How about voc 07+12trainval in 07 test dataset map ？I only have 2080Ti*2, can I reproduce this value？ about asff

[GOATmessi8]

For VOC, our experiments show that asff and rfb could not benefit a lot as on COCO. Both the baseline model and ASFF model can get about 84.0 mAP on 07 test.
With two gpus, the low batchsize may downgrade the accuracy, you may turn-off the random resizing and use bigger batchsize.

[lovepan1]

OK. If i get result, i will reply U ,thanks a lot.

[jwnirvana]

OK. If i get result, i will reply U ,thanks a lot.

could you tell me whether you get result，thx

[lovepan1]

sorry, I'm traveling outside in these days，so，I didn't get results

[lovepan1]

hello, i use 2*2080Ti, when i trained to 120 epochs , map arrived 0.72 in voc2007test. i think the reasons of lower map is my gpu.
when i trained voc datasets, gradient explosions always occured, even if i use warm up lr, so, i use 1e-6 or smaller lr to train my model.
after 120 epochs, the best map is 0.72, and if i continue training, map will r
educe .
how about your map?
@jwnirvana @ruinmessi