Can't do repreduction

[CHELSEA234]

Hi,

I downloaded your code and ran it on my local, and I did averaging over 3 times experiment printing results. I got some numerical values (Fig1), acceptable yet still worse than ones in your paper (Fig2), can you tell me where I made mistakes? BTW, in figure 1, the every row's last value means mean error of all 25 frames.

Best, Xiao Guo

[CHELSEA234]

Here is more complete comparison between my questionable reproduction and your published results:

I hope you could give me some hints at this, really appreciate your help!

---

Can you share your save checkpoint file?

[chaneyddtt]

Hi Xiao Guo, Sorry for the late reply. I don't have the original checkpoint file. But I trained my model just now and reached an average error of 1.77 after training for 24000 iterations . I can sent you the checkpoint file if you need. Or you can train more than 20000 iterations(30000 for example) by yourself and choose the best one, which i think can reach a similar average result.

[zaverichintan]

Hello @chaneyddtt , I will try to follow the steps you mentioned for the number of iterations. Could you upload the checkpoint file somewhere or send link to the uploaded version?

[chaneyddtt]

Hi @zaverichintan, I upload the model here. Please note that I trained the network again and the result for each action may be slightly different, but the overall average is the same.

[zaverichintan]

Do I need to run the training once to generate the sample folder? I have to copy the files you provided and copy them to a folder called Model? And then to how to run the AC_main.py with the new model?

[chaneyddtt]

Yes, or you can just put the model somewhere you like and specify the path in the AC_main.py. Note that you also have to specify the 'checkpoint' to keep consistent, for example 24000 for the provided model. Then you specify the 'is_sampling' as 'True' before you run the AC_main.py.

[zaverichintan]

It throws the error:

ValueError: Dimensions must be equal, but are 54 and 70 for 'VAESkeleton/add_1' (op: 'Add') with input shapes: [?,1,54,1], [?,1,70,1].

[chaneyddtt]

Do you set the parameter ''--dataset" to cmu? Because the input dimension is 70 for the cmu dataset, while it is 54 for the human3.6m.

[zaverichintan]

Is it possible to get the checkpoint for the Human3.6m dataset, and run the testing on it?

[chaneyddtt]

Yes, the checkpoint I provide is for the human3.6m dataset. Did you set the 'whole_sample_dim' in the Autoenc_gan to [None,50+args.output_length,54,1]. The input dimension should be 54 for the human3.6m dataset.

[zaverichintan]

Now it's running without any errors. Thanks