GST training question.

[CAI23sbP]

Hi! @Shuijing725 , i read your reference. But there is no script about how to make train dataset. Could you tell me how to create dataset? (In my situation, actually i use other platform about DRL environment. )

[LifGorg]

I have the same question as @CAI23sbP :)

When I run "sh run/create_datasets_eth_ucy.sh", it couldn't find any create_datasets_eth_ucy.py file that's written in it.

Ted Huang https://github.com/tedhuang96/gst/tree/main/scripts has a full dataset creation file, and I am trying to change it to generate batch dataset myself

[CAI23sbP]

Oh thank you for your answer @LifGorg !. I will try that :). By the way, i have three questions about your code.

1. When you take a prediction data from function (generate_ob) and you calculate collision penalty in that function. why did you create like that? why did not calculate reward in reward function?
2. You created dummy human for zero detection in scan rage, i don't understand why you did.
3. This algorithm can training and testing at fixed env? It mean is you create observation space [visible_mask],[spatial_edges],[detected_human_num] are fixed size.

[Shuijing725]

Hi @CAI23sbP,

Sorry for the late reply. If you want an ORCA dataset, you can run collect_data.py. If you want real pedestrian data, then I recommend looking into Ted Huang's repo as suggested by @LifGorg.

1. To accelerate training, we calculate r_pred for multiple parallel environments in a batch manner. The true prediction reward calculation is in this line.
2. If no human is detected, I feed one dummy human to the attention network as a workaround to maintain consistency. Of course, there are other workarounds.
3. The [visible_mask],[spatial_edges],[detected_human_num] are fixed because they are created according to the maximum number of humans. This design choice is because gym environment requires me to have a fixed observation space. If fewer humans are detected, the undetected entries are filled with dummy values and their gradients are masked out during backprop.