Decoder output == nan

[rui-love]

When I use my dataset to train the model, the decoder output all are nan. The only thing I do is changing the data and road net. And I get the results from the 100 example trajectories.

[chenyuqi990215]

can you provide more details, so that I can help you? Do the parameters contain nan after the gradient descend? or Do the inputs contain nan? or Do the log operations cause nan?

[rui-love]

Thank you very much. The parameters don't contain any nan value. The input are all 0. When I use the torch.autograd.set_detect_anomaly(True)

May be the log operations cause nan? How can I deal with it?

[rui-love]

sorry for the waiting. I have reply the detail of my experiment on the github issues.

…

-----原始邮件----- 发件人:"Yuqi Chen" ***@***.***> 发送时间:2023-12-27 16:28:25 (星期三) 收件人: chenyuqi990215/RNTrajRec ***@***.***> 抄送: 瑞の爱 ***@***.***>, Author ***@***.***> 主题: Re: [chenyuqi990215/RNTrajRec] Decoder output == nan (Issue #14) can you provide more details, so that I can help you? Do the parameters contain nan after the gradient descend? or Do the inputs contain nan? or Do the log operations cause nan? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[rui-love]

hello! do you need more information to deal with this problem？

[chenyuqi990215]

Maybe you can set 'search_dist' and 'neighbor_dist' in multi_main.py to a large enough values which is larger than the maximum GPS error of your dataset.

[rui-love]

There is the same situation, is there any other parameters should be care about?

[maxwang967]

it happens when i'm using a alternative dataset, while it's ok on other baseline models.

[maxwang967]

There is the same situation, is there any other parameters should be care about?

The issue can be sovled by add a small constant (e.g., 1e-6) to 'x_exp_sum' in both mask_log_softmax funciton and mask_graph_log_softmax function.

[maxwang967]

There is the same situation, is there any other parameters should be care about?

The issue can be sovled by add a small constant (e.g., 1e-6) to 'x_exp_sum' in both mask_log_softmax funciton and mask_graph_log_softmax function.

@chenyuqi990215 Please consider modify the source code to make it more robust, thank you!

[chenyuqi990215]

Thank you for your advice.