retinanet_r50_fpn_ssdd.yml

backbone: {'type': 'resnet50',
           'pretrained': True}

neck: {'type': 'fpn',
       'init_method': 'xavier_init',
       'extra_conv_init_method': 'xavier_init'}

head: {'type': 'retinanet',
       'num_stacked': 4,
       'cls_branch_init_method': 'normal_init',
       'reg_branch_init_method': 'normal_init'}

loss: {'cls': {'alpha': 0.25, 'gamma': 2.0},
       'reg': {'type': 'smooth'}}


assigner: {'pos_iou_thr': 0.3,
           'neg_iou_thr': 0.2,
           'min_pos_iou': 0.0,
           'low_quality_match': True}

# warmup settings
warm_up: True
warmup_epoch: 2
warmup_lr: 0.00001

# data settings
dataset: SSDD
classes: ['ship']
image_size: 512
keep_ratio: False
batch_size: 16
augment: False

data_path: '{Data Root Path.}'
output_path: '{Project Root Path.}'

optimizer: adam
lr: 0.0001
epoch: 100
evaluation_train_start: 101
evaluation_val_start: 101
save_interval: 4
val_interval: 4
eval_method: voc
freeze_bn: True
device: [2]

# anchor settings
base_size: 4
ratios: [0.2, 0.5, 1.0, 2.0, 5.0]
#ratios: [0.5, 1.0, 2.0]
scales_per_octave: 3
angle: 0  # opencv version > 4.5.1

rotation_nms_thr: 0.5
score_thr: 0.05

tensorboard: 'tensorboard'
checkpoint: 'checkpoints'
log: 'log.log'