surfing2003 / image-classification-level1-35

image-classification-level1-35 created by GitHub Classroom

Mask image classification task

Team: 7Features (level1-35)

Contents

1. Requirements
2. Project files
3. Preprocessing
4. Train

Requirements

torch == 1.9.0+cu102
torchvision == 0.10.0+cu102
timm == 0.42.12
albumentations == 1.0.3
sklearn == 0.24.2
cv2 == 4.5.3
PIL == 8.1.0
pandas_streaming == 0.2.175
numpy == 1.19.2
pandas == 1.1.5
facenet_pytorch
retina-face

Project files

• main.py
• train.py - function for train
• preprocessing.py - make cropped image wtih facenet, retina-face
• dataset.py - class and function for data load
• model.py - models for training
• optimizer.py - optimizer for training
• Loss.py - loss for training
• utill.py - Defining functions necessary for the overall process
• config.ini - Setting the necessary parameters for the overall learning process

Preprocessing

• Using facenet and retina-face
• Crop only the human face

# Change values to suit your situation.
train_path = '/opt/ml/input/data/eval/info.csv'
img_path = '/opt/ml/input/data/eval/images' 
train_data = True
PATH = '/opt/ml/input/data/train/cropped_images' if train_data else '/opt/ml/input/data/eval/cropped_images'

# Execution
python preprocessing.py

Train

1. config.ini setting

[augmentation]
; augmentation do or not
augmentation = 0
; augmentation data use or not
load_augmentation = 0
; # of augmentation
aug_num = 4
; target class of augmentation
aug_targets = [8, 11, 14, 17]

; path setting
[path]
train_csv_path = ../../input/data/train/train.csv
train_images_path = ../../input/data/train/cropped_images
model_save_path = ../../ ; Create a "models" folder in that location
result_save_path = ../../ ; Create a "results" folder in that location
save_name = ;<model name>

; wandb setting
[wandb]
wandb_group_name = ;<group name>
wandb_name = ;<run name>
wandb_entity = ;<userid>
wandb_project_name = ;<projectname>

[training]
optimizer = AdamW
scheduler = CosineAnnealingLR
loss = Crossentropy_focal_labelsmoothing
loss1_weight = 0.9
loss2_weight = 0.1
loss3_weight = 0.3
model_name = swsl_resnext50_32x4d
early_stopping = 5
k_fold_num = 5
epoches = 100
lr = 1e-4
batch_size = 32
prediction_type = Age ; Mask, Age, Gender, All
;(E.g. prediction_type: Age -> A model that judges Age [<30, 30<=&<60, 60<=])
learning_type = All ; Mask, None, All 
;(E.g. prediction_type: Age, learning_type: Mask -> A model that judges only the image with a mask in Age)
num_classes = 3 ; # of discriminant classes according to model

2. main execution

• -c is mandatory

main.py -c <config file path>
E.g. main.py -c config.ini

Inference

• -m : your models location
• -d : info.csv location
• -s : save location (Basically, current folder, name is results)
• -t is option: None or Fold

inference.py -m <model path> -d <file path> -s <save path> -t <inference_type>