mxnet-ssd only get 73.8map on voc2007

I train the ssd vgg16 300 on voc2007+voc2012 trainval, batchsize 64 and trained 240 epoches as the default settings, but I only get 73.8map,how to train to get 77map as the readme says.

Sep 05 '17 12:09 lzx1413

batchsize=32, lr-steps=80,160,240, increase end-epoch, better train on single gpu. YMMV

Sep 05 '17 17:09 zhreshold

Thanks.

Sep 06 '17 00:09 lzx1413

still can not get the 77 map..can you show me the training log, @zhreshold thanks

Sep 09 '17 01:09 lzx1413

also cannot reproduce the map @zhreshold , would mind get the dealt training optimizer stage? !

Sep 10 '17 14:09 Yezipiaomu

I don't have the logs now, what mAP did you guys get, have you tried some other network like resnet or inception?

Sep 10 '17 17:09 zhreshold

about 73map, I have only tried vgg16 reduced @zhreshold

Sep 11 '17 01:09 lzx1413

vgg model map 0.738 , Batchsize 32 , start lr at 0.004, and In four gpu

Sep 11 '17 01:09 Yezipiaomu

I guess something is broken. Will update later.

Sep 11 '17 02:09 zhreshold

It's the freeze parameters that leads to the result ?!

Sep 11 '17 03:09 Yezipiaomu

77.3 map In Resnet50 data_shape 512, would mind tell me how to improve?

Sep 14 '17 13:09 Yezipiaomu

77.142 map in Vgg-reduced 512, lr = 0.004 batch_size =32 @zhreshold Also cannot reproduce the result.

Sep 17 '17 08:09 Yezipiaomu

@zhreshold Any update on this?

Feb 27 '18 00:02 siyiding1216