Training doesn't work on custom datasets

[pouvoirdasha]

Hi! First of all, thank you for your work!
I have tried to train the program on a custom dataset which I have created following your instructions. I have tried both - .mp4 and folders with .png, and none of them work - I get several errors when I try to run it:

• caught valueerror in dataloader worker process 0
• value error 'a' cannot be empty unless no samples are taken

The problem seems to be raised because of torch, any ideas how I can solve it?

[pouvoirdasha]

Okay, I tried also training on mgif dataset that you have in your paper and here is what I got :

[AliaksandrSiarohin]

For mgif, everything seems ok to me, just need to wait longer. Actually hard to say, you should check if nvidia-smi using the gpu.
For custom dataset, probably your naming is incorect, or you should disable id_sampling in the config. (see

first-order-model/config/taichi-256.yaml

Line 18 in f4ff6da

id_sampling: True

)

[fjesikfjdskl]

Hi! First of all, thank you for your work! I have tried to train the program on a custom dataset which I have created following your instructions. I have tried both - .mp4 and folders with .png, and none of them work - I get several errors when I try to run it:

• caught valueerror in dataloader worker process 0
• value error 'a' cannot be empty unless no samples are taken

The problem seems to be raised because of torch, any ideas how I can solve it?

This problem can be solved by looking at the file name in line 103 of frames_dataset. Can you tell me what the training steps are for your custom dataset？

[fjesikfjdskl]

@AliaksandrSiarohin Hello, can you tell me that custom dataset training only needs to place.mp4 or.png files in the folder and modify the configuration file, and finally run run.py? I only did the training test on a few hundred images placed in the price folder, but I got the log.txt result as shown below. Is this result normal?

[pouvoirdasha]

Hi! Thanks forfor your replies, in the config file I have specified 30 epochs, 2 repeats and [15,25] as epoch milestones (I have a dataset of videos split into two folders test and train, and each of folders contains separate folders (video1, ... video_n) with the videos split frame by frame in 900 png files each approx). I have tried different values for epochs/repeats etc but it doesn't seem to change the issue...

I sue nvidia geforce rtx 3090 GPUs to train the model, and even for the mgif I haven't got any results after approx. an hour of training (log.txt is empty)

[fjesikfjdskl]

@pouvoirdasha How do you choose between test sets and training sets? I put all the training images in a folder and started running python run.py.

[bajiua]

你好！首先，感谢您的工作！ 我尝试在按照您的指示创建的自定义数据集上训练该程序。我已经尝试过 - .mp4 和带有 .png 的文件夹，但它们都不起作用 - 当我尝试运行它时，我遇到了几个错误：

• 在数据加载器工作进程 0 中捕获值错误
• 值错误“a”不能为空，除非没有采样

这个问题似乎是因为火炬而提出的，我有什么想法可以解决它吗？

Hello,I had the same problem when training on mgif datasets. Have you solved it? How was it resolved? I am looking forward to your reply and would appreciate it very much if you could reply me.

[bajiua]

你好！首先，感谢您的工作！我尝试在按照您的指示创建的自定义数据集上训练该程序。我已经尝试过 - .mp4 和带有 .png 的文件夹，但它们都不起作用 - 当我尝试运行它时，我遇到了几个错误：

• 在数据加载器工作进程 0 中捕获值错误
• 值错误“a”不能为空，除非没有采样

这个问题似乎是因为火炬而提出的，我有什么想法可以解决它吗？

这个问题可以通过查看frames_dataset第103行的文件名来解决。您能告诉我您的自定义数据集的训练步骤是什么吗？

Hello, I saw that you said that this problem could be solved by looking at the file name in line 103 of the frames_dataset. How to modify it? I would appreciate it if you could reply me.

[G-force78]

In line 70

if os.path.exists(os.path.join(root_dir, 'train')):

change the path to your training directory if os.path.exists(os.path.join(root_dir, 'yourpathhere/train')):

[bajiua]

在第 70 行

如果 os.path.exists(os.path.join(root_dir, 'train')):

如果 os.path.exists(os.path.join(root_dir, 'yourpathhere/train')) 则更改训练目录的路径：

First of all, thank you, but after I changed it, I also modified it in the yaml file, but it didn't work. Is there anything else I need to modify?

[G-force78]

I'm not sure sorry I couldnt get training to work at all, no matter what I did it would not recognise frames_dataset as a module.. If you like you can try swapping the files out for ones I've modified
here where I solved some of the problems.. Good luck
https://github.com/G-force78/articulated-animation

[bajiua]

我不确定抱歉，我根本无法接受培训，无论我做什么，它都不会将frames_dataset识别为模块。如果您愿意，您可以尝试将文件换成我在这里修改过的 文件一些问题..祝你好运 https://github.com/G-force78/铰接动画

Thank you. Did you also encounter the same problem as me in first-order-try? Are the changes you made in articulated animation already runnable?