CV_Sementic-Segmentation

Classification Pipeline in Computer Vision (Pytorch)

Environments

Directory

CV_classification
├── datasets
│   ├── __init__.py
│   ├── augmentation.py
│   └── factory.py
├── models
│   ├── __init__.py
│   ├── resnet.py
│   └── loss.py
├── log.py
├── main.py
├── train.py
├── run.sh
├── requirements.txt
├── README.md
└── LICENSE

Pipeline

0. Set seed
1. Make directory to save results
2. Build model
3. Build dataset with augmentations
   • Train dataset
   • Validation dataset (optional)
   • Test dataset
4. Make dataLoader
5. Define optimizer (model parameters)
6. Define loss function
7. Training model
   • Checkpoint model using evaluation on validation dataset
   • Log training history using logging or wandb in save folder
8. Testing model

example

bash run.sh CIFAR10 10

Config Parameters

Wandb 관련 설정

Argument	Description	Default	Possible value
use_wandb	Wandb 사용 여부	True	True,False
use_cm	Confusion metrix 사용 여부	True	True,False
entity	Wandb 엔티티 명	"connect-cv-04"	---
project_name	Wandb 프로젝트 명	"Image_classification_mask"	---
exp_name	실험명	"exp"	---
exp_num	실험 번호	0	---
user_name	실험자	"my_name"	"KDH","KJY","HJH","KDK"

실험 관련 설정

Argument	Description	Default	Possible value
datadir	input 경로	'../input	---
train_file	train csv 이름	"train.csv"	---
valid_file	valid csv 이름	"valid.csv"	---
transform	Transform 목록	['resize','randomrotation', 'totensor', 'normalize']	---
seed	Random seed	223	---
model_name	Model_names	“CustomModel”	---
model_param	Model_names	{pretrained : True, backbone : "resnet18"}	---
num_classes	Class 개수	18	---
batch_size	Batch size	32	---
opt_name	Optimizer	"Adam"	"Adam"
loss	loss 종류	"crossentropy"	"crossentropy","focalloss","f1loss","bceloss","mseloss"
loss_param	loss parm	"미정"	---
lr	learning rate	5e-6	---
lr_sheduler	"Learning rate scheduler	"StepLR"	"StepLR","ReduceLROnPlateau"
lr_sheduler_param	Lr scheduler parameter	"미정"	---
weight_decay	Weight Decay	5e-4	---
epochs	epoch	100	---
savedir	모델 저장 위치	"./checkpoint"	---
grad_accum_steps	---	1	---
mixed_precision	---	"fp16"	---
patience	Early Stopping	100	---

Contributors ✨