first commit

dff2c686 · renzhc · 8f9dd0ed · dff2c686 · dff2c686 · dff2c686
Commit dff2c686 authored Sep 03, 2024 by renzhc
20 changed files
--- a/configs/regnet/regnetx-3.2gf_8xb64_in1k.py
+++ b/configs/regnet/regnetx-3.2gf_8xb64_in1k.py
+_base_ = ['./regnetx-400mf_8xb128_in1k.py']
+
+# model settings
+model = dict(
+    backbone=dict(type='RegNet', arch='regnetx_3.2gf'),
+    head=dict(in_channels=1008, ))
+
+# dataset settings
+train_dataloader = dict(batch_size=64)
+
+# schedule settings
+# for batch_size 512, use lr = 0.4
+optim_wrapper = dict(optimizer=dict(lr=0.4))
+
+# NOTE: `auto_scale_lr` is for automatically scaling LR
+# based on the actual training batch size.
+# base_batch_size = (8 GPUs) x (64 samples per GPU)
+auto_scale_lr = dict(base_batch_size=512)
--- a/configs/regnet/regnetx-4.0gf_8xb64_in1k.py
+++ b/configs/regnet/regnetx-4.0gf_8xb64_in1k.py
+_base_ = ['./regnetx-400mf_8xb128_in1k.py']
+
+# model settings
+model = dict(
+    backbone=dict(type='RegNet', arch='regnetx_4.0gf'),
+    head=dict(in_channels=1360, ))
+
+# dataset settings
+train_dataloader = dict(batch_size=64)
+
+# schedule settings
+# for batch_size 512, use lr = 0.4
+optim_wrapper = dict(optimizer=dict(lr=0.4))
+
+# NOTE: `auto_scale_lr` is for automatically scaling LR
+# based on the actual training batch size.
+# base_batch_size = (8 GPUs) x (64 samples per GPU)
+auto_scale_lr = dict(base_batch_size=512)
--- a/configs/regnet/regnetx-400mf_8xb128_in1k.py
+++ b/configs/regnet/regnetx-400mf_8xb128_in1k.py
+_base_ = [
+    '../_base_/models/regnet/regnetx_400mf.py',
+    '../_base_/datasets/imagenet_bs32.py',
+    '../_base_/schedules/imagenet_bs1024_coslr.py',
+    '../_base_/default_runtime.py'
+]
+
+# dataset settings
+data_preprocessor = dict(
+    # BGR format normalization parameters
+    mean=[103.53, 116.28, 123.675],
+    std=[57.375, 57.12, 58.395],
+    to_rgb=False,  # The checkpoints from PyCls requires BGR format inputs.
+)
+
+# lighting params, in order of BGR, from repo. pycls
+EIGVAL = [0.2175, 0.0188, 0.0045]
+EIGVEC = [
+    [-0.5836, -0.6948, 0.4203],
+    [-0.5808, -0.0045, -0.814],
+    [-0.5675, 0.7192, 0.4009],
+]
+
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='RandomResizedCrop', scale=224),
+    dict(type='RandomFlip', prob=0.5, direction='horizontal'),
+    dict(
+        type='Lighting',
+        eigval=EIGVAL,
+        eigvec=EIGVEC,
+        alphastd=25.5,  # because the value range of images is [0,255]
+        to_rgb=False),
+    dict(type='PackInputs'),
+]
+
+train_dataloader = dict(batch_size=128, dataset=dict(pipeline=train_pipeline))
+val_dataloader = dict(batch_size=128)
+test_dataloader = dict(batch_size=128)
+
+# schedule settings
+
+# sgd with nesterov, base ls is 0.8 for batch_size 1024,
+optim_wrapper = dict(optimizer=dict(lr=0.8, nesterov=True))
+
+# runtime settings
+
+# Precise BN hook will update the bn stats, so this hook should be executed
+# before CheckpointHook(priority of 'VERY_LOW') and
+# EMAHook(priority of 'NORMAL') So set the priority of PreciseBNHook to
+# 'ABOVENORMAL' here.
+custom_hooks = [
+    dict(
+        type='PreciseBNHook',
+        num_samples=8192,
+        interval=1,
+        priority='ABOVE_NORMAL')
+]
--- a/configs/regnet/regnetx-6.4gf_8xb64_in1k.py
+++ b/configs/regnet/regnetx-6.4gf_8xb64_in1k.py
+_base_ = ['./regnetx-400mf_8xb128_in1k.py']
+
+# model settings
+model = dict(
+    backbone=dict(type='RegNet', arch='regnetx_6.4gf'),
+    head=dict(in_channels=1624, ))
+
+# dataset settings
+train_dataloader = dict(batch_size=64)
+
+# schedule settings
+# for batch_size 512, use lr = 0.4
+optim_wrapper = dict(optimizer=dict(lr=0.4))
+
+# NOTE: `auto_scale_lr` is for automatically scaling LR
+# based on the actual training batch size.
+# base_batch_size = (8 GPUs) x (64 samples per GPU)
+auto_scale_lr = dict(base_batch_size=512)
--- a/configs/regnet/regnetx-8.0gf_8xb64_in1k.py
+++ b/configs/regnet/regnetx-8.0gf_8xb64_in1k.py
+_base_ = ['./regnetx-400mf_8xb128_in1k.py']
+
+# model settings
+model = dict(
+    backbone=dict(type='RegNet', arch='regnetx_8.0gf'),
+    head=dict(in_channels=1920, ))
+
+# dataset settings
+train_dataloader = dict(batch_size=64)
+
+# schedule settings
+# for batch_size 512, use lr = 0.4
+optim_wrapper = dict(optimizer=dict(lr=0.4))
+
+# NOTE: `auto_scale_lr` is for automatically scaling LR
+# based on the actual training batch size.
+# base_batch_size = (8 GPUs) x (64 samples per GPU)
+auto_scale_lr = dict(base_batch_size=512)
--- a/configs/regnet/regnetx-800mf_8xb128_in1k.py
+++ b/configs/regnet/regnetx-800mf_8xb128_in1k.py
+_base_ = ['./regnetx-400mf_8xb128_in1k.py']
+
+# model settings
+model = dict(
+    backbone=dict(type='RegNet', arch='regnetx_800mf'),
+    head=dict(in_channels=672, ))
--- a/configs/replknet/README.md
+++ b/configs/replknet/README.md
+# RepLKNet
+
+> [Scaling Up Your Kernels to 31x31: Revisiting Large Kernel Design in CNNs](https://arxiv.org/abs/2203.06717)
+
+<!-- [ALGORITHM] -->
+
+## Abstract
+
+We revisit large kernel design in modern convolutional neural networks (CNNs). Inspired by recent advances in vision transformers (ViTs), in this paper, we demonstrate that using a few large convolutional kernels instead of a stack of small kernels could be a more powerful paradigm. We suggested five guidelines, e.g., applying re-parameterized large depth-wise convolutions, to design efficient highperformance large-kernel CNNs. Following the guidelines, we propose RepLKNet, a pure CNN architecture whose kernel size is as large as 31×31, in contrast to commonly used 3×3. RepLKNet greatly closes the performance gap between CNNs and ViTs, e.g., achieving comparable or superior results than Swin Transformer on ImageNet and a few typical downstream tasks, with lower latency. RepLKNet also shows nice scalability to big data and large models, obtaining 87.8% top-1 accuracy on ImageNet and 56.0% mIoU on ADE20K, which is very competitive among the state-of-the-arts with similar model sizes. Our study further reveals that, in contrast to small-kernel CNNs, large kernel CNNs have much larger effective receptive fields and higher shape bias rather than texture bias.
+
+<div align=center>
+<img src="https://user-images.githubusercontent.com/48375204/197546040-cdf078c3-7fbd-400f-8b27-01668c8dfebf.png" width="60%"/>
+</div>
+
+## How to use it?
+
+<!-- [TABS-BEGIN] -->
+
+**Predict image**
+
+```python
+from mmpretrain import inference_model, get_model
+
+model = get_model('replknet-31B_3rdparty_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('replknet-31B_3rdparty_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/replknet/replknet-31B_32xb64_in1k.py https://download.openmmlab.com/mmclassification/v0/replknet/replknet-31B_3rdparty_in1k_20221118-fd08e268.pth
+```
+
+**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 RepLKNet
+
+backbone = RepLKNet(arch='31B')
+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                            |
+| :--------------------------------------------- | :----------: | :--------: | :-------: | :-------: | :-------: | :-----------------------------------------: | :------------------------------------------------------------: |
+| `replknet-31B_3rdparty_in1k`\*                 | From scratch |   79.86    |   15.64   |   83.48   |   96.57   |    [config](replknet-31B_32xb64_in1k.py)    | [model](https://download.openmmlab.com/mmclassification/v0/replknet/replknet-31B_3rdparty_in1k_20221118-fd08e268.pth) |
+| `replknet-31B_3rdparty_in1k-384px`\*           | From scratch |   79.86    |   45.95   |   84.84   |   97.34   | [config](replknet-31B_32xb64_in1k-384px.py) | [model](https://download.openmmlab.com/mmclassification/v0/replknet/replknet-31B_3rdparty_in1k-384px_20221118-03a170ce.pth) |
+| `replknet-31B_in21k-pre_3rdparty_in1k`\*       | ImageNet-21k |   79.86    |   15.64   |   85.20   |   97.56   |    [config](replknet-31B_32xb64_in1k.py)    | [model](https://download.openmmlab.com/mmclassification/v0/replknet/replknet-31B_in21k-pre_3rdparty_in1k_20221118-54ed5c46.pth) |
+| `replknet-31B_in21k-pre_3rdparty_in1k-384px`\* | ImageNet-21k |   79.86    |   45.95   |   85.99   |   97.75   | [config](replknet-31B_32xb64_in1k-384px.py) | [model](https://download.openmmlab.com/mmclassification/v0/replknet/replknet-31B_in21k-pre_3rdparty_in1k-384px_20221118-76c92b24.pth) |
+| `replknet-31L_in21k-pre_3rdparty_in1k-384px`\* | ImageNet-21k |   172.67   |   97.24   |   86.63   |   98.00   | [config](replknet-31L_32xb64_in1k-384px.py) | [model](https://download.openmmlab.com/mmclassification/v0/replknet/replknet-31L_in21k-pre_3rdparty_in1k-384px_20221118-dc3fc07c.pth) |
+| `replknet-XL_meg73m-pre_3rdparty_in1k-320px`\* |    MEG73M    |   335.44   |  129.57   |   87.57   |   98.39   | [config](replknet-XL_32xb64_in1k-320px.py)  | [model](https://download.openmmlab.com/mmclassification/v0/replknet/replknet-XL_meg73m-pre_3rdparty_in1k-320px_20221118-88259b1d.pth) |
+
+*Models with * are converted from the [official repo](https://github.com/DingXiaoH/RepLKNet-pytorch/blob/main/replknet.py). The config files of these models are only for inference. We haven't reproduce the training results.*
+
+## Citation
+
+```bibtex
+@inproceedings{ding2022scaling,
+  title={Scaling up your kernels to 31x31: Revisiting large kernel design in cnns},
+  author={Ding, Xiaohan and Zhang, Xiangyu and Han, Jungong and Ding, Guiguang},
+  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
+  pages={11963--11975},
+  year={2022}
+}
+```
--- a/configs/replknet/deploy/replknet-31B-deploy_32xb64_in1k-384px.py
+++ b/configs/replknet/deploy/replknet-31B-deploy_32xb64_in1k-384px.py
+_base_ = '../replknet-31B_32xb64_in1k-384px.py'
+
+model = dict(backbone=dict(small_kernel_merged=True))
--- a/configs/replknet/deploy/replknet-31B-deploy_32xb64_in1k.py
+++ b/configs/replknet/deploy/replknet-31B-deploy_32xb64_in1k.py
+_base_ = '../replknet-31B_32xb64_in1k.py'
+
+model = dict(backbone=dict(small_kernel_merged=True))
--- a/configs/replknet/deploy/replknet-31L-deploy_32xb64_in1k-384px.py
+++ b/configs/replknet/deploy/replknet-31L-deploy_32xb64_in1k-384px.py
+_base_ = '../replknet-31L_32xb64_in1k-384px.py'
+
+model = dict(backbone=dict(small_kernel_merged=True))
--- a/configs/replknet/deploy/replknet-XL-deploy_32xb64_in1k-320px.py
+++ b/configs/replknet/deploy/replknet-XL-deploy_32xb64_in1k-320px.py
+_base_ = '../replknet-XL_32xb64_in1k-320px.py'
+
+model = dict(backbone=dict(small_kernel_merged=True))
--- a/configs/replknet/metafile.yml
+++ b/configs/replknet/metafile.yml
+Collections:
+  - Name: RepLKNet
+    Metadata:
+      Training Data: ImageNet-1k
+      Architecture:
+        - Large-Kernel Convolution
+        - VGG-style Neural Network
+    Paper:
+      URL: https://arxiv.org/abs/2203.06717
+      Title: 'Scaling Up Your Kernels to 31x31: Revisiting Large Kernel Design in CNNs'
+    README: configs/replknet/README.md
+    Code:
+      URL: https://github.com/open-mmlab/mmpretrain/blob/v1.0.0rc3/mmcls/models/backbones/replknet.py
+      Version: v1.0.0rc3
+
+Models:
+  - Name: replknet-31B_3rdparty_in1k
+    In Collection: RepLKNet
+    Config: configs/replknet/replknet-31B_32xb64_in1k.py
+    Metadata:
+      FLOPs: 15636547584
+      Parameters: 79864168
+    Results:
+    - Dataset: ImageNet-1k
+      Task: Image Classification
+      Metrics:
+        Top 1 Accuracy: 83.48
+        Top 5 Accuracy: 96.57
+    Weights: https://download.openmmlab.com/mmclassification/v0/replknet/replknet-31B_3rdparty_in1k_20221118-fd08e268.pth
+    Converted From:
+      Weights: https://drive.google.com/u/0/uc?id=1azQUiCxK9feYVkkrPqwVPBtNsTzDrX7S&export=download
+      Code: https://github.com/DingXiaoH/RepLKNet-pytorch/blob/main/replknet.py
+
+  - Name: replknet-31B_3rdparty_in1k-384px
+    In Collection: RepLKNet
+    Config: configs/replknet/replknet-31B_32xb64_in1k-384px.py
+    Metadata:
+      FLOPs: 45952303104
+      Parameters: 79864168
+    Results:
+    - Dataset: ImageNet-1k
+      Task: Image Classification
+      Metrics:
+        Top 1 Accuracy: 84.84
+        Top 5 Accuracy: 97.34
+    Weights: https://download.openmmlab.com/mmclassification/v0/replknet/replknet-31B_3rdparty_in1k-384px_20221118-03a170ce.pth
+    Converted From:
+      Weights: https://drive.google.com/u/0/uc?id=1vo-P3XB6mRLUeDzmgv90dOu73uCeLfZN&export=download
+      Code: https://github.com/DingXiaoH/RepLKNet-pytorch/blob/main/replknet.py
+
+  - Name: replknet-31B_in21k-pre_3rdparty_in1k
+    In Collection: RepLKNet
+    Config: configs/replknet/replknet-31B_32xb64_in1k.py
+    Metadata:
+      Training Data:
+        - ImageNet-21k
+        - ImageNet-1k
+      FLOPs: 15636547584
+      Parameters: 79864168
+    Results:
+    - Dataset: ImageNet-1k
+      Task: Image Classification
+      Metrics:
+        Top 1 Accuracy: 85.20
+        Top 5 Accuracy: 97.56
+    Weights: https://download.openmmlab.com/mmclassification/v0/replknet/replknet-31B_in21k-pre_3rdparty_in1k_20221118-54ed5c46.pth
+    Converted From:
+      Weights: https://drive.google.com/u/0/uc?id=1DslZ2voXZQR1QoFY9KnbsHAeF84hzS0s&export=download
+      Code: https://github.com/DingXiaoH/RepLKNet-pytorch/blob/main/replknet.py
+
+  - Name: replknet-31B_in21k-pre_3rdparty_in1k-384px
+    In Collection: RepLKNet
+    Config: configs/replknet/replknet-31B_32xb64_in1k-384px.py
+    Metadata:
+      Training Data:
+        - ImageNet-21k
+        - ImageNet-1k
+      FLOPs: 45952303104
+      Parameters: 79864168
+    Results:
+    - Dataset: ImageNet-1k
+      Task: Image Classification
+      Metrics:
+        Top 1 Accuracy: 85.99
+        Top 5 Accuracy: 97.75
+    Weights: https://download.openmmlab.com/mmclassification/v0/replknet/replknet-31B_in21k-pre_3rdparty_in1k-384px_20221118-76c92b24.pth
+    Converted From:
+      Weights: https://drive.google.com/u/0/uc?id=1Sc46BWdXXm2fVP-K_hKKU_W8vAB-0duX&export=download
+      Code: https://github.com/DingXiaoH/RepLKNet-pytorch/blob/main/replknet.py
+
+  - Name: replknet-31L_in21k-pre_3rdparty_in1k-384px
+    In Collection: RepLKNet
+    Config: configs/replknet/replknet-31L_32xb64_in1k-384px.py
+    Metadata:
+      Training Data:
+        - ImageNet-21k
+        - ImageNet-1k
+      FLOPs: 97240006656
+      Parameters: 172671016
+    Results:
+    - Dataset: ImageNet-1k
+      Task: Image Classification
+      Metrics:
+        Top 1 Accuracy: 86.63
+        Top 5 Accuracy: 98.00
+    Weights: https://download.openmmlab.com/mmclassification/v0/replknet/replknet-31L_in21k-pre_3rdparty_in1k-384px_20221118-dc3fc07c.pth
+    Converted From:
+      Weights: https://drive.google.com/u/0/uc?id=1JYXoNHuRvC33QV1pmpzMTKEni1hpWfBl&export=download
+      Code: https://github.com/DingXiaoH/RepLKNet-pytorch/blob/main/replknet.py
+
+  - Name: replknet-XL_meg73m-pre_3rdparty_in1k-320px
+    In Collection: RepLKNet
+    Config: configs/replknet/replknet-XL_32xb64_in1k-320px.py
+    Metadata:
+      Training Data:
+        - MegData-73M
+        - ImageNet-1k
+      FLOPs: 129570201600
+      Parameters: 335435752
+    Results:
+    - Dataset: ImageNet-1k
+      Task: Image Classification
+      Metrics:
+        Top 1 Accuracy: 87.57
+        Top 5 Accuracy: 98.39
+    Weights: https://download.openmmlab.com/mmclassification/v0/replknet/replknet-XL_meg73m-pre_3rdparty_in1k-320px_20221118-88259b1d.pth
+    Converted From:
+      Weights: https://drive.google.com/u/0/uc?id=1tPC60El34GntXByIRHb-z-Apm4Y5LX1T&export=download
+      Code: https://github.com/DingXiaoH/RepLKNet-pytorch/blob/main/replknet.py
--- a/configs/replknet/replknet-31B_32xb64_in1k-384px.py
+++ b/configs/replknet/replknet-31B_32xb64_in1k-384px.py
+_base_ = [
+    '../_base_/models/replknet-31B_in1k.py',
+    '../_base_/datasets/imagenet_bs16_pil_bicubic_384.py',
+    '../_base_/schedules/imagenet_bs256_coslr.py',
+    '../_base_/default_runtime.py'
+]
+
+# schedule settings
+param_scheduler = dict(
+    type='CosineAnnealingLR', T_max=300, by_epoch=True, begin=0, end=300)
+
+train_cfg = dict(by_epoch=True, max_epochs=300)
--- a/configs/replknet/replknet-31B_32xb64_in1k.py
+++ b/configs/replknet/replknet-31B_32xb64_in1k.py
+_base_ = [
+    '../_base_/models/replknet-31B_in1k.py',
+    '../_base_/datasets/imagenet_bs32_pil_bicubic.py',
+    '../_base_/schedules/imagenet_bs256_coslr.py',
+    '../_base_/default_runtime.py'
+]
+
+# schedule settings
+param_scheduler = dict(
+    type='CosineAnnealingLR', T_max=300, by_epoch=True, begin=0, end=300)
+
+train_cfg = dict(by_epoch=True, max_epochs=300)
--- a/configs/replknet/replknet-31L_32xb64_in1k-384px.py
+++ b/configs/replknet/replknet-31L_32xb64_in1k-384px.py
+_base_ = [
+    '../_base_/models/replknet-31L_in1k.py',
+    '../_base_/datasets/imagenet_bs16_pil_bicubic_384.py',
+    '../_base_/schedules/imagenet_bs256_coslr.py',
+    '../_base_/default_runtime.py'
+]
+
+# schedule settings
+param_scheduler = dict(
+    type='CosineAnnealingLR', T_max=300, by_epoch=True, begin=0, end=300)
+
+train_cfg = dict(by_epoch=True, max_epochs=300)
--- a/configs/replknet/replknet-XL_32xb64_in1k-320px.py
+++ b/configs/replknet/replknet-XL_32xb64_in1k-320px.py
+_base_ = [
+    '../_base_/models/replknet-XL_in1k.py',
+    '../_base_/datasets/imagenet_bs8_pil_bicubic_320.py',
+    '../_base_/schedules/imagenet_bs256_coslr.py',
+    '../_base_/default_runtime.py'
+]
+
+# schedule settings
+param_scheduler = dict(
+    type='CosineAnnealingLR', T_max=300, by_epoch=True, begin=0, end=300)
+
+train_cfg = dict(by_epoch=True, max_epochs=300)
--- a/configs/repmlp/README.md
+++ b/configs/repmlp/README.md
+# RepMLP
+
+> [RepMLP: Re-parameterizing Convolutions into Fully-connected Layers for Image Recognition](https://arxiv.org/abs/2105.01883)
+
+<!-- [ALGORITHM] -->
+
+## Abstract
+
+We propose RepMLP, a multi-layer-perceptron-style neural network building block for image recognition, which is composed of a series of fully-connected (FC) layers. Compared to convolutional layers, FC layers are more efficient, better at modeling the long-range dependencies and positional patterns, but worse at capturing the local structures, hence usually less favored for image recognition. We propose a structural re-parameterization technique that adds local prior into an FC to make it powerful for image recognition. Specifically, we construct convolutional layers inside a RepMLP during training and merge them into the FC for inference. On CIFAR, a simple pure-MLP model shows performance very close to CNN. By inserting RepMLP in traditional CNN, we improve ResNets by 1.8% accuracy on ImageNet, 2.9% for face recognition, and 2.3% mIoU on Cityscapes with lower FLOPs. Our intriguing findings highlight that combining the global representational capacity and positional perception of FC with the local prior of convolution can improve the performance of neural network with faster speed on both the tasks with translation invariance (e.g., semantic segmentation) and those with aligned images and positional patterns (e.g., face recognition).
+
+<div align=center>
+<img src="https://user-images.githubusercontent.com/18586273/155455288-a17a5c48-11af-4b74-995a-cf7183f0e2d2.png" width="80%"/>
+</div>
+
+## How to use it?
+
+<!-- [TABS-BEGIN] -->
+
+**Predict image**
+
+```python
+from mmpretrain import inference_model, get_model
+
+model = get_model('repmlp-base_3rdparty_8xb64_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('repmlp-base_3rdparty_8xb64_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/repmlp/repmlp-base_8xb64_in1k.py https://download.openmmlab.com/mmclassification/v0/repmlp/repmlp-base_3rdparty_8xb64_in1k_20220330-1cb1f11b.pth
+```
+
+**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 RepMLPNet
+
+backbone = RepMLPNet(arch='B', img_size=224, reparam_conv_kernels=(1, 3))
+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                                |
+| :---------------------------------------- | :----------: | :--------: | :-------: | :-------: | :-------: | :---------------------------------------: | :-------------------------------------------------------------------: |
+| `repmlp-base_3rdparty_8xb64_in1k`\*       | From scratch |   68.24    |   6.71    |   80.41   |   95.14   |    [config](repmlp-base_8xb64_in1k.py)    | [model](https://download.openmmlab.com/mmclassification/v0/repmlp/repmlp-base_3rdparty_8xb64_in1k_20220330-1cb1f11b.pth) |
+| `repmlp-base_3rdparty_8xb64_in1k-256px`\* | From scratch |   96.45    |   9.69    |   81.11   |   95.50   | [config](repmlp-base_8xb64_in1k-256px.py) | [model](https://download.openmmlab.com/mmclassification/v0/repmlp/repmlp-base_3rdparty_8xb64_in1k-256px_20220330-7c5a91ce.pth) |
+
+*Models with * are converted from the [official repo](https://github.com/DingXiaoH/RepMLP/blob/072d8516beba83d75dfe6ebb12f625abad4b53d5/repmlpnet.py#L278). The config files of these models are only for inference. We haven't reproduce the training results.*
+
+## Citation
+
+```bibtex
+@article{ding2021repmlp,
+  title={Repmlp: Re-parameterizing convolutions into fully-connected layers for image recognition},
+  author={Ding, Xiaohan and Xia, Chunlong and Zhang, Xiangyu and Chu, Xiaojie and Han, Jungong and Ding, Guiguang},
+  journal={arXiv preprint arXiv:2105.01883},
+  year={2021}
+}
+```
--- a/configs/repmlp/metafile.yml
+++ b/configs/repmlp/metafile.yml
+Collections:
+  - Name: RepMLP
+    Metadata:
+      Training Data: ImageNet-1k
+      Architecture:
+        - Multi-layer Perceptron
+        - Re-parameterization Convolution
+    Paper:
+      URL: https://arxiv.org/abs/2105.01883
+      Title: 'RepMLP: Re-parameterizing Convolutions into Fully-connected Layers for Image Recognition'
+    README: configs/repmlp/README.md
+    Code:
+      URL: https://github.com/open-mmlab/mmpretrain/blob/v0.21.0/mmcls/models/backbones/repmlp.py
+      Version: v0.21.0
+
+Models:
+  - Name: repmlp-base_3rdparty_8xb64_in1k
+    In Collection: RepMLP
+    Config: configs/repmlp/repmlp-base_8xb64_in1k.py
+    Metadata:
+      FLOPs: 6710000000  # 6.71 G
+      Parameters: 68240000  # 68.24 M
+    Results:
+      - Dataset: ImageNet-1k
+        Metrics:
+          Top 1 Accuracy: 80.41
+          Top 5 Accuracy: 95.14
+        Task: Image Classification
+    Weights: https://download.openmmlab.com/mmclassification/v0/repmlp/repmlp-base_3rdparty_8xb64_in1k_20220330-1cb1f11b.pth
+    Converted From:
+      Weights: https://github.com/DingXiaoH/RepMLP
+      Code: https://github.com/DingXiaoH/RepMLP/blob/072d8516beba83d75dfe6ebb12f625abad4b53d5/repmlpnet.py#L274
+  - Name: repmlp-base_3rdparty_8xb64_in1k-256px
+    In Collection: RepMLP
+    Config: configs/repmlp/repmlp-base_8xb64_in1k-256px.py
+    Metadata:
+      FLOPs: 9690000000  # 9.69 G
+      Parameters: 96450000  # 96.45M
+    Results:
+      - Dataset: ImageNet-1k
+        Metrics:
+          Top 1 Accuracy: 81.11
+          Top 5 Accuracy: 95.50
+        Task: Image Classification
+    Weights: https://download.openmmlab.com/mmclassification/v0/repmlp/repmlp-base_3rdparty_8xb64_in1k-256px_20220330-7c5a91ce.pth
+    Converted From:
+      Weights: https://github.com/DingXiaoH/RepMLP
+      Code: https://github.com/DingXiaoH/RepMLP/blob/072d8516beba83d75dfe6ebb12f625abad4b53d5/repmlpnet.py#L278
--- a/configs/repmlp/repmlp-base_8xb64_in1k-256px.py
+++ b/configs/repmlp/repmlp-base_8xb64_in1k-256px.py
+_base_ = [
+    '../_base_/models/repmlp-base_224.py',
+    '../_base_/datasets/imagenet_bs64_pil_resize.py',
+    '../_base_/schedules/imagenet_bs4096_AdamW.py',
+    '../_base_/default_runtime.py'
+]
+
+# model settings
+model = dict(backbone=dict(img_size=256))
+
+# dataset settings
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='RandomResizedCrop', scale=256),
+    dict(type='RandomFlip', prob=0.5, direction='horizontal'),
+    dict(type='PackInputs'),
+]
+
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='ResizeEdge', scale=292, edge='short', backend='pillow'),
+    dict(type='CenterCrop', crop_size=256),
+    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
+optim_wrapper = dict(clip_grad=dict(max_norm=1.0))
+
+# NOTE: `auto_scale_lr` is for automatically scaling LR
+# based on the actual training batch size.
+# base_batch_size = (8 GPUs) x (64 samples per GPU)
+auto_scale_lr = dict(base_batch_size=512)
--- a/configs/repmlp/repmlp-base_8xb64_in1k.py
+++ b/configs/repmlp/repmlp-base_8xb64_in1k.py
+_base_ = [
+    '../_base_/models/repmlp-base_224.py',
+    '../_base_/datasets/imagenet_bs64_pil_resize.py',
+    '../_base_/schedules/imagenet_bs1024_adamw_swin.py',
+    '../_base_/default_runtime.py'
+]
+
+# dataset settings
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    # resizing to (256, 256) here, different from resizing shorter edge to 256
+    dict(type='Resize', scale=(256, 256), backend='pillow'),
+    dict(type='CenterCrop', crop_size=224),
+    dict(type='PackInputs'),
+]
+
+val_dataloader = dict(dataset=dict(pipeline=test_pipeline))
+test_dataloader = dict(dataset=dict(pipeline=test_pipeline))
+
+# schedule settings
+optim_wrapper = dict(clip_grad=dict(max_norm=5.0))
+
+# NOTE: `auto_scale_lr` is for automatically scaling LR
+# based on the actual training batch size.
+# base_batch_size = (8 GPUs) x (64 samples per GPU)
+auto_scale_lr = dict(base_batch_size=512)