Hi,
\nI see the agent filter 0.5 is given only for train data and Eval data, but not for test data.Is it because we are not allowed to know the ground truth of test data?
\nRegards,
\nThe Brown Iceman
Hi,
\nI see the agent filter 0.5 is given only for train data and Eval data, but not for test data.Is it because we are not allowed to know the ground truth of test data?
\nRegards,
\nThe Brown Iceman
Hii @deepakrajpurushothaman
\ni went through the config file and this thing( agent filter) confused me. what is it?
Its just a filter for the dataset. Threshold of 0.5 means it includes frames with agent of interest. In our case agent of interests are pedestrians, cyclists and other cars.
\nhttp://www.l5kit.org/API/l5kit.data.filter.html
", "rawMarkdown": "Its just a filter for the dataset. Threshold of 0.5 means it includes frames with agent of interest. In our case agent of interests are pedestrians, cyclists and other cars. \n\nhttp://www.l5kit.org/API/l5kit.data.filter.html", "votes": null }, { "id": "996770", "postDate": "09/03/2020 14:40:44", "content": "thank you @deepakrajpurushothaman
", "rawMarkdown": "thank you @deepakrajpurushothaman", "votes": null }, { "id": "1000525", "postDate": "09/06/2020 15:29:41", "content": "@deepakrajpurushothaman can the value of threshold be considered as a hyperparameter or we have to take it fixed as 0.5? Also, how the probabilities of the agents are calculated (on which they are filtered) ? are we computing that probs during training or its already fixed in the dataset? Thanks!
", "rawMarkdown": "deepakrajpurushothaman can the value of threshold be considered as a hyperparameter or we have to take it fixed as 0.5? Also, how the probabilities of the agents are calculated (on which they are filtered) ? are we computing that probs during training or its already fixed in the dataset? Thanks!", "votes": null }, { "id": "1000582", "postDate": "09/06/2020 16:25:49", "content": "@pawankumarsahu My understanding is that, Its not a parameter. Its a type filter. So you could Fix at 0.5 and do your thing. There is no probability, its labeled data and already assigned probability of 1.0. Basically what happens when you set to 0.5 is that you are focusing on relevant agents like cars, cyclists and pedestrian.
", "rawMarkdown": "pawankumarsahu My understanding is that, Its not a parameter. Its a type filter. So you could Fix at 0.5 and do your thing. There is no probability, its labeled data and already assigned probability of 1.0. Basically what happens when you set to 0.5 is that you are focusing on relevant agents like cars, cyclists and pedestrian.", "votes": null }, { "id": "1000795", "postDate": "09/06/2020 19:54:36", "content": "@deepakrajpurushothaman thanks for that clear explanation!
", "rawMarkdown": "deepakrajpurushothaman thanks for that clear explanation!", "votes": null }, { "id": "1003082", "postDate": "09/08/2020 16:42:40", "content": "Hi @deepakrajpurushothaman, for the test set you are provided with a mask already (the output of the select_agents script) so you don't have to filter yourself. The reason is that we selected the agents in a way that doesn't leak any future GT (in particular, we removed all frames in each scene after a given point T and set only agents at T-1 as True in the mask).
Hii @deepakrajpurushothaman
\ni went through the config file and this thing( agent filter) confused me. what is it?
Its just a filter for the dataset. Threshold of 0.5 means it includes frames with agent of interest. In our case agent of interests are pedestrians, cyclists and other cars.
\nhttp://www.l5kit.org/API/l5kit.data.filter.html
", "votes": null, "replies": [] }, { "id": 996770, "author_name": "debasish05", "author_url": "", "post_date": "09/03/2020 14:40:44", "content": "thank you @deepakrajpurushothaman
", "votes": null, "replies": [] } ] }, { "id": 1000525, "author_name": "pawankumarsahu", "author_url": "", "post_date": "09/06/2020 15:29:41", "content": "@deepakrajpurushothaman can the value of threshold be considered as a hyperparameter or we have to take it fixed as 0.5? Also, how the probabilities of the agents are calculated (on which they are filtered) ? are we computing that probs during training or its already fixed in the dataset? Thanks!
", "votes": null, "replies": [ { "id": 1000582, "author_name": "deepakrajpurushothaman", "author_url": "", "post_date": "09/06/2020 16:25:49", "content": "@pawankumarsahu My understanding is that, Its not a parameter. Its a type filter. So you could Fix at 0.5 and do your thing. There is no probability, its labeled data and already assigned probability of 1.0. Basically what happens when you set to 0.5 is that you are focusing on relevant agents like cars, cyclists and pedestrian.
", "votes": null, "replies": [] }, { "id": 1000795, "author_name": "pawankumarsahu", "author_url": "", "post_date": "09/06/2020 19:54:36", "content": "@deepakrajpurushothaman thanks for that clear explanation!
", "votes": null, "replies": [] } ] }, { "id": 1003082, "author_name": "lucabergamini", "author_url": "", "post_date": "09/08/2020 16:42:40", "content": "Hi @deepakrajpurushothaman, for the test set you are provided with a mask already (the output of the select_agents script) so you don't have to filter yourself. The reason is that we selected the agents in a way that doesn't leak any future GT (in particular, we removed all frames in each scene after a given point T and set only agents at T-1 as True in the mask).