The run model has poor training effect

[todayThursday]

Thanks for your effort, I am trying to reproduce the code. But I have a few questions: There is a weight value input of the value function r in the matadata.txt file. When I trained the walk model without any modification, the result was not bad in the end. But when I trained run or other models, the training could not reach the normal result for a long time.

[InckieUltra]

Hi, I'm reproducing this paper too.
The walk.bvh is good, but the training of kick.bvh meets some difficulty.
I think it's mainly because the training speed.
The paper says it needs about 12 hours to get 10 million tuples, and the paper uses i7-7700 and 1070ti.
But I spend 11 hours training, getting only about 200*2000=0.4 million tuples. I use r5-9600x and 4070tis. There seems to be some problems with my training speed.
I want to ask what's your training speed? Thanks!

[todayThursday]

Hi, I'm reproducing this paper too. The walk.bvh is good, but the training of kick.bvh meets some difficulty. I think it's mainly because the training speed. The paper says it needs about 12 hours to get 10 million tuples, and the paper uses i7-7700 and 1070ti. But I spend 11 hours training, getting only about 200*2000=0.4 million tuples. I use r5-9600x and 4070tis. There seems to be some problems with my training speed. I want to ask what's your training speed? Thanks!

sorry, I have not trained the kick model, so I cannot match the training speed with you. When I was training walk, I trained for about six hours, and my device was i9-4080s, and about 10 million tuples were collected. You also mentioned that your walk training effect was very good. I would like to ask if you can observe and see from the UI interface, do you have some data comparison to show that its effect is very good? I had a terrible time training run's model, and I was wondering if you had any such problems. Did you make any changes to the code in the meantime?

[todayThursday]

I am a beginner, maybe we can add a contact information to learn from each other

[InckieUltra]

Hi, I'm reproducing this paper too. The walk.bvh is good, but the training of kick.bvh meets some difficulty. I think it's mainly because the training speed. The paper says it needs about 12 hours to get 10 million tuples, and the paper uses i7-7700 and 1070ti. But I spend 11 hours training, getting only about 200*2000=0.4 million tuples. I use r5-9600x and 4070tis. There seems to be some problems with my training speed. I want to ask what's your training speed? Thanks!

sorry, I have not trained the kick model, so I cannot match the training speed with you. When I was training walk, I trained for about six hours, and my device was i9-4080s, and about 10 million tuples were collected. You also mentioned that your walk training effect was very good. I would like to ask if you can observe and see from the UI interface, do you have some data comparison to show that its effect is very good? I had a terrible time training run's model, and I was wondering if you had any such problems. Did you make any changes to the code in the meantime?

Sorry, I didnt calculate any metrics to compare with reference motion, and I didnt train the running motion too.
I said the walking moiton is good just because the model is capable of maintaining balance and mimic the walking motion.
And the training speed problem is solved after rebuilding. Thanks for your information.

I am a beginner, maybe we can add a contact information to learn from each other

Sure, this is my tencent QQ: 1242869192.