SimpleAICV

Open-source / Simple / Lightweight / Easy-to-use / Extensible

---

• 📢 News!
• Simplicity
• Introduction
• All task training results
• Training GPU server
• Environments
• Download pretrained models and experiments checkpoints/logs
• Download datasets
• How to use gradio demo
• How to use inference demo
• How to train or test model
• Prepare datasets
  • CIFAR100
  • ImageNet 1K(ILSVRC2012)
  • ImageNet 21K(Winter 2021 release)
  • COCO2017
  • Objects365(v2,2020)
  • VOC2007&VOC2012
  • ADE20K
  • SAMACOCO
• Reference
• My column
• Citation

📢 News!

• 2026/02/02: update dinov3 backbone implementation in SimpleAICV/detection/models/backbones.
• 2026/02/02: update SAM(segment_anything)/SAM_Matting model training pipeline and jupyter example in 13.interactive_segmentation_training.
• 2026/02/02: update SAM2(segment_anything2)/SAM2_Matting model training pipeline and jupyter example in 14.video_interactive_segmentation_training.
• 2026/02/02: update universal_segmentation/universal_matting model training pipeline in 16.universal_segmentation_training.
• 2026/02/02: updata all task gradio demo in gradio_demo.
• 2026/02/02: updata all task inference demo in inference_demo.

Simplicity

This repository maintains a lightweight codebase.It requiring only Python and PyTorch as core dependencies(no third-party frameworks like MMCV).

Introduction

This repository provides simple training and testing examples for following tasks:

task	support dataset	support model
00.classification_training	CIFAR100 ImageNet1K(ILSVRC2012) ImageNet21K(Winter 2021 release)	DarkNet ResNet Convformer VAN ViT
01.distillation_training	ImageNet1K(ILSVRC2012)	DML loss(ResNet) KD loss(ResNet)
02.masked_image_modeling_training	ImageNet1K(ILSVRC2012)	MAE(ViT)
03.detection_training	COCO2017 Objects365(v2,2020) VOC2007&VOC2012	RetinaNet FCOS DETR
04.semantic_segmentation_training	ADE20K COCO2017	pfan_semantic_segmentation
05.instance_segmentation_training	COCO2017	SOLOv2 YOLACT
06.salient_object_detection_training	combine dataset	pfan_segmentation
07.human_matting_training	combine dataset	pfan_matting
08.ocr_text_detection_training	combine dataset	DBNet
09.ocr_text_recognition_training	combine dataset	CTC_Model
10.face_detection_training	combine dataset	RetinaFace
11.face_parsing_training	CelebAMask-HQ FaceSynthetics	pfan_face_parsing
12.human_parsing_training	CIHP LIP	pfan_human_parsing
13.interactive_segmentation_training	combine dataset	SAM(segment_anything) SAM_Matting
14.video_interactive_segmentation_training	combine dataset	SAM2(segment_anything2) SAM2_Matting
16.universal_segmentation_training	combine dataset	universal_segmentation universal_matting

All task training results

See all task training results in RESULTS.md.

Training GPU server

1、1-8 RTX 4090D(24GB) GPUs, Python3.12, Pytorch2.5.1, CUDA12.4, Ubuntu22.04(for most experiments).

2、8 RTX PRO 6000(96GB) GPUs, Python3.12, Pytorch2.8.0, CUDA12.8, Ubuntu22.04(for 13.interactive_segmentation_training/14.video_interactive_segmentation_training/16.universal_segmentation_training).

Environments

1、Python and Pytorch Supported Version: Python>=3.12, Pytorch>=2.5.1.

2、Most Experiments only support Single-Node Single-GPU training/Single-Node Multi-GPU DDP training, but 13.interactive_segmentation_training/14.video_interactive_segmentation_training also support Multi-Node Multi-GPU DDP training(Requires InfiniBand/RoCE).

3、Create a conda environment:

conda create -n SimpleAICV python=3.12

4、Install PyTorch:

conda install pytorch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 pytorch-cuda=12.4 -c pytorch -c nvidia

To install a different PyTorch version, find command from here:

https://pytorch.org/get-started/previous-versions/

5、Install other Packages:

pip install -r requirements.txt

Download pretrained models and experiments checkpoints/logs

You can download all my pretrained models and experiments checkpoints/logs from huggingface or Baidu-Netdisk.

If you only need the pretrained models (model.state_dict()), you can download the pretrained_models folder.

# huggingface
https://huggingface.co/zgcr654321/00.classification_training/tree/main
https://huggingface.co/zgcr654321/01.distillation_training/tree/main
https://huggingface.co/zgcr654321/02.masked_image_modeling_training/tree/main
https://huggingface.co/zgcr654321/03.detection_training/tree/main
https://huggingface.co/zgcr654321/04.semantic_segmentation_training/tree/main
https://huggingface.co/zgcr654321/05.instance_segmentation_training/tree/main
https://huggingface.co/zgcr654321/06.salient_object_detection_training/tree/main
https://huggingface.co/zgcr654321/07.human_matting_training/tree/main
https://huggingface.co/zgcr654321/08.ocr_text_detection_training/tree/main
https://huggingface.co/zgcr654321/09.ocr_text_recognition_training/tree/main
https://huggingface.co/zgcr654321/10.face_detection_training/tree/main
https://huggingface.co/zgcr654321/11.face_parsing_training/tree/main
https://huggingface.co/zgcr654321/12.human_parsing_training/tree/main
https://huggingface.co/zgcr654321/13.interactive_segmentation_training/tree/main
https://huggingface.co/zgcr654321/14.video_interactive_segmentation_training/tree/main
https://huggingface.co/zgcr654321/16.universal_segmentation_training/tree/main
https://huggingface.co/zgcr654321/pretrained_models/tree/main

# Baidu-Netdisk
链接:https://pan.baidu.com/s/17oSFXgIy1vxUdPUhTzRkdw?pwd=3l99
提取码：3l99

Download datasets

You can download all datasets from Baidu-Netdisk.

# Baidu-Netdisk
链接: https://pan.baidu.com/s/1zjwdVNliOMS3xwuuY41gcA?pwd=z9sa 
提取码: z9sa

How to use gradio demo

cd to gradio_demo folder,we have:

00.gradio_classify_single_image.py
03.gradio_detect_single_image.py
04.gradio_semantic_segment_single_image.py
05.gradio_instance_segment_single_image.py
06.gradio_salient_object_detection_single_image.py
07.gradio_human_matting_single_image.py
08.gradio_ocr_text_detect_single_image.py
09.gradio_ocr_text_recognition_single_image.py
10.gradio_face_detect_single_image.py
11.gradio_face_parsing_single_image.py
12.gradio_human_parsing_single_image.py
13.0.0.gradio_sam_point_target_single_image.py
13.0.1.gradio_sam_circle_target_single_image.py
16.0.gradio_universal_segment_single_image.py
16.1.gradio_universal_matting_single_image.py

For example,you can run 03.gradio_detect_single_image.py(please prepare pretrained model weight first and modify pretrained model load path):

python 03.gradio_detect_single_image.py

How to use inference demo

cd to inference_demo folder,we have:

00.inference_classify_single_image.py
03.inference_detect_single_image.py
04.inference_semantic_segment_single_image.py
05.inference_instance_segment_single_image.py
06.inference_salient_object_detection_single_image.py
07.inference_human_matting_single_image.py
08.inference_ocr_text_detect_single_image.py
09.inference_ocr_text_recognition_single_image.py
10.inference_face_detect_single_image.py
11.inference_face_parsing_single_image.py
12.inference_human_parsing_single_image.py
13.0.inference_sam_single_image.py
16.0.inference_universal_segment_single_image.py
16.1.inference_universal_matting_single_image.py

For example,you can run 03.inference_detect_single_image.py(please prepare pretrained model weight first and modify pretrained model load path):

python 03.inference_detect_single_image.py

How to train or test model

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

For example,you can enter in folder 00.classification_training/imagenet/resnet50.

If you want to train model from scratch,please delete checkpoints and log folders first,then run train.sh:

CUDA_VISIBLE_DEVICES=0,1 torchrun \
    --nproc_per_node=2 \
    --master_addr 127.0.1.0 \
    --master_port 10000 \
    ../../../tools/train_classification_model.py \
    --work-dir ./

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

CUDA_VISIBLE_DEVICES=0 torchrun \
    --nproc_per_node=1 \
    --master_addr 127.0.1.1 \
    --master_port 10001 \
    ../../../tools/test_classification_model.py \
    --work-dir ./

CUDA_VISIBLE_DEVICES is used to specify gpu_ids for training.Please make sure the number of nproc_per_node equal to the number of gpus.Make sure master_addr/master_port are unique for each training.

Checkpoints/log folders are saved in your training/testing experiment folder directory.Also, You can modify super parameters in train_config.py/test_config.py.

Prepare datasets

CIFAR100

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

ImageNet 1K(ILSVRC2012)

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.

ImageNet 21K(Winter 2021 release)

Make sure the folder architecture as follows:

ImageNet21K
|
|-----train-----------10450 sub classes folders
|-----val-------------10450 sub classes folders
|-----small_classes---10450 sub classes folders
|-----imagenet21k_miil_tree.pth
Please make sure the same class has same class folder name in train and val folders.

COCO2017

Make sure the folder architecture as follows:

COCO2017
|                |----captions_train2017.json
|                |----captions_val2017.json
|--annotations---|----instances_train2017.json
|                |----instances_val2017.json
|                |----person_keypoints_train2017.json
|                |----person_keypoints_val2017.json
|                 
|                |----train2017
|----images------|----val2017

Objects365(v2,2020)

Make sure the folder architecture as follows:

objects365_2020
|
|                |----zhiyuan_objv2_train.json
|--annotations---|----zhiyuan_objv2_val.json
|                |----sample_2020.json
|                 
|                |----train all train patch folders
|----images------|----val   all val patch folders
                 |----test  all test patch folders

VOC2007&VOC2012

Make sure the folder architecture as follows:

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

ADE20K

Make sure the folder architecture as follows:

ADE20K
|                 |----training
|---images--------|----validation
|                 |----testing
|        
|                 |----training
|---annotations---|----validation

SAMACOCO

Make sure the folder architecture as follows:

SAMA-COCO
|                |----sama_coco_train.json
|                |----sama_coco_validation.json
|--annotations---|----train_labels.json
|                |----validation_labels.json
|                |----test_labels.json
|                |----image_info_test2017.json
|                |----image_info_test-dev2017.json
|                 
|                |----train
|----images------|----validation

Reference

https://github.com/facebookresearch/dinov3
https://github.com/facebookresearch/segment-anything
https://github.com/facebookresearch/sam2
https://github.com/tue-mps/EoMT

My column

https://www.zhihu.com/column/c_1692623656205897728

Citation

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

@inproceedings{zgcr,
 title={SimpleAICV-pytorch-training-examples},
 author={zgcr},
 year={2020-2030}
}