sh_albert_cls.sh get a weired score 0.499

[HuipengXu]

Has anyone ran sh_albert_cls.sh and got a strange score, keeping 0.499 unchanged

[tufeiming]

Has anyone ran sh_albert_cls.sh and got a strange score, keeping 0.499 unchanged

Has anyone ran sh_albert_cls.sh and got a strange score, keeping 0.499 unchanged

Which model do you use? Because I use RTX 3060 graphics card for training, I choose albert-base-v2, and the result I get is acc = 0.6097869114798282.

[HuipengXu]

Has anyone ran sh_albert_cls.sh and got a strange score, keeping 0.499 unchanged

Has anyone ran sh_albert_cls.sh and got a strange score, keeping 0.499 unchanged

Which model do you use? Because I use RTX 3060 graphics card for training, I choose albert-base-v2, and the result I get is acc = 0.6097869114798282.

I also use albert-base-v2. After downsampling, the ratio of positive and negative samples is approximately equal to 40000:50000. In the first 3000 steps, the score is approximately equal to 0.71, the highest is 0.76, and then it starts to decline. What’s even stranger is that after I remove the downsampling, the score is no longer 0.499, but 0.61

[tufeiming]

Has anyone ran sh_albert_cls.sh and got a strange score, keeping 0.499 unchanged

Has anyone ran sh_albert_cls.sh and got a strange score, keeping 0.499 unchanged

Which model do you use? Because I use RTX 3060 graphics card for training, I choose albert-base-v2, and the result I get is acc = 0.6097869114798282.

I also use albert-base-v2. After downsampling, the ratio of positive and negative samples is approximately equal to 40000:50000. In the first 3000 steps, the score is approximately equal to 0.71, the highest is 0.76, and then it starts to decline. What’s even stranger is that after I remove the downsampling, the score is no longer 0.499, but 0.61

我今天又把epoch改为5跑了下，结果也是acc = 0.4992840899519919，不清楚原因。。。另外执行后面的sh_albert_av.sh需要先执行sh_albert_cls.sh吗？

[HuipengXu]

Has anyone ran sh_albert_cls.sh and got a strange score, keeping 0.499 unchanged

Has anyone ran sh_albert_cls.sh and got a strange score, keeping 0.499 unchanged

Which model do you use? Because I use RTX 3060 graphics card for training, I choose albert-base-v2, and the result I get is acc = 0.6097869114798282.

I also use albert-base-v2. After downsampling, the ratio of positive and negative samples is approximately equal to 40000:50000. In the first 3000 steps, the score is approximately equal to 0.71, the highest is 0.76, and then it starts to decline. What’s even stranger is that after I remove the downsampling, the score is no longer 0.499, but 0.61

我今天又把epoch改为5跑了下，结果也是acc = 0.4992840899519919，不清楚原因。。。另外执行后面的sh_albert_av.sh需要先执行sh_albert_cls.sh吗？

应该不需要吧，我执行 sh_albert_av.sh 好像是正常的，你试过对负样本下采样吗，我这边确实可以到 0.76 左右，但应该还是不正常

[JiTingyu]

Has anyone ran sh_albert_cls.sh and got a strange score, keeping 0.499 unchanged

Has anyone ran sh_albert_cls.sh and got a strange score, keeping 0.499 unchanged

Which model do you use? Because I use RTX 3060 graphics card for training, I choose albert-base-v2, and the result I get is acc = 0.6097869114798282.

I also use albert-base-v2. After downsampling, the ratio of positive and negative samples is approximately equal to 40000:50000. In the first 3000 steps, the score is approximately equal to 0.71, the highest is 0.76, and then it starts to decline. What’s even stranger is that after I remove the downsampling, the score is no longer 0.499, but 0.61

我今天又把epoch改为5跑了下，结果也是acc = 0.4992840899519919，不清楚原因。。。另外执行后面的sh_albert_av.sh需要先执行sh_albert_cls.sh吗？

应该不需要吧，我执行 sh_albert_av.sh 好像是正常的，你试过对负样本下采样吗，我这边确实可以到 0.76 左右，但应该还是不正常

请问是如何进行下采样的

[HuipengXu]

Has anyone ran sh_albert_cls.sh and got a strange score, keeping 0.499 unchanged

Has anyone ran sh_albert_cls.sh and got a strange score, keeping 0.499 unchanged

Which model do you use? Because I use RTX 3060 graphics card for training, I choose albert-base-v2, and the result I get is acc = 0.6097869114798282.

I also use albert-base-v2. After downsampling, the ratio of positive and negative samples is approximately equal to 40000:50000. In the first 3000 steps, the score is approximately equal to 0.71, the highest is 0.76, and then it starts to decline. What’s even stranger is that after I remove the downsampling, the score is no longer 0.499, but 0.61

我今天又把epoch改为5跑了下，结果也是acc = 0.4992840899519919，不清楚原因。。。另外执行后面的sh_albert_av.sh需要先执行sh_albert_cls.sh吗？

应该不需要吧，我执行 sh_albert_av.sh 好像是正常的，你试过对负样本下采样吗，我这边确实可以到 0.76 左右，但应该还是不正常

请问是如何进行下采样的

就是对负样本随机下采样

[ryanpram]

Hi i have same problem when run sh_albert_cls.sh, the score remains unchanged whatever the param is

[HuipengXu]

Hi i have same problem when run sh_albert_cls.sh, the score remains unchanged whatever the param is

Have you solved it now？

[ryanpram]

Hi i have same problem when run sh_albert_cls.sh, the score remains unchanged whatever the param is

Have you solved it now？

Unfortunately not yet. Hope author or anybody can help

[cooelf]

Dear, sorry for the late reply. It seems that albert-base-v2 does not converge according to your results. I did not try the base models, and the given hyper-parameters are tuned for xxlarge models. Maybe you can try to use the settings in google-research/albert#235