hukai97 / simpleaicv-pytorch-imagenet-coco-training

SimpleAICV:pytorch training example on ImageNet(ILSVRC2012)/COCO2017/VOC2007+2012 datasets.Include ResNet/DarkNet/RetinaNet/FCOS/CenterNet/TTFNet/YOLOv3/YOLOv4/YOLOv5/YOLOX.

• My ZhiHu column
• Environments
• Prepare datasets
• Download my pretrained models
• Train and test model
• Classification training results
  • ILSVRC2012(ImageNet) training results
  • CIFAR100 training results
• Detection training results
  • COCO2017 training results
  • VOC2007 and VOC2012 training results
• Distillation training results
  • ImageNet training results
• Citation

My ZhiHu column

https://www.zhihu.com/column/c_1249719688055193600

Environments

This repository only support one server one gpu card/one server multi gpu cards.

environments: Ubuntu 20.04.3 LTS,30 core AMD EPYC 7543 32-Core Processor, 2*RTX A5000, Python Version:3.8, CUDA Version:11.3

Please make sure your Python version>=3.7. Use pip or conda to install those Packages:

torch==1.10.0
torchvision==0.11.1
torchaudio==0.10.0
onnx==1.11.0
onnx-simplifier==0.3.6
numpy
Cython
pycocotools
opencv-python
tqdm
thop
yapf
apex

How to install apex?

apex needs to be installed separately.For torch1.10,modify apex/apex/amp/utils.py:

if cached_x.grad_fn.next_functions[1][0].variable is not x:

to

if cached_x.grad_fn.next_functions[0][0].variable is not x:

Then use the following orders to install apex:

git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --no-cache-dir ./

Using apex to train can reduce video memory usage by 25%-30%, but the training speed will be slower, the trained model has the same performance as not using apex.

Prepare datasets

If you want to reproduce my imagenet pretrained models,you need download ILSVRC2012 dataset,and make sure the folder architecture as follows:

ILSVRC2012
|
|-----train----1000 sub classes folders
|-----val------1000 sub classes folders
Please make sure the same class has same class folder name in train and val folders.

If you want to reproduce my cifar100 pretrained models,you need download cifar100 dataset,and make sure the folder architecture as follows:

CIFAR100
|
|-----train unzip from cifar-100-python.tar.gz
|-----test  unzip from cifar-100-python.tar.gz
|-----meta  unzip from cifar-100-python.tar.gz

If you want to reproduce my COCO pretrained models,you need download COCO2017 dataset,and make sure the folder architecture as follows:

COCO2017
|
|-----annotations----all .json file (label file)
|                 
|                |----train2017
|----images------|----val2017

If you want to reproduce my VOC pretrained models,you need download VOC2007+VOC2012 dataset,and make sure the folder architecture as follows:

VOCdataset
|                 |----Annotations
|                 |----ImageSets
|----VOC2007------|----JPEGImages
|                 |----SegmentationClass
|                 |----SegmentationObject
|        
|                 |----Annotations
|                 |----ImageSets
|----VOC2012------|----JPEGImages
|                 |----SegmentationClass
|                 |----SegmentationObject

Download my pretrained models

You can download all my pretrained models from google drive or BAIDUWANGPAN:

https://drive.google.com/drive/folders/1oif1oma3BvJ54bEB_487U8mmbToNI4Jh?usp=sharing

链接：https://pan.baidu.com/s/1IN81YQWkfVGq2bg6IhFztw 
提取码：ruzk

Train and test model

If you want to train or test model,you need enter a training folder directory,then run train.sh and test.sh.

For example,you can enter classification_training/imagenet/resnet50. If you want to train this model from scratch,please delete checkpoints and log folders first,then run train.sh:

CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.run --nproc_per_node=2 --master_addr 127.0.1.0 --master_port 10000 ../../../tools/train_classification_model.py --work-dir ./

CUDA_VISIBLE_DEVICES is used to specify the gpu ids for this training.Please make sure the number of nproc_per_node equal to the number of gpu cards. Make sure master_addr/master_port are unique for each training.

if you want to test this model,you need have a pretrained model first,modify trained_model_path in test_config.py,then run test.sh:

CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.run --nproc_per_node=2 --master_addr 127.0.1.0 --master_port 10000 ../../../tools/test_classification_model.py --work-dir ./

Also, You can modify super parameters in train_config.py/test_config.py.

Classification training results

ILSVRC2012(ImageNet) training results

Network	macs	params	input size	gpu num	batch	warm up	lr decay	apex	syncbn	epochs	Top-1
ResNet18	1.819G	11.690M	224x224	2 RTX A5000	256	0	multistep	True	False	100	70.490
ResNet34half	949.323M	5.585M	224x224	2 RTX A5000	256	0	multistep	True	False	100	67.690
ResNet34	3.671G	21.798M	224x224	2 RTX A5000	256	0	multistep	True	False	100	73.950
ResNet50half	1.063G	6.918M	224x224	2 RTX A5000	256	0	multistep	True	False	100	72.048
ResNet50	4.112G	25.557M	224x224	2 RTX A5000	256	0	multistep	True	False	100	76.334
ResNet101	7.834G	44.549M	224x224	2 RTX A5000	256	0	multistep	True	False	100	77.716
ResNet152	11.559G	60.193M	224x224	2 RTX A5000	256	0	multistep	True	False	100	78.318
ResNet50-200epoch	4.112G	25.557M	224x224	2 RTX A5000	256	5	cosinelr	True	False	200	77.326
ResNet50-autoaugment	4.112G	25.557M	224x224	2 RTX A5000	256	5	cosinelr	True	False	200	77.692
ResNet50-randaugment	4.112G	25.557M	224x224	2 RTX A5000	256	5	cosinelr	True	False	200	77.578
DarkNetTiny	412.537M	2.087M	256x256	2 RTX A5000	256	0	multistep	True	False	100	54.720
DarkNet19	3.663G	20.842M	256x256	2 RTX A5000	256	0	multistep	True	False	100	73.830
DarkNet53	9.322G	41.610M	256x256	2 RTX A5000	256	0	multistep	True	False	100	76.796
Yolov4CspDarkNetTiny	977.589M	4.143M	256x256	2 RTX A5000	256	0	multistep	True	False	100	64.340
Yolov4CspDarkNet53	6.584G	27.642M	256x256	2 RTX A5000	256	0	multistep	True	False	100	77.418
Yolov5nBackbone	205.613M	937.480K	256x256	2 RTX A5000	256	0	multistep	True	False	100	55.474
Yolov5sBackbone	759.354M	3.225M	256x256	2 RTX A5000	256	0	multistep	True	False	100	66.486
Yolov5mBackbone	2.230G	7.556M	256x256	2 RTX A5000	256	0	multistep	True	False	100	72.090
Yolov5lBackbone	4.932G	14.315M	256x256	2 RTX A5000	256	0	multistep	True	False	100	73.186
Yolov5xBackbone	9.243G	23.961M	256x256	2 RTX A5000	256	0	multistep	True	False	100	73.618
YoloxnBackbone	104.508M	716.968K	256x256	2 RTX A5000	256	0	multistep	True	False	100	57.350
YoloxtBackbone	504.979M	2.757M	256x256	2 RTX A5000	256	0	multistep	True	False	100	66.246
YoloxsBackbone	876.729M	4.726M	256x256	2 RTX A5000	256	0	multistep	True	False	100	69.092
YoloxmBackbone	2.683G	13.122M	256x256	2 RTX A5000	256	0	multistep	True	False	100	72.378
YoloxlBackbone	6.072G	28.101M	256x256	2 RTX A5000	256	0	multistep	True	False	100	73.976
YoloxxBackbone	11.548G	51.583M	256x256	2 RTX A5000	256	0	multistep	True	False	100	74.484

You can find more model training details in classification_training/imagenet/.

CIFAR100 training results

Network	macs	params	input size	gpu num	batch	warm up	lr decay	apex	syncbn	epochs	Top-1
ResNet18Cifar	556.706M	11.220M	32x32	1 RTX A5000	128	0	multistep	True	False	200	78.180
ResNet34halfCifar	291.346M	5.350M	32x32	1 RTX A5000	128	0	multistep	True	False	200	76.690
ResNet34Cifar	1.162G	21.328M	32x32	1 RTX A5000	128	0	multistep	True	False	200	79.310
ResNet50halfCifar	328.447M	5.991M	32x32	1 RTX A5000	128	0	multistep	True	False	200	77.170
ResNet50Cifar	1.305G	23.705M	32x32	1 RTX A5000	128	0	multistep	True	False	200	76.950
ResNet101Cifar	2.520G	42.697M	32x32	1 RTX A5000	128	0	multistep	True	False	200	78.270
ResNet152Cifar	3.737G	58.341M	32x32	1 RTX A5000	128	0	multistep	True	False	200	78.700

You can find more model training details in classification_training/cifar100/.

Detection training results

COCO2017 training results

Trained on COCO2017_train dataset, tested on COCO2017_val dataset.

mAP is IoU=0.5:0.95,area=all,maxDets=100,mAP(COCOeval,stats[0]).

RetinaNet Paper:https://arxiv.org/abs/1708.02002

FCOS Paper:https://arxiv.org/abs/1904.01355

CenterNet Paper:https://arxiv.org/abs/1904.07850

TTFNet Paper:https://arxiv.org/abs/1909.00700

YOLOv3 Paper:https://arxiv.org/abs/1804.02767

YOLOv4 Paper:https://arxiv.org/abs/2004.10934

YOLOv5 Code:https://github.com/ultralytics/yolov5

YOLOX Paper:https://arxiv.org/abs/2107.08430

How to use yolov3 anchor clustering method to generate a set of custom anchors for your own dataset?

I provide a script in simpleAICV/detection/yolov3_anchor_cluster.py,and I give two examples for generate anchors on COCO2017 and VOC2007+2012 datasets.If you want to generate anchors for your dataset,just modify the part of input code,get width and height of all annotaion boxes,then use the script to compute anchors.The anchors size will change with different datasets or different input resizes.

Network	resize-style	input size	macs	params	gpu num	batch	warm up	lr decay	apex	syncbn	epochs	mAP
ResNet50-RetinaNet	RetinaStyle-400	400x667	63.093G	37.969M	2 RTX A5000	32	0	multistep	True	False	13	32.067
ResNet50-RetinaNet	RetinaStyle-800	800x1333	250.069G	37.969M	2 RTX A5000	8	0	multistep	True	False	13	35.647
ResNet50-RetinaNet	YoloStyle-640	640x640	95.558G	37.969M	2 RTX A5000	32	0	multistep	True	False	13	32.971
ResNet50-FCOS	RetinaStyle-400	400x667	54.066G	32.291M	2 RTX A5000	32	0	multistep	True	False	13	34.046
ResNet50-FCOS	RetinaStyle-800	800x1333	214.406G	32.291M	2 RTX A5000	8	0	multistep	True	False	13	37.857
ResNet50-FCOS	YoloStyle-640	640x640	81.943G	32.291M	2 RTX A5000	32	0	multistep	True	False	13	35.055
ResNet18DCN-CenterNet	YoloStyle-512	512x512	14.854G	12.889M	2 RTX A5000	64	0	multistep	True	False	140	27.813
ResNet18DCN-TTFNet-3x	YoloStyle-512	512x512	16.063G	13.737M	2 RTX A5000	64	0	multistep	True	False	39	28.155
ResNet18DCN-TTFNet-70	YoloStyle-512	512x512	16.063G	13.737M	2 RTX A5000	64	0	multistep	True	False	70	29.675

You can find more model training details in detection_training/coco/.

VOC2007 and VOC2012 training results

Trained on VOC2007 trainval dataset + VOC2012 trainval dataset, tested on VOC2007 test dataset.

mAP is IoU=0.50,area=all,maxDets=100,mAP.

Network	resize-style	input size	macs	params	gpu num	batch	warm up	lr decay	apex	syncbn	epochs	mAP
ResNet50-RetinaNet	RetinaStyle-400	400x667	56.093G	36.724M	2 RTX A5000	32	0	multistep	True	False	13	79.804
ResNet50-RetinaNet	YoloStyle-640	640x640	84.947G	36.724M	2 RTX A5000	32	0	multistep	True	False	13	80.565
ResNet50-FCOS	RetinaStyle-400	400x667	53.288G	32.153M	2 RTX A5000	32	0	multistep	True	False	13	79.894
ResNet50-FCOS	YoloStyle-640	640x640	80.764G	32.153M	2 RTX A5000	32	0	multistep	True	False	13	80.510

You can find more model training details in detection_training/voc/.

Distillation training results

ImageNet training results

KD loss Paper:https://arxiv.org/abs/1503.02531

DKD loss Paper:https://arxiv.org/abs/2203.08679

DML loss Paper:https://arxiv.org/abs/1706.00384

Teacher Network	Student Network	method	Freeze Teacher	input size	gpu num	batch	warm up	lr decay	apex	syncbn	epochs	Teacher Top-1	Student Top-1
ResNet34	ResNet18	CE+KD	True	224x224	2 RTX A5000	256	0	multistep	True	False	100	/	71.848
ResNet34	ResNet18	CE+DKD	True	224x224	2 RTX A5000	256	0	multistep	True	False	100	/	71.856
ResNet34	ResNet18	CE+DML	False	224x224	2 RTX A5000	256	0	multistep	True	False	100	74.318	71.678
ResNet152	ResNet50	CE+KD	True	224x224	2 RTX A5000	256	0	multistep	True	False	100	/	76.830
ResNet152	ResNet50	CE+DKD	True	224x224	2 RTX A5000	256	0	multistep	True	False	100	/	77.692
ResNet152	ResNet50	CE+DML	False	224x224	2 RTX A5000	256	0	multistep	True	False	100	79.462	77.618

You can find more model training details in distillation_training/imagenet/.

Citation

If you find my work useful in your research, please consider citing:

@inproceedings{zgcr,
 title={SimpleAICV-ImageNet-CIFAR-COCO-VOC-training},
 author={zgcr},
 year={2022}
}