train VGG-SSD from scratch get a better result

KimSoybean / ScratchDet

The code and models for paper: "ScratchDet: Exploring to Train Single-Shot Object Detectors from Scratch"

229 stars 38 forks source link

train VGG-SSD from scratch get a better result #4

Open ujsyehao opened 5 years ago

ujsyehao commented 5 years ago

Hi, I follow your experiment setting, train VGG-SSD from scratch(no pretraining, no BN in the backbone, no BN in the head), use the following hyper param(caffe):

train batch size: 32, iter: 1, test batch size: 8
base_lr: 0.001
max_iter: 120000
type: SDG
weight_decay: 0.0005
gamma: 0.1

At 110k iterations, I get 68.8% mAP, It is higher than 67.6% mAP(paper reported).

CHKswiftly commented 5 years ago

Can your share your VGG-SSD code with me? there is some problems with the code which i want to repeat,