加载yolo-s.pth预训练模型后训练coco，精度一直在下降

[zhanghaochen-817]

Cuda            = True
seed            = 11
distributed     = False
sync_bn         = False
fp16            = False
classes_path    = 'data/classes/coco_classes.txt'
model_path = 'checkpoint/yolox_s.pth'
input_shape     = [416, 416]
phi             = 's'
mosaic              = True
mosaic_prob         = 0.5
mixup               = True
mixup_prob          = 0.5
special_aug_ratio   = 0.7
Init_Epoch          = 0
Freeze_Epoch        = 50
Freeze_batch_size   = 16
UnFreeze_Epoch      = 300
Unfreeze_batch_size = 16
Freeze_Train        = True
Init_lr             = 1e-5
Min_lr              = Init_lr * 0.01
optimizer_type      = "sgd"
momentum            = 0.937
weight_decay        = 5e-4
lr_decay_type       = "cos"
save_period         = 10
save_dir            = 'logs'
eval_flag           = True
eval_period         = 2
num_workers         = 0

以上是我的参数设定，最终的初始lr为5e-4。
训练图