- My column
- Introduction
- All task training results
- Environments
- Download my pretrained models and experiments records
- Prepare datasets
- How to train and test model
- How to use gradio demo
- Citation
https://www.zhihu.com/column/c_1692623656205897728
This repository provides simple training and testing examples for the following tasks:
| task | support dataset | support network |
|---|---|---|
| Image classification task | CIFAR100 ImageNet1K(ILSVRC2012) ImageNet21K(Winter 2021 release) ACCV2022 |
ResNet DarkNet RepVGG RegNetX ViT VAN |
| Object detection task | VOC2007 and VOC2012 COCO2017 Objects365(v2,2020) |
RetinaNet FCOS CenterNet TTFNet DETR DINO-DETR |
| Semantic segmentation task | ADE20K | DeepLabv3+ U2Net |
| Instance segmentation task | COCO2017 | YOLACT SOLOv2 |
| Knowledge distillation task | ImageNet1K(ILSVRC2012) | KD loss(for ResNet) DML loss(for ResNet) |
| Contrastive learning task | ImageNet1K(ILSVRC2012) | DINO(for ResNet) |
| Masked image modeling task | ImageNet1K(ILSVRC2012) ACCV2022 |
MAE(for ViT) |
| OCR text detection task | ICDAR | DBNet |
| OCR text recognition task | / | CTC Model |
| Image inpainting task | CelebA-HQ Places365-standard Places365-challenge |
AOT-GAN TRANSX-LKA-AOT-GAN |
| diffusion model task | CIFAR10 CIFAR100 CelebA-HQ FFHQ |
DDPM DDIM PLMS |
See all task training results in results.md.
1、This repository only supports running on ubuntu(verison>=18.04 LTS).
2、This repository only support one machine one gpu/one machine multi gpus mode with pytorch DDP training.
3、Please make sure your Python environment version>=3.7.
4、Please make sure your pytorch version>=1.10.
5、If you want to use torch.complie() function,please make sure your pytorch version>=2.0.
Use pip or conda to install those Packages in your Python environment:
torch
torchvision
pillow
numpy
Cython
colormath
pycocotools
opencv-python
scipy
eniops
scikit-image
pyclipper
shapely
imagesize
nltk
tqdm
onnx
onnx-simplifier
thop==0.0.31.post2005241907
gradio==3.32.0
yapf
If you want to use dino-detr model,install MultiScaleDeformableAttention Packge in your Python environment:
cd to simpleAICV/detection/compile_multiscale_deformable_attention,then run commands:
chmod +x make.sh
./make.sh
You can download all my pretrained models and all my experiments records/checkpoints from huggingface or Baidu-Netdisk.
If you only want to download all my pretrained models(model.state_dict()),you can download pretrained_models folder.
# huggingface
https://huggingface.co/zgcr654321/classification_training/tree/main
https://huggingface.co/zgcr654321/contrastive_learning_training/tree/main
https://huggingface.co/zgcr654321/detection_training/tree/main
https://huggingface.co/zgcr654321/image_inpainting_training/tree/main
https://huggingface.co/zgcr654321/diffusion_model_training/tree/main
https://huggingface.co/zgcr654321/distillation_training/tree/main
https://huggingface.co/zgcr654321/instance_segmentation_training/tree/main
https://huggingface.co/zgcr654321/masked_image_modeling_training/tree/main
https://huggingface.co/zgcr654321/ocr_text_detection_training/tree/main
https://huggingface.co/zgcr654321/ocr_text_recognition_training/tree/main
https://huggingface.co/zgcr654321/semantic_segmentation_training/tree/main
https://huggingface.co/zgcr654321/pretrained_models/tree/main
# Baidu-Netdisk
链接：https://pan.baidu.com/s/1yhEwaZhrb2NZRpJ5eEqHBw
提取码：rgdo
Make sure the folder architecture as follows:
CIFAR10
|
|-----batches.meta unzip from cifar-10-python.tar.gz
|-----data_batch_1 unzip from cifar-10-python.tar.gz
|-----data_batch_2 unzip from cifar-10-python.tar.gz
|-----data_batch_3 unzip from cifar-10-python.tar.gz
|-----data_batch_4 unzip from cifar-10-python.tar.gz
|-----data_batch_5 unzip from cifar-10-python.tar.gz
|-----readme.html unzip from cifar-10-python.tar.gz
|-----test_batch unzip from cifar-10-python.tar.gz
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
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.
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.
Make sure the folder architecture as follows:
ACCV2022
|
|-----train-------------5000 sub classes folders
|-----testa-------------60000 images
|-----accv2022_broken_list.json
Make sure the folder architecture as follows:
VOCdataset
| |----Annotations
| |----ImageSets
|----VOC2007------|----JPEGImages
| |----SegmentationClass
| |----SegmentationObject
|
| |----Annotations
| |----ImageSets
|----VOC2012------|----JPEGImages
| |----SegmentationClass
| |----SegmentationObject
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
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
Make sure the folder architecture as follows:
ADE20K
| |----training
|---images--------|----validation
| |----testing
|
| |----training
|---annotations---|----validation
Make sure the folder architecture as follows:
CelebA-HQ
| |----female
|---train---------|----male
|
| |----female
|---val-----------|----male
Make sure the folder architecture as follows:
FFHQ
|
|---images
|---ffhq-dataset-v1.json
|---ffhq-dataset-v2.json
Make sure the folder architecture as follows:
Places365-standard/challenge
|
| |---train_large all sub folders
|---high_resolution_images---|---val_large all images
| |---test_large all images
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 classification_training/imagenet/resnet50.
If you want to restart train this model,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 ./
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 ./
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 using gpu cards.Make sure master_addr/master_port are unique for each training.
All checkpoints/log are saved in training/testing experiment folder directory.
Also, You can modify super parameters in train_config.py/test_config.py.
cd to gradio_demo,we have classification/detection/semantic_segmentation/instance_segmentation demo.
For example,you can run detection gradio demo:
chmod +x run_gradio_detect_single_image.sh
./run_gradio_detect_single_image.sh
If you find my work useful in your research, please consider citing:
@inproceedings{zgcr,
title={SimpleAICV-pytorch-training-examples},
author={zgcr},
year={2020-2024}
}