SasonChan / image_classification_pytorch

Pytorch实践简单， 方便， 快速训练 图像分类模型

Image Classification Tool

---

As a computer vision engineer, it was designed for my daily work

This project will be updated occasionally.

Update

2021.10.15

1. add transform setting in config.yaml
2. add new arguments parser, and function update_cfg

2021.10.13

1. reconstruct whole code structure

2021.2.6

1. add labelsmooth flag in config.py and train.py and utils/loss.py

2021.1.7

1. preprocess.py add new column now class_name
2. add test_tta.py (not finished), train_.py(support different validation)

2021.1.5

1. add ScoreCam.pyto visualize CNN layer
2. add confusion_matrix to train.py, add new argument --cal_mtx default is True

2021.1.4

1. modified the model/models, to change the Classifier
2. modified the train.py
3. modified the test.py

2020.12.31

1. update preprocess_data for easily produce csv file
2. update test.py for easily doing inference

2020.12.29 upload to github

supported models

2020.12.29

1. ResNet all models
2. ResNest all models
3. efficientNet B4

---

How to train

1. data prepare

1. set you image data be like :

data
  | --- task_name
            |---train
                  |---class1
                        |---- img1.jpg
                        |---- img2.jpg
                        |---- ....
                  |---class2
                  |---...
                    
            |--- val
                  |---class1
                        |---- img1.jpg
                        |---- img2.jpg
                        |---- ....
                  |---class2
                  |---...

2.using data_load/img2csv.py to create train and val csv file

#create train csv
python data_load/img2csv.py --img_path data/your_task_name/train/ --classes class1,class2... --csv data/your_task_name_train.csv
#create vaL csv
python data_load/img2csv.py --img_path data/your_task_name/ --classes class1,class2... --csv data/your_task_name_val.csv

2. start training

you can copy a default cfg to modify it in experiments dir:

config.yaml remember to set None if you do not want to any of img transforms

fold_num: 5
seed: 719
model_arch: resnet50
img_size: 224 
epochs: 100
loss: labelsmooth  #BCEwithlogits, labelsmooth, c
train_batchsize: 256
valid_batchsize: 128
test_batchsize: 128 #for test tt
T_0: 10  #for cosine lr schedule
lr: 0.001
min_lr: 0.00001
weight_decay: 0.00001
num_workers: 4
accum_iter: 2 # suppoprt to do batch accumulation for backprop with effectively larger batch siz
verbose_step: 1
tta : 3
valid_every_x_epoch: 5
default_save_path: output_model
gpus : 0
# 'model_arch': "tf_efficientnet_b4_ns"
# 'model_arch': 'RepVGG-B1g2'

#img augmentation
train_aug:
  HorizontalFlip : [0.5]
  VerticalFlip : [0.5]
  HueSaturationValue : [0.2, 0.2, 0.2, 0.5]
  ShiftScaleRotate : [0.5]
  CoarseDropout : [0.5]
  Cutout : [0.5]
  MotionBlur : [3, 0.5]
val_aug:
  HorizontalFlip : None
  VerticalFlip : None
  HueSaturationValue : None
  ShiftScaleRotate : None
  CoarseDropout : [0.5]
  Cutout : [0.5]
tta_aug:
  HorizontalFlip : None
  VerticalFlip : None
  HueSaturationValue : None
  ShiftScaleRotate : None
  CoarseDropout : [0.5]
  Cutout : [0.5]

train commnad example

python train.py --tpath data/yourtask/train/ --tcsv data/yourtask_train.csv --vpath data/yourtask/val/ --vcsv data/yourtask_val.csv --cfg experiment/config_task.yaml --bsize 128 --gpus '0,1,2,3,4,5,6,7'

• -bsize : training-batchsize
• -cfg : your config yaml
• -gpus : set gpu be like '0,1,2,3' ps.start from 0

3. after training

1. training confusion matrix
2. trained weight
3. tensorboard log

above will be save into output_model directory by date_modelname

---

Inference

not yet update

---

ScoreCam

not yet update