pretrained model

[suguliwei]

Can you share the pretraind model of imagenet?That will be a lot helpful for me! Thanks!

[rog93]

@suguliwei sorry, I will not, but I can share some of my training details here, my solver.proto is like below: lr_policy: "multistep" stepvalue: 300000 stepvalue: 600000 stepvalue: 900000 base_lr: 0.045 momentum: 0.9 max_iter: 1000000 weight_decay: 4e-5 gamma: 0.1 solver_type: SGD batch_size: 128 And I use Inception augmentation as well. You can try to train you own model.

[chunniunai220ml]

@RuiminChen can you explain why you take multistep policy?just according to test?and how to understand the stepvalue? what is the Inception augmentation?can you provide some informaion?mean=[128.0],scale=0.0078????and some others ???

[rog93]

@chunniunai220ml 1.why multistep policy? just personal experience. 2.how to understand the stepvalue? ImageNet Classfication training set has about 1280k images，300k iteration (128 batchsize) ≈ 30 epochs 3.Inception augmentation means ImageNet data augmentation strategy proposed by InceptionV1，you can get more details here: https://arxiv.org/abs/1409.4842 4.BGR mean values [103.94,116.78,123.68] are subtracted and scale=0.017 is used as std values for image preprocessing. Hope these helpful for you.

[chunniunai220ml]

@RuiminChen thanks for you reply timely. I have read the paper,and I want to know the data augmentation how important it is. the paper set resize_ratio:8-100%，aspect_ratio：[3/4,4/3],and distortions refer a paper,it also use random interpolation,have you done all of these measures in data layer? more importantly,you add these data augmentation in caffe layer or python layer or some other ways?