Can anyone reproduce the result by using a smaller batch size?

[zzf005]

In my case, only got EAO: 0.339/A: 0.601/R: 0.356 with batch size=40 which different from the result (0.380/0.609/0.276). Can someone help me solve this problem? Thank you.

[guangqianqin]

i use the same batch with author ,but got bad result.How many epochs do you use?after training train_siammask.py,the result is?

[zzf005]

@guangqianqin After 20 epochs, the best model (base) I got will be epoch 9 and the EAO is 0.293. What was the result you got? Had you tried to run the train_siammask_refine.py?

[guangqianqin]

The top one of EAO is 0.289;but top one of accuracy is 0.623 and its EAO is 0.287.whats more,after train_siammask_refine,the best model is generated by the second base model.Maybe i will try to change the optimizer( try to use lookahead +Radam).

[Damon2019]

@zzf005 @guangqianqin hi,How did you train your own dataset?could you give me some suggest?plz!

[hmmlencat]

@zzf005 @guangqianqin hi,How did you train your own dataset?could you give me some suggest?plz!

[Damon2019]

@zzf005 @guangqianqin hi,How did you train your own dataset?could you give me some suggest?plz!

I succeed

[ggy112]

@zzf005 @guangqianqin hi,How did you train your own dataset?could you give me some suggest?plz!

I succeed

您好，我想请教一下您是怎么成功训练siammask的,能否指点一下？谢谢。

[SamihaSara]

@zzf005 @guangqianqin hi,How did you train your own dataset? Could you give me some suggest?plz!

I succeed

How did you manage to do so? can you please mention the steps? I tried by reducing the dataset with only bag and ball1 in data/VOT2016 and test/VOT2016 but getting a divide by zero error

[HaoyuanXu]

EAO的最高位为0.289;但精度的最高位为0.623，其EAO为0.287。更重要的是，在train_siammask_refine之后，第二个基本模型生成了最佳模型。也许我会尝试更改优化器（尝试使用前瞻+ Radam）。

请问您复现出论文的结果了嘛

[guangqianqin]

没有，效果差很多，我在siamrpn(resnet)上做的改进，这个效果挺好

---Original--- From: "HaoyuanXu"<[email protected]> Date: Tue, Mar 17, 2020 22:03 PM To: "foolwood/SiamMask"<[email protected]>; Cc: "Jimzhou82sub"<[email protected]>;"Mention"<[email protected]>; Subject: Re: [foolwood/SiamMask] Can anyone reproduce the result by using a smaller batch size? (#102)

EAO的最高位为0.289;但精度的最高位为0.623，其EAO为0.287。更重要的是，在train_siammask_refine之后，第二个基本模型生成了最佳模型。也许我会尝试更改优化器（尝试使用前瞻+ Radam）。

请问您复现出论文的结果了嘛

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.

[zosel260]

after training, my final results was only EAO=0.35, my batch size were:

• train base model : 64(e1 ~ 10), 16(e11 ~ 20)
• train refinement : 32(e1 ~ 20)

My result is far from the result EOA=0.42 obtained from provided model(SiamMask_VOT_LD.pth) I reported here: https://github.com/foolwood/SiamMask/issues/176