I was not able to reproduce the published mAP with SSD VGG16_reduced model with VOC0712 train-validation and VOC07 test data (300x300 image size) with MxNet (Expected 0.77, Actual 0.73).

I used an effective batch size of 32 on 2 GPUs with a learning rate of 0.001. Rest of the hyper-parameters were as documented in GitHub.

Trained on AWS Deep Learning AMI Ubuntu Linux - 2.4_Oct2017 (ami-37bb714d) p2.8xlarge

ssd_caffe_test.txt
ssd_caffe_train.txt
ssd_mxnet_trainval.txt