RegNet

open-mmlab / mmdetection

Last updated on Feb 23, 2021
Faster R-CNN (RegNetX-3.2GF-FPN, 1x, pytorch)
Memory (M) 4500.0
FLOPs
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, Convolution, FPN, RoIPool, RegNetX
lr sched 1x
Memory (M) 4500.0
SHOW MORE
SHOW LESS
Faster R-CNN (RegNetX-3.2GF-FPN, 2x, pytorch)
Memory (M) 4500.0
FLOPs
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, Convolution, FPN, RoIPool, RegNetX
lr sched 2x
Memory (M) 4500.0
SHOW MORE
SHOW LESS
Faster R-CNN (RegNetX-3.2GF-FPN, 3x, pytorch)
Memory (M) 5000.0
FLOPs
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, Convolution, FPN, RoIPool, RegNetX
lr sched 3x
Memory (M) 5000.0
SHOW MORE
SHOW LESS
Mask R-CNN R-CNN (RegNetX-3.2GF-FPN-DCN-C3-C5, 1x, pytorch)
Memory (M) 5000.0
FLOPs
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, Convolution, Dense Connections, Deformable Convolution, FPN, RegNetX, RoIAlign
lr sched 1x
Memory (M) 5000.0
SHOW MORE
SHOW LESS
Mask R-CNN (RegNetX-12GF-FPN, 1x, pytorch)
Memory (M) 7400.0
FLOPs
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, Convolution, Dense Connections, FPN, RegNetX, RoIAlign
lr sched 1x
Memory (M) 7400.0
SHOW MORE
SHOW LESS
Mask R-CNN (RegNetX-3.2GF-FPN, 1x, pytorch)
Memory (M) 5000.0
FLOPs
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, Convolution, Dense Connections, FPN, RegNetX, RoIAlign
lr sched 1x
Memory (M) 5000.0
SHOW MORE
SHOW LESS
Mask R-CNN (RegNetX-3.2GF-FPN, 3x, pytorch)
Memory (M) 5000.0
FLOPs
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, Convolution, Dense Connections, FPN, RegNetX, RoIAlign
lr sched 3x
Memory (M) 5000.0
SHOW MORE
SHOW LESS
Mask R-CNN (RegNetX-4.0GF-FPN, 1x, pytorch)
Memory (M) 5500.0
FLOPs
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, Convolution, Dense Connections, FPN, RegNetX, RoIAlign
lr sched 1x
Memory (M) 5500.0
SHOW MORE
SHOW LESS
Mask R-CNN (RegNetX-6.4GF-FPN, 1x, pytorch)
Memory (M) 6100.0
FLOPs
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, Convolution, Dense Connections, FPN, RegNetX, RoIAlign
lr sched 1x
Memory (M) 6100.0
SHOW MORE
SHOW LESS
Mask R-CNN (RegNetX-8.0GF-FPN, 1x, pytorch)
Memory (M) 6400.0
FLOPs
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, Convolution, Dense Connections, FPN, RegNetX, RoIAlign
lr sched 1x
Memory (M) 6400.0
SHOW MORE
SHOW LESS
RetinaNet (RegNetX-1.6GF-FPN, 1x, pytorch)
Memory (M) 3300.0
FLOPs
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture RegNetX, FPN, Focal Loss
lr sched 1x
Memory (M) 3300.0
SHOW MORE
SHOW LESS
RetinaNet (RegNetX-3.2GF-FPN, 1x, pytorch)
Memory (M) 4200.0
FLOPs
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture RegNetX, FPN, Focal Loss
lr sched 1x
Memory (M) 4200.0
SHOW MORE
SHOW LESS
RetinaNet (RegNetX-800MF-FPN, 1x, pytorch)
Memory (M) 2500.0
FLOPs
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture RegNetX, FPN, Focal Loss
lr sched 1x
Memory (M) 2500.0
SHOW MORE
SHOW LESS
Faster R-CNN (R-50-FPN, 1x, pytorch)
Memory (M) 4000.0
inference time (s/im) 0.05495
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, Convolution, FPN, RoIPool, ResNet
lr sched 1x
Memory (M) 4000.0
Backbone Layers 50
inference time (s/im) 0.05495
SHOW MORE
SHOW LESS
Mask R-CNN (R-101-FPN, 1x, pytorch)
Memory (M) 6400.0
inference time (s/im) 0.09709
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, Convolution, Dense Connections, FPN, ResNet, RoIAlign
lr sched 1x
Memory (M) 6400.0
Backbone Layers 101
inference time (s/im) 0.09709
SHOW MORE
SHOW LESS
Mask R-CNN (R-50-FPN, 1x, pytorch)
Memory (M) 4400.0
inference time (s/im) 0.08333
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, Convolution, Dense Connections, FPN, ResNet, RoIAlign
lr sched 1x
Memory (M) 4400.0
Backbone Layers 50
inference time (s/im) 0.08333
SHOW MORE
SHOW LESS
Mask R-CNN (X-101-32x4d-FPN, 1x, pytorch)
Memory (M) 7600.0
inference time (s/im) 0.10638
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, ResNeXt, Convolution, Dense Connections, FPN, RoIAlign
lr sched 1x
Memory (M) 7600.0
Backbone Layers 101
inference time (s/im) 0.10638
SHOW MORE
SHOW LESS
RetinaNet (R-50-FPN, 1x, pytorch)
Memory (M) 3800.0
inference time (s/im) 0.06024
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture ResNet, FPN, Focal Loss
lr sched 1x
Memory (M) 3800.0
Backbone Layers 50
inference time (s/im) 0.06024
SHOW MORE
SHOW LESS
README.md

Designing Network Design Spaces

Introduction

[BACKBONE]

We implement RegNetX and RegNetY models in detection systems and provide their first results on Mask R-CNN, Faster R-CNN and RetinaNet.

The pre-trained modles are converted from model zoo of pycls.

@article{radosavovic2020designing,
    title={Designing Network Design Spaces},
    author={Ilija Radosavovic and Raj Prateek Kosaraju and Ross Girshick and Kaiming He and Piotr Dollár},
    year={2020},
    eprint={2003.13678},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

Usage

To use a regnet model, there are two steps to do:

  1. Convert the model to ResNet-style supported by MMDetection
  2. Modify backbone and neck in config accordingly

Convert model

We already prepare models of FLOPs from 400M to 12G in our model zoo.

For more general usage, we also provide script regnet2mmdet.py in the tools directory to convert the key of models pretrained by pycls to ResNet-style checkpoints used in MMDetection.

python -u tools/model_converters/regnet2mmdet.py ${PRETRAIN_PATH} ${STORE_PATH}

This script convert model from PRETRAIN_PATH and store the converted model in STORE_PATH.

Modify config

The users can modify the config's depth of backbone and corresponding keys in arch according to the configs in the pycls model zoo. The parameter in_channels in FPN can be found in the Figure 15 & 16 of the paper (wi in the legend). This directory already provides some configs with their performance, using RegNetX from 800MF to 12GF level. For other pre-trained models or self-implemented regnet models, the users are responsible to check these parameters by themselves.

Note: Although Fig. 15 & 16 also provide w0, wa, wm, group_w, and bot_mul for arch, they are quantized thus inaccurate, using them sometimes produces different backbone that does not match the key in the pre-trained model.

Results

Mask R-CNN

Backbone Style Lr schd Mem (GB) Inf time (fps) box AP mask AP Config Download
R-50-FPN pytorch 1x 4.4 12.0 38.2 34.7 config model | log
RegNetX-3.2GF-FPN pytorch 1x 5.0 40.3 36.6 config model | log
RegNetX-4.0GF-FPN pytorch 1x 5.5 41.5 37.4 config model | log
R-101-FPN pytorch 1x 6.4 10.3 40.0 36.1 config model | log
RegNetX-6.4GF-FPN pytorch 1x 6.1 41.0 37.1 config model | log
X-101-32x4d-FPN pytorch 1x 7.6 9.4 41.9 37.5 config model | log
RegNetX-8.0GF-FPN pytorch 1x 6.4 41.7 37.5 config model | log
RegNetX-12GF-FPN pytorch 1x 7.4 42.2 38 config model | log
RegNetX-3.2GF-FPN-DCN-C3-C5 pytorch 1x 5.0 40.3 36.6 config model | log

Faster R-CNN

Backbone Style Lr schd Mem (GB) Inf time (fps) box AP Config Download
R-50-FPN pytorch 1x 4.0 18.2 37.4 config model | log
RegNetX-3.2GF-FPN pytorch 1x 4.5 39.9 config model | log
RegNetX-3.2GF-FPN pytorch 2x 4.5 41.1 config model | log

RetinaNet

Backbone Style Lr schd Mem (GB) Inf time (fps) box AP Config Download
R-50-FPN pytorch 1x 3.8 16.6 36.5 config model | log
RegNetX-800MF-FPN pytorch 1x 2.5 35.6 config model | log
RegNetX-1.6GF-FPN pytorch 1x 3.3 37.3 config model | log
RegNetX-3.2GF-FPN pytorch 1x 4.2 39.1 config model | log

Pre-trained models

We also train some models with longer schedules and multi-scale training. The users could finetune them for downstream tasks.

Method Backbone Style Lr schd Mem (GB) Inf time (fps) box AP mask AP Config Download
Faster RCNN RegNetX-3.2GF-FPN pytorch 3x 5.0 42.2 - config model | log
Mask RCNN RegNetX-3.2GF-FPN pytorch 3x 5.0 43.1 38.7 config model | log

Notice

  1. The models are trained using a different weight decay, i.e., weight_decay=5e-5 according to the setting in ImageNet training. This brings improvement of at least 0.7 AP absolute but does not improve the model using ResNet-50.
  2. RetinaNets using RegNets are trained with learning rate 0.02 with gradient clip. We find that using learning rate 0.02 could improve the results by at least 0.7 AP absolute and gradient clip is necessary to stabilize the training. However, this does not improve the performance of ResNet-50-FPN RetinaNet.

Results

Object Detection on COCO minival

Object Detection on COCO minival
MODEL BOX AP
Mask R-CNN (RegNetX-3.2GF-FPN, 3x, pytorch) 43.1
Faster R-CNN (RegNetX-3.2GF-FPN, 3x, pytorch) 42.2
Mask R-CNN (RegNetX-12GF-FPN, 1x, pytorch) 42.2
Mask R-CNN (X-101-32x4d-FPN, 1x, pytorch) 41.9
Mask R-CNN (RegNetX-8.0GF-FPN, 1x, pytorch) 41.7
Mask R-CNN (RegNetX-4.0GF-FPN, 1x, pytorch) 41.5
Faster R-CNN (RegNetX-3.2GF-FPN, 2x, pytorch) 41.1
Mask R-CNN (RegNetX-6.4GF-FPN, 1x, pytorch) 41.0
Mask R-CNN R-CNN (RegNetX-3.2GF-FPN-DCN-C3-C5, 1x, pytorch) 40.3
Mask R-CNN (RegNetX-3.2GF-FPN, 1x, pytorch) 40.3
Mask R-CNN (R-101-FPN, 1x, pytorch) 40.0
Faster R-CNN (RegNetX-3.2GF-FPN, 1x, pytorch) 39.9
RetinaNet (RegNetX-3.2GF-FPN, 1x, pytorch) 39.1
Mask R-CNN (R-50-FPN, 1x, pytorch) 38.2
Faster R-CNN (R-50-FPN, 1x, pytorch) 37.4
RetinaNet (RegNetX-1.6GF-FPN, 1x, pytorch) 37.3
RetinaNet (R-50-FPN, 1x, pytorch) 36.5
RetinaNet (RegNetX-800MF-FPN, 1x, pytorch) 35.6
Instance Segmentation on COCO minival
MODEL MASK AP
Mask R-CNN (RegNetX-12GF-FPN, 1x, pytorch) 38.0
Mask R-CNN (RegNetX-8.0GF-FPN, 1x, pytorch) 37.5
Mask R-CNN (X-101-32x4d-FPN, 1x, pytorch) 37.5
Mask R-CNN (RegNetX-4.0GF-FPN, 1x, pytorch) 37.4
Mask R-CNN (RegNetX-6.4GF-FPN, 1x, pytorch) 37.1
Mask R-CNN (RegNetX-3.2GF-FPN, 1x, pytorch) 36.6
Mask R-CNN R-CNN (RegNetX-3.2GF-FPN-DCN-C3-C5, 1x, pytorch) 36.6
Mask R-CNN (R-101-FPN, 1x, pytorch) 36.1
Mask R-CNN (R-50-FPN, 1x, pytorch) 34.7
Contact us on: hello@paperswithcode.com . Papers With Code is a free resource with all data licensed under CC-BY-SA.
Terms Data policy Cookies policy