zero mAPs in the first 5 epochs

[sandylaker]

Hi,

I tried to train efficientdet_d1 on COCO dataset from scratch. Here are my settings:

• hardware: 8x A100 GPUs;
• batch size: 18 images on each GPU;
• lr: linearly scaled to 0.18 = (0.16 / 128) * 18 * 8.
• others: no amp, no fused momentum, 5 epochs for warm-up.

I also modified the train.py to enable monitoring the classification and regression losses. Although the losses kept decreasing in the first 5 epochs, the mAPs on validation set were near 0 all the time. Have you also observed this issue? Is it normal?

[rwightman]

@sandylaker If you have EMA enabled that's expected, and will likely remain like that for 30-100 epochs depending on the decay constant. If you do not have ema enabled then something is likely wrong.

[jaideep11061982]

I too face similar issue with eff det 2.
Loss dsnt comes down ,I.copied almost same dataset that worked last for me in kaggle gwd competition . Difference is only no of classes
I hope I pass in box cord in right format .
Ymin , x min , ymax x max ,
where x max is x +width , y max is ymin +height

Also Class loss plus box loss dsnt adds to total loss .
0.5... + .002 box loss ,total loss 0.7

Any help or pointers