Predict code execution error

[Pluteapple]

Hello! I'm sorry to bother you. In the process of reproducing the dyna dataset, the following error occurs when the predicet code is executed. Please tell me how to solve this problem. I would be very grateful if I could receive your reply.

[ChocolateBiscuit]

Try explicitly casting n_tens_grid into a long tensor.

[Pluteapple]

I modified it according to the method you provided. As shown in the screenshot, because the obtained value is empty, an error still occurs here. When I iterated for five cycles before, I performed inference and could generate an obj file. However, when I retrained for 200 cycles, I modified the command line and directly reported the original error, and the training results of five previous runs could not be inferred.

[ChocolateBiscuit]

It seems that in this case, as the error indicates, max_len should be converted into an integer.

[Pluteapple]

I tried it. I trained 50 times on my own data set with 300 obj formats. Finally, the inference is generated as follows. I would like to ask you what are the requirements for the training cycle and the size of the data set? There are also options for cube 0.20, 0.25, etc.

[ChocolateBiscuit]

We are able to train on as small of a dataset as COSEG vases (300). I recommend using either cube_0.05 or cube_0.03. They are the highest resolution grids. I also recommend using the write_to_obj functionality of the meshes and checking that the surface meshes you train on look reasonable.

[Pluteapple]

Thank you for your patient reply, I will try again, thank you very much

[Pluteapple]

Hello, I use the A40 graphics card to debug the code. I changed the grids to 0.05 according to your last prompt, and the following video memory error will be reported directly.