train.py

from super_gradients.common.object_names import Models
from super_gradients.training import Trainer
from super_gradients.training import models
from super_gradients.training.dataloaders import coco2017_train_yolo_nas, coco2017_val_yolo_nas
from super_gradients.training.losses import PPYoloELoss
from super_gradients.training.metrics import DetectionMetrics_050
from super_gradients.training.models.detection_models.pp_yolo_e import PPYoloEPostPredictionCallback
from super_gradients.training.utils.distributed_training_utils import setup_device

if __name__ == '__main__':
    # Download model
    model = models.get(Models.YOLO_NAS_S, pretrained_weights='coco', num_classes=2)

    # Initialize trainer
    trainer = Trainer(experiment_name='yolo-nas-plastic', ckpt_root_dir='checkpoints/sg_yolonas')

    # Datasets or dataloaders
    train_dataloader = coco2017_train_yolo_nas(
        dataset_params={
            'data_dir': 'dataset',
            'subdir': '.',
            'json_file': '_annotations.coco.json'
        },
        dataloader_params={
            'batch_size': 2,
            'num_workers': 0
        }
    )

    valid_dataloader = coco2017_val_yolo_nas(
        dataset_params={
            'data_dir': 'dataset',
            'subdir': '.',
            'json_file': '_annotations.coco.json'
        },
        dataloader_params={
            'batch_size': 2,
            'num_workers': 0
        }
    )

    # Training parameters
    train_params = {
        'average_best_models'   : True,
        'max_epochs'            : 100,
        'initial_lr'            : 5e-4,
        'warmup_mode'           : 'LinearEpochLRWarmup',
        'warmup_initial_lr'     : 1e-6,
        'lr_warmup_epochs'      : 3,
        'lr_mode'               : 'cosine',
        'cosine_final_lr_ratio' : 0.1,
        'optimizer'             : 'Adam',
        'optimizer_params'      : {
            'weight_decay'      : 0.0001
            },
        'zero_weight_decay_on_bias_and_bn': True,
        'ema'                   : True,
        'ema_params'            : {
            'decay'             : 0.9,
            'decay_type'        : 'threshold'
        },
        'mixed_precision'       : False,
        'loss'                  : PPYoloELoss(
            num_classes=2,
            reg_max=16
        ),
        'valid_metrics_list'    : [
            DetectionMetrics_050(
                score_thres=0.1,
                top_k_predictions=100,
                num_cls=2, # Include background
                normalize_targets=True,
                post_prediction_callback=PPYoloEPostPredictionCallback(
                    score_threshold=0.01,
                    nms_top_k=1000,
                    max_predictions=100,
                    nms_threshold=0.7
                )
            )
        ],
        'metric_to_watch': 'mAP@0.50'
    }

    trainer.train(
        model=model,
        training_params=train_params,
        train_loader=train_dataloader,
        valid_loader=valid_dataloader
    )