How to train the d0 on coco_human_data ? (only for person detection)

[xscjun]

Thanks for your share of the great work. I evaluate the d0 on the coco_human_val ,and get map : 0.432. When I trained the data with the pre-trained model (efficientdet-d0.pth) , after 100 epoch ,only get the map 0.342,and the loss is no longer drop .

I trained with the script first: 
python train.py -c 0 --batch_size 512 --lr 1e-2 --num_workers 24 
        --num_epochs 500  --load_weights efficientdet-d0.pth

From epoch34,the learning rate drop to 1e-3 automatically,   
From epoch 59 ，the loss is no longer drop .then  I changed the lr to 1e-4，  but the loss is still  no longger drop（now 100epoch but still no dropping） 。

How can I change the Hyperparameter， to  get a better loss graph and map  ？ 
change adamw to sgd ? drop the lr ? other parameter ?

[xscjun]

I have read from other issue " training on coco takes time and requires you to change parameters frequently. " Can you share you training parameters and how to change them during the training,please ？

[zylo117]

You should train with --head_only, and reduce lr. You can also monitor loss graph in tensorboard. If your mAP drops, it's very likely that you are suffering overfitting

[xscjun]

@zylo117 Yes ,I can reach i higher map(0.411) with --head_only. I just want to know how to try train the backbone . my loss graph : (I think it's overfitting from 48k )

Second time I change the lr to 1e-4 when 48k , and the loss graph like that :

How should I change the parameters ,please ?

[zylo117]

backbone training is not necessary unless no matter what you do, the mAP can't get any higher.

but, I think by reducing lr, the mAP can be a little bit higher.

Person is one of the hard objects in Coco, so 0.411 is high enough for d0.

[xscjun]

@zylo117 Thank you for your reply, I change the lr to 1e-3,
train from the weight you provided,efficientdet-d0.pth After 17epoch , the map : 0.409, then , the loss is no longer drop.

I wander ， If I train the model on another dataset(only for person ,100000-bboxes),should I freeze the backbone ?

[zylo117]

Why did you increase lr instead of reducing? And always freeze backbone first.