LG_Camera_Competition

Summary

Data Path

​```
${LG_Folder}
├── train.py
├── inference.py
├── preprocess.py
├── utils.py
├── dataloader.py
|
├── saved_models
|   └── 46e_31.9065_s.pth
|   └── 41e_32.1909_s.pth
|
├── submission
|   └── submission.zip
|
├── data
|   └── train_input_img
|   └── train_label_img
|   └── test_input_img
|   └── train.csv
|   └── test.csv
| 
├── img
|   └── augmented_img_xx.png
|   └── original_img_xx.png
|   
└── environment.yml
​```

Environments Settings

• CUDA version >= 11.1

• Ubuntu 18.04

$ conda env create -n lg --file environment.yml
$ conda activate lg
$ pip install git+https://github.com/ildoonet/pytorch-gradual-warmup-lr.git

Preprocess Script

$ python preprocess.py

Training Script

$ python train.py --gpu=0,1 --img_size=512 --batch_size=64 --exp_name=512_models
$ python train.py --gpu=0,1 --img_size=768 --batch_size=32 --exp_name=768_models

optional arguments:
    --img_size:           size image input, default 512
    --num_workers:        num workers of dataloader, default 8
    --encoder_type:       backbone model, default 'se_resnext50_32x4d'
    --decoder_type:       decoder model, default 'Unet'
    --scheduler:          scheduler type, default 'GradualWarmupSchedulerV2'
    --encoder_lr:         learning rate of encoder, default 3e-5
    --min_lr:             minimum learning rate, default 1e-6
    --batch_size:         batch size training, default 32
    --weight_decay:       weight decay, default 1e-6
    --amp:                use apex, default True  
    --gpu:                gpu numbers
    --exp_name:           experiment name  

Inference Script

$ python inference.py --gpu=0,1

optional arguments:
    --gpu:                gpu numbers

Make augmented images Script

$ python sample_augmentation.py

Result

	실험	CV	Public	Private
v1	Baseline (Input Resolution 512)	31.52	30.56	31.12
v2	+Aug	31.90	30.93	31.07
v3	+Aug + AdamP	32.04	31.06	31.27
v4	+Aug + AdamP + Loss	32.19	30.93	31.41
v5	+Aug + AdamP + Loss + Inference Resolution (512 -> 768)	32.29	31.44	31.69
v6	+Aug + AdamP + Loss + Inference Resolution (768 -> 1024)	31.98	31.88	31.49
v7	Ensemble (512model - v5 + 768model – v6)	32.49	31.80 (+-0.08)	31.80 (+-0.08)

LeaderBoard