Projecs Based on MMRazor

There are many research works and pre-trained models built on MMRazor. We list some of them as examples of how to use MMRazor slimmable models for downstream frameworks. As the page might not be completed, please feel free to contribute more efficient mmrazor-models to update this page.

Description

This is an implementation of MMRazor Searchable Backbone Application, we provide detection configs and models for MMRazor in MMYOLO.

Backbone support

Here are the Neural Architecture Search(NAS) Models that come from MMRazor which support YOLO Series. If you are looking for MMRazor models only for Backbone, you could refer to MMRazor ModelZoo and corresponding repository.

Usage

Prerequisites

MMRazor v1.0.0rc2 or higher (dev-1.x)

Install MMRazor using MIM.

mim install mmengine
mim install "mmrazor>=1.0.0rc2"

Install MMRazor from source

git clone -b dev-1.x https://github.com/open-mmlab/mmrazor.git
cd mmrazor
# Install MMRazor
mim install -v -e .

Training commands

In MMYOLO's root directory, if you want to use single GPU for training, run the following command to train the model:

CUDA_VISIBLE_DEVICES=0 PORT=29500 ./tools/dist_train.sh configs/razor/subnets/yolov5_s_spos_shufflenetv2_syncbn_8xb16-300e_coco.py

If you want to use several of these GPUs to train in parallel, you can use the following command:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 PORT=29500 ./tools/dist_train.sh configs/razor/subnets/yolov5_s_spos_shufflenetv2_syncbn_8xb16-300e_coco.py

Testing commands

In MMYOLO's root directory, run the following command to test the model:

CUDA_VISIBLE_DEVICES=0 PORT=29500 ./tools/dist_test.sh configs/razor/subnets/yolov5_s_spos_shufflenetv2_syncbn_8xb16-300e_coco.py ${CHECKPOINT_PATH}

Results and Models

Here we provide the baseline version of YOLO Series with NAS backbone.

Model	size	box AP	Params(M)	FLOPs(G)	Config	Download
yolov5-s	640	37.7	7.235	8.265	config	model \| log
yolov5_s_spos_shufflenetv2	640	38.0	7.04(-2.7%)	7.03	config	model \| log
yolov6-s	640	44.0	18.869	24.253	config	model \| log
yolov6_l_attentivenas_a6	640	45.3	18.38(-2.6%)	8.49	config	model \| log
RTMDet-tiny	640	41.0	4.8	8.1	config	model \| log
rtmdet_tiny_ofa_lat31	960	41.3	3.91(-18.5%)	6.09	config	model \| log

Note:

For fair comparison, the training configuration is consistent with the original configuration and results in an improvement of about 0.2-0.5% AP.
yolov5_s_spos_shufflenetv2 achieves 38.0% AP with only 7.042M parameters, directly instead of the backbone, and outperforms yolov5_s with a similar size by more than 0.3% AP.
With the efficient backbone of yolov6_l_attentivenas_a6, the input channels of YOLOv6RepPAFPN are reduced. Meanwhile, modify the deepen_factor and the neck is made deeper to restore the AP.
with the rtmdet_tiny_ofa_lat31 backbone with only 3.315M parameters and 3.634G flops, we can modify the input resolution to 960, with a similar model size compared to rtmdet_tiny and exceeds rtmdet_tiny by 0.4% AP, reducing the size of the whole model to 3.91 MB.