\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3722352%2Fb8b77b30ad92549cd6131090347821c2%2F1401603111927_.pic_hd.jpg?generation=1603112114960662&alt=media\")
When we are training the multi-mode baseline, we got confidences as a part of the output. However, we have no true label in the data, as shown in the following screenshot, for the backpropagation. How can we adjust the multi-mode network to predict the mode of agents without any true label? Are the confidences we predict is totally meaningless?
\n
target_positions : The coordinates (in agent reference system) of the AV in the future. Unit is meters
\nhttps://github.com/lyft/l5kit/blob/03eb9e037d23940a134e27c9f124021e18982020/data_format.md
target_positions and target_availabilities are the labels you're looking for
So we do not actually need to predict the mode of agents like what the baseline I cite above?
", "rawMarkdown": "So we do not actually need to predict the mode of agents like what the baseline I cite above? " }, { "id": 1053897, "postDate": "2020-10-19T13:19:45.153Z", "content": "You can choose whether to predict multiple modes, or just a single mode. Multiple modes make sense in this context as that's the format in which we are being measured.
\nYou predict multiple modes, but the negative log likelihood metric measures these against a single truth. See here
\nThis notebook is a good starting point.
", "rawMarkdown": "You can choose whether to predict multiple modes, or just a single mode. Multiple modes make sense in this context as that's the format in which we are being measured. \n\nYou predict multiple modes, but the [negative log likelihood metric](https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/overview/evaluation) measures these against a single truth. See [here](https://github.com/lyft/l5kit/blob/20ab033c01610d711c3d36e1963ecec86e8b85b6/l5kit/l5kit/evaluation/metrics.py#L4)\n\n[This](https://www.kaggle.com/corochann/lyft-training-with-multi-mode-confidence) notebook is a good starting point.", "votes": 4 }, { "id": 1053899, "postDate": "2020-10-19T13:26:39.123Z", "content": "It seems that I misunderstood what the conf_0,1,2 actually used for… Thank you so much.
When we are training the multi-mode baseline, we got confidences as a part of the output. However, we have no true label in the data, as shown in the following screenshot, for the backpropagation. How can we adjust the multi-mode network to predict the mode of agents without any true label? Are the confidences we predict is totally meaningless?
\n
target_positions : The coordinates (in agent reference system) of the AV in the future. Unit is meters
\nhttps://github.com/lyft/l5kit/blob/03eb9e037d23940a134e27c9f124021e18982020/data_format.md
target_positions and target_availabilities are the labels you're looking for
So we do not actually need to predict the mode of agents like what the baseline I cite above?
", "votes": 0, "replies": [] }, { "id": 1053897, "author_name": "fergusoci", "author_url": "", "post_date": "2020-10-19T13:19:45.153000", "content": "You can choose whether to predict multiple modes, or just a single mode. Multiple modes make sense in this context as that's the format in which we are being measured.
\nYou predict multiple modes, but the negative log likelihood metric measures these against a single truth. See here
\nThis notebook is a good starting point.
", "votes": 4, "replies": [] }, { "id": 1053899, "author_name": "Yannik", "author_url": "", "post_date": "2020-10-19T13:26:39.123000", "content": "It seems that I misunderstood what the conf_0,1,2 actually used for… Thank you so much.