Questions about results reproduction

[anonymous-pusher]

Hello, thank you for releasing this amazing work.
I would like to reproduce the results on the humandml3d dataset and have some questions about it:

• About the hyperparameters, in the paper you use the number of iterations, while the configs in the code use END_EPOCH. Do they mean the same thing for you, or is it that nb_iterations = END_EPOCH * nb_batches ?

I also have a question about the real time application of the method; is it possible to use the method for motion captioning in real time, for example a motion is being played and the description of motion is quite in sync as the motion is being generated, so there is no big lag between the two?

[blue-blue272]

+1

[Spark001]

@mJones00 Have you solve the reproducing problem?
I set END_STEPS instead of END_EPOCH, then train 150K steps for stage1, 300K for stage2, 300K for stage3. But I only get 40.6 for RP_1 as follow:

[www-Ye]

@mJones00 Have you solve the reproducing problem?繁殖问题解决了吗？ I set END_STEPS instead of END_EPOCH, then train 150K steps for stage1, 300K for stage2, 300K for stage3. But I only get 40.6 for RP_1 as follow:我设置 END_STEPS 而不是 END_EPOCH，然后为 stage1 训练 150K 步骤，为 stage2 训练 300K，为 stage3 训练 300K。但我只得到 RP_1 的 40.6，如​​下：

Hello, I used the official checkpoint's VAE along with 300 epochs for stage 2 and 100 epochs for stage 3, but the model I trained only achieved around 0.3 R Top1, not even reaching 0.4. Have you made any modifications?