Skip to content

Latest commit

 

History

History
100 lines (82 loc) · 3.84 KB

README.md

File metadata and controls

100 lines (82 loc) · 3.84 KB
Raw

Pytorch ReXNet with center loss train code

Introduce

This project is Clova AI ReXNet with center loss train code.

How to train?

You have to prepare AffectNet Dataset. Data loader code is focused on AffectNet.

Here is link to download dataset. Download

    $  python -m rexnet train --train_data data/affectnet \
                              --eval_data data/affectnet \
                              --batch_train 256 \
                              --batch_eval 256 \
                              --epochs 400 \
                              --save_epochs 10 \
                              --eval_epochs 5 \

The detail of command-line is as follows:

    usage: rexnet train [-h] --train_data TRAIN_DATA --valid_data VALID_DATA
                [--train_batch_size TRAIN_BATCH_SIZE]
                [--valid_batch_size VALID_BATCH_SIZE]
                [--train_shuffle TRAIN_SHUFFLE]
                [--valid_shuffle VALID_SHUFFLE]
                [--num_workers NUM_WORKERS] [--epochs EPOCHS]
                [--save_epochs SAVE_EPOCHS] [--valid_epochs VALID_EPOCHS]
                [--base_lr BASE_LR] [--lr_min LR_MIN]
                [--lr_decay LR_DECAY] [--warmup_lr_init WARMUP_LR_INIT]
                [--warmup_t WARMUP_T] [--cooldown_epochs COOLDOWN_EPOCHS]
                [--momentum MOMENTUM] [--nesterov NESTEROV]
                [--model_save_path MODEL_SAVE_PATH]
                [--checkpoint_path CHECKPOINT_PATH] [--gpus GPUS]
                [--center_loss CENTER_LOSS]
                [--center_loss_lambda CENTER_LOSS_LAMBDA]
                [--center_loss_alpha CENTER_LOSS_ALPHA]

    optional arguments:
    -h, --help            show this help message and exit

    Dataset:
    --train_data TRAIN_DATA
                            affectnet train data file path
    --valid_data VALID_DATA
                            affectnet valid data file path

    Dataset Config:
    --train_batch_size TRAIN_BATCH_SIZE
                            train batch size
    --valid_batch_size VALID_BATCH_SIZE
                            valid batch size
    --train_shuffle TRAIN_SHUFFLE
                            train data shuffle
    --valid_shuffle VALID_SHUFFLE
                            valid data shuffle
    --num_workers NUM_WORKERS
                            number of workers for data load

    Train Config:
    --epochs EPOCHS       num of total epochs
    --save_epochs SAVE_EPOCHS
                            interval epohcs of saving
    --valid_epochs VALID_EPOCHS
                            interval epochs of valid
    --base_lr BASE_LR     base lr value
    --lr_min LR_MIN       minimum value of lr
    --lr_decay LR_DECAY   lr decay value
    --warmup_lr_init WARMUP_LR_INIT
                            base warmup lr
    --warmup_t WARMUP_T   warmup epochs
    --cooldown_epochs COOLDOWN_EPOCHS
                            cooldown epochs

    Optimizer Config:
    --momentum MOMENTUM   momentum of SGD
    --nesterov NESTEROV

    Saving Config:
    --model_save_path MODEL_SAVE_PATH
                            model save path
    --checkpoint_path CHECKPOINT_PATH
                            checkpoint save path

    Others:
    --gpus GPUS           number of gpu devices
    --center_loss CENTER_LOSS
                            use center loss
    --center_loss_lambda CENTER_LOSS_LAMBDA
                            center loss lambda
    --center_loss_alpha CENTER_LOSS_ALPHA
                            center loss alpha

Contact Info