification
+      Metrics:
+        Top 1 Accuracy: 78.58
+        Top 5 Accuracy: 94.23
+    Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B2_8xb32_in1k_20221213-d8b420ef.pth
+  - Name: repvgg-B2g4_8xb32_in1k
+    In Collection: RepVGG
+    Config: configs/repvgg/repvgg-B2g4_8xb32_in1k.py
+    Metadata:
+      FLOPs: 11329464832
+      Parameters: 55777512
+    Results:
+    - Dataset: ImageNet-1k
+      Task: Image Classification
+      Metrics:
+        Top 1 Accuracy: 79.44
+        Top 5 Accuracy: 94.72
+    Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B2g4_8xb32_in1k_20221213-0c1990eb.pth
+  - Name: repvgg-B3_8xb32_in1k
+    In Collection: RepVGG
+    Config: configs/repvgg/repvgg-B3_8xb32_in1k.py
+    Metadata:
+      FLOPs: 26206448128
+      Parameters: 110960872
+    Results:
+    - Dataset: ImageNet-1k
+      Task: Image Classification
+      Metrics:
+        Top 1 Accuracy: 80.58
+        Top 5 Accuracy: 95.33
+    Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B3_8xb32_in1k_20221213-927a329a.pth
+  - Name: repvgg-B3g4_8xb32_in1k
+    In Collection: RepVGG
+    Config: configs/repvgg/repvgg-B3g4_8xb32_in1k.py
+    Metadata:
+      FLOPs: 16062065152
+      Parameters: 75626728
+    Results:
+    - Dataset: ImageNet-1k
+      Task: Image Classification
+      Metrics:
+        Top 1 Accuracy: 80.26
+        Top 5 Accuracy: 95.15
+    Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B3g4_8xb32_in1k_20221213-e01cb280.pth
+  - Name: repvgg-D2se_3rdparty_in1k
+    In Collection: RepVGG
+    Config: configs/repvgg/repvgg-D2se_8xb32_in1k.py
+    Metadata:
+      FLOPs: 32838581760
+      Parameters: 120387572
+    Results:
+    - Dataset: ImageNet-1k
+      Task: Image Classification
+      Metrics:
+        Top 1 Accuracy: 81.81
+        Top 5 Accuracy: 95.94
+    Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-D2se_3rdparty_4xb64-autoaug-lbs-mixup-coslr-200e_in1k_20210909-cf3139b7.pth
+    Converted From:
+      Weights: https://drive.google.com/drive/folders/1Avome4KvNp0Lqh2QwhXO6L5URQjzCjUq
+      Code: https://github.com/DingXiaoH/RepVGG/blob/9f272318abfc47a2b702cd0e916fca8d25d683e7/repvgg.py#L250
--- a/configs/repvgg/repvgg-A0_8xb32_in1k.py
+++ b/configs/repvgg/repvgg-A0_8xb32_in1k.py
+_base_ = [
+    '../_base_/models/repvgg-A0_in1k.py',
+    '../_base_/datasets/imagenet_bs32_pil_resize.py',
+    '../_base_/schedules/imagenet_bs256_coslr.py',
+    '../_base_/default_runtime.py'
+]
+
+val_dataloader = dict(batch_size=256)
+test_dataloader = dict(batch_size=256)
+
+# schedule settings
+optim_wrapper = dict(
+    paramwise_cfg=dict(
+        bias_decay_mult=0.0,
+        custom_keys={
+            'branch_3x3.norm': dict(decay_mult=0.0),
+            'branch_1x1.norm': dict(decay_mult=0.0),
+            'branch_norm.bias': dict(decay_mult=0.0),
+        }))
+
+# schedule settings
+param_scheduler = dict(
+    type='CosineAnnealingLR',
+    T_max=120,
+    by_epoch=True,
+    begin=0,
+    end=120,
+    convert_to_iter_based=True)
+
+train_cfg = dict(by_epoch=True, max_epochs=120)
+
+default_hooks = dict(
+    checkpoint=dict(type='CheckpointHook', interval=1, max_keep_ckpts=3))
--- a/configs/repvgg/repvgg-A0_deploy_in1k.py
+++ b/configs/repvgg/repvgg-A0_deploy_in1k.py
+_base_ = './repvgg-A0_8xb32_in1k.py'
+
+model = dict(backbone=dict(deploy=True))
--- a/configs/repvgg/repvgg-A1_8xb32_in1k.py
+++ b/configs/repvgg/repvgg-A1_8xb32_in1k.py
+_base_ = './repvgg-A0_8xb32_in1k.py'
+
+model = dict(backbone=dict(arch='A1'))
--- a/configs/repvgg/repvgg-A2_8xb32_in1k.py
+++ b/configs/repvgg/repvgg-A2_8xb32_in1k.py
+_base_ = './repvgg-A0_8xb32_in1k.py'
+
+model = dict(backbone=dict(arch='A2'), head=dict(in_channels=1408))
--- a/configs/repvgg/repvgg-B0_8xb32_in1k.py
+++ b/configs/repvgg/repvgg-B0_8xb32_in1k.py
+_base_ = './repvgg-A0_8xb32_in1k.py'
+
+model = dict(backbone=dict(arch='B0'), head=dict(in_channels=1280))
--- a/configs/repvgg/repvgg-B1_8xb32_in1k.py
+++ b/configs/repvgg/repvgg-B1_8xb32_in1k.py
+_base_ = './repvgg-A0_8xb32_in1k.py'
+
+model = dict(backbone=dict(arch='B1'), head=dict(in_channels=2048))
--- a/configs/repvgg/repvgg-B1g2_8xb32_in1k.py
+++ b/configs/repvgg/repvgg-B1g2_8xb32_in1k.py
+_base_ = './repvgg-A0_8xb32_in1k.py'
+
+model = dict(backbone=dict(arch='B1g2'), head=dict(in_channels=2048))
--- a/configs/repvgg/repvgg-B1g4_8xb32_in1k.py
+++ b/configs/repvgg/repvgg-B1g4_8xb32_in1k.py
+_base_ = './repvgg-A0_8xb32_in1k.py'
+
+model = dict(backbone=dict(arch='B1g4'), head=dict(in_channels=2048))
--- a/configs/repvgg/repvgg-B2_8xb32_in1k.py
+++ b/configs/repvgg/repvgg-B2_8xb32_in1k.py
+_base_ = './repvgg-A0_8xb32_in1k.py'
+
+model = dict(backbone=dict(arch='B2'), head=dict(in_channels=2560))
--- a/configs/repvgg/repvgg-B2g4_8xb32_in1k.py
+++ b/configs/repvgg/repvgg-B2g4_8xb32_in1k.py
+_base_ = './repvgg-B3_8xb32_in1k.py'
+
+model = dict(backbone=dict(arch='B2g4'), head=dict(in_channels=2560))
--- a/configs/repvgg/repvgg-B3_8xb32_in1k.py
+++ b/configs/repvgg/repvgg-B3_8xb32_in1k.py
+_base_ = [
+    '../_base_/models/repvgg-B3_lbs-mixup_in1k.py',
+    '../_base_/datasets/imagenet_bs32_pil_resize.py',
+    '../_base_/schedules/imagenet_bs256_coslr.py',
+    '../_base_/default_runtime.py'
+]
+
+# schedule settings
+optim_wrapper = dict(
+    paramwise_cfg=dict(
+        bias_decay_mult=0.0,
+        custom_keys={
+            'branch_3x3.norm': dict(decay_mult=0.0),
+            'branch_1x1.norm': dict(decay_mult=0.0),
+            'branch_norm.bias': dict(decay_mult=0.0),
+        }))
+
+data_preprocessor = dict(
+    # RGB format normalization parameters
+    mean=[123.675, 116.28, 103.53],
+    std=[58.395, 57.12, 57.375],
+    # convert image from BGR to RGB
+    to_rgb=True,
+)
+
+bgr_mean = data_preprocessor['mean'][::-1]
+bgr_std = data_preprocessor['std'][::-1]
+
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='RandomResizedCrop', scale=224, backend='pillow'),
+    dict(type='RandomFlip', prob=0.5, direction='horizontal'),
+    dict(
+        type='RandAugment',
+        policies='timm_increasing',
+        num_policies=2,
+        total_level=10,
+        magnitude_level=7,
+        magnitude_std=0.5,
+        hparams=dict(pad_val=[round(x) for x in bgr_mean])),
+    dict(type='PackInputs'),
+]
+
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='ResizeEdge', scale=256, edge='short', backend='pillow'),
+    dict(type='CenterCrop', crop_size=224),
+    dict(type='PackInputs'),
+]
+
+train_dataloader = dict(dataset=dict(pipeline=train_pipeline))
+val_dataloader = dict(dataset=dict(pipeline=test_pipeline))
+test_dataloader = dict(dataset=dict(pipeline=test_pipeline))
+
+# schedule settings
+param_scheduler = dict(
+    type='CosineAnnealingLR',
+    T_max=200,
+    by_epoch=True,
+    begin=0,
+    end=200,
+    convert_to_iter_based=True)
+
+train_cfg = dict(by_epoch=True, max_epochs=200)
+
+default_hooks = dict(
+    checkpoint=dict(type='CheckpointHook', interval=1, max_keep_ckpts=3))
--- a/configs/repvgg/repvgg-B3g4_8xb32_in1k.py
+++ b/configs/repvgg/repvgg-B3g4_8xb32_in1k.py
+_base_ = './repvgg-B3_8xb32_in1k.py'
+
+model = dict(backbone=dict(arch='B3g4'))
--- a/configs/repvgg/repvgg-D2se_8xb32_in1k.py
+++ b/configs/repvgg/repvgg-D2se_8xb32_in1k.py
+_base_ = './repvgg-B3_8xb32_in1k.py'
+
+model = dict(backbone=dict(arch='D2se'), head=dict(in_channels=2560))
+
+param_scheduler = [
+    # warm up learning rate scheduler
+    dict(
+        type='LinearLR',
+        start_factor=0.0001,
+        by_epoch=True,
+        begin=0,
+        end=5,
+        # update by iter
+        convert_to_iter_based=True),
+    # main learning rate scheduler
+    dict(
+        type='CosineAnnealingLR',
+        T_max=295,
+        eta_min=1.0e-6,
+        by_epoch=True,
+        begin=5,
+        end=300)
+]
+
+train_cfg = dict(by_epoch=True, max_epochs=300)
+
+default_hooks = dict(
+    checkpoint=dict(type='CheckpointHook', interval=1, max_keep_ckpts=3))
--- a/configs/res2net/README.md
+++ b/configs/res2net/README.md
+# Res2Net
+
+> [Res2Net: A New Multi-scale Backbone Architecture](https://arxiv.org/abs/1904.01169)
+
+<!-- [ALGORITHM] -->
+
+## Abstract
+
+Representing features at multiple scales is of great importance for numerous vision tasks. Recent advances in backbone convolutional neural networks (CNNs) continually demonstrate stronger multi-scale representation ability, leading to consistent performance gains on a wide range of applications. However, most existing methods represent the multi-scale features in a layer-wise manner. In this paper, we propose a novel building block for CNNs, namely Res2Net, by constructing hierarchical residual-like connections within one single residual block. The Res2Net represents multi-scale features at a granular level and increases the range of receptive fields for each network layer. The proposed Res2Net block can be plugged into the state-of-the-art backbone CNN models, e.g., ResNet, ResNeXt, and DLA. We evaluate the Res2Net block on all these models and demonstrate consistent performance gains over baseline models on widely-used datasets, e.g., CIFAR-100 and ImageNet. Further ablation studies and experimental results on representative computer vision tasks, i.e., object detection, class activation mapping, and salient object detection, further verify the superiority of the Res2Net over the state-of-the-art baseline methods.
+
+<div align=center>
+<img src="https://user-images.githubusercontent.com/26739999/142573547-cde68abf-287b-46db-a848-5cffe3068faf.png" width="50%"/>
+</div>
+
+## How to use it?
+
+<!-- [TABS-BEGIN] -->
+
+**Predict image**
+
+```python
+from mmpretrain import inference_model
+
+predict = inference_model('res2net50-w14-s8_3rdparty_8xb32_in1k', 'demo/bird.JPEG')
+print(predict['pred_class'])
+print(predict['pred_score'])
+```
+
+**Use the model**
+
+```python
+import torch
+from mmpretrain import get_model
+
+model = get_model('res2net50-w14-s8_3rdparty_8xb32_in1k', pretrained=True)
+inputs = torch.rand(1, 3, 224, 224)
+out = model(inputs)
+print(type(out))
+# To extract features.
+feats = model.extract_feat(inputs)
+print(type(feats))
+```
+
+**Test Command**
+
+Prepare your dataset according to the [docs](https://mmpretrain.readthedocs.io/en/latest/user_guides/dataset_prepare.html#prepare-dataset).
+
+Test:
+
+```shell
+python tools/test.py configs/res2net/res2net50-w14-s8_8xb32_in1k.py https://download.openmmlab.com/mmclassification/v0/res2net/res2net50-w14-s8_3rdparty_8xb32_in1k_20210927-bc967bf1.pth
+```
+
+<!-- [TABS-END] -->
+
+## Models and results
+
+### Image Classification on ImageNet-1k
+
+| Model                                     |   Pretrain   | Params (M) | Flops (G) | Top-1 (%) | Top-5 (%) |                  Config                   |                               Download                                |
+| :---------------------------------------- | :----------: | :--------: | :-------: | :-------: | :-------: | :---------------------------------------: | :-------------------------------------------------------------------: |
+| `res2net50-w14-s8_3rdparty_8xb32_in1k`\*  | From scratch |   25.06    |   4.22    |   78.14   |   93.85   | [config](res2net50-w14-s8_8xb32_in1k.py)  | [model](https://download.openmmlab.com/mmclassification/v0/res2net/res2net50-w14-s8_3rdparty_8xb32_in1k_20210927-bc967bf1.pth) |
+| `res2net50-w26-s8_3rdparty_8xb32_in1k`\*  | From scratch |   48.40    |   8.39    |   79.20   |   94.36   | [config](res2net50-w26-s8_8xb32_in1k.py)  | [model](https://download.openmmlab.com/mmclassification/v0/res2net/res2net50-w26-s8_3rdparty_8xb32_in1k_20210927-f547a94b.pth) |
+| `res2net101-w26-s4_3rdparty_8xb32_in1k`\* | From scratch |   45.21    |   8.12    |   79.19   |   94.44   | [config](res2net101-w26-s4_8xb32_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/res2net/res2net101-w26-s4_3rdparty_8xb32_in1k_20210927-870b6c36.pth) |
+
+*Models with * are converted from the [official repo](https://github.com/Res2Net/Res2Net-PretrainedModels/blob/master/res2net.py#L181). The config files of these models are only for inference. We haven't reproduce the training results.*
+
+## Citation
+
+```bibtex
+@article{gao2019res2net,
+  title={Res2Net: A New Multi-scale Backbone Architecture},
+  author={Gao, Shang-Hua and Cheng, Ming-Ming and Zhao, Kai and Zhang, Xin-Yu and Yang, Ming-Hsuan and Torr, Philip},
+  journal={IEEE TPAMI},
+  year={2021},
+  doi={10.1109/TPAMI.2019.2938758},
+}
+```
--- a/configs/res2net/metafile.yml
+++ b/configs/res2net/metafile.yml
+Collections:
+  - Name: Res2Net
+    Metadata:
+      Training Data: ImageNet-1k
+      Training Techniques:
+        - SGD with Momentum
+        - Weight Decay
+      Architecture:
+        - Batch Normalization
+        - Convolution
+        - Global Average Pooling
+        - ReLU
+        - Res2Net Block
+    Paper:
+      Title: 'Res2Net: A New Multi-scale Backbone Architecture'
+      URL: https://arxiv.org/abs/1904.01169
+    README: configs/res2net/README.md
+    Code:
+      URL: https://github.com/open-mmlab/mmpretrain/blob/v0.17.0/mmcls/models/backbones/res2net.py
+      Version: v0.17.0
+
+Models:
+  - Name: res2net50-w14-s8_3rdparty_8xb32_in1k
+    Metadata:
+      FLOPs: 4220000000
+      Parameters: 25060000
+    In Collection: Res2Net
+    Results:
+      - Dataset: ImageNet-1k
+        Metrics:
+          Top 1 Accuracy: 78.14
+          Top 5 Accuracy: 93.85
+        Task: Image Classification
+    Weights: https://download.openmmlab.com/mmclassification/v0/res2net/res2net50-w14-s8_3rdparty_8xb32_in1k_20210927-bc967bf1.pth
+    Converted From:
+      Weights: https://1drv.ms/u/s!AkxDDnOtroRPdOTqhF8ne_aakDI?e=EVb8Ri
+      Code: https://github.com/Res2Net/Res2Net-PretrainedModels/blob/master/res2net.py#L221
+    Config: configs/res2net/res2net50-w14-s8_8xb32_in1k.py
+  - Name: res2net50-w26-s8_3rdparty_8xb32_in1k
+    Metadata:
+      FLOPs: 8390000000
+      Parameters: 48400000
+    In Collection: Res2Net
+    Results:
+      - Dataset: ImageNet-1k
+        Metrics:
+          Top 1 Accuracy: 79.20
+          Top 5 Accuracy: 94.36
+        Task: Image Classification
+    Weights: https://download.openmmlab.com/mmclassification/v0/res2net/res2net50-w26-s8_3rdparty_8xb32_in1k_20210927-f547a94b.pth
+    Converted From:
+      Weights: https://1drv.ms/u/s!AkxDDnOtroRPdTrAd_Afzc26Z7Q?e=slYqsR
+      Code: https://github.com/Res2Net/Res2Net-PretrainedModels/blob/master/res2net.py#L201
+    Config: configs/res2net/res2net50-w26-s8_8xb32_in1k.py
+  - Name: res2net101-w26-s4_3rdparty_8xb32_in1k
+    Metadata:
+      FLOPs: 8120000000
+      Parameters: 45210000
+    In Collection: Res2Net
+    Results:
+      - Dataset: ImageNet-1k
+        Metrics:
+          Top 1 Accuracy: 79.19
+          Top 5 Accuracy: 94.44
+        Task: Image Classification
+    Weights: https://download.openmmlab.com/mmclassification/v0/res2net/res2net101-w26-s4_3rdparty_8xb32_in1k_20210927-870b6c36.pth
+    Converted From:
+      Weights: https://1drv.ms/u/s!AkxDDnOtroRPcJRgTLkahL0cFYw?e=nwbnic
+      Code: https://github.com/Res2Net/Res2Net-PretrainedModels/blob/master/res2net.py#L181
+    Config: configs/res2net/res2net101-w26-s4_8xb32_in1k.py
--- a/configs/res2net/res2net101-w26-s4_8xb32_in1k.py
+++ b/configs/res2net/res2net101-w26-s4_8xb32_in1k.py
+_base_ = [
+    '../_base_/models/res2net101-w26-s4.py',
+    '../_base_/datasets/imagenet_bs32_pil_resize.py',
+    '../_base_/schedules/imagenet_bs256.py', '../_base_/default_runtime.py'
+]