I think this file can help me to understand which columns should be predicted in the output, but I have no idea to deal with it.
", "messages": [ { "id": "1015930", "postDate": "09/18/2020 14:30:33", "content": "I think this file can help me to understand which columns should be predicted in the output, but I have no idea to deal with it.
", "rawMarkdown": "I think this file can help me to understand which columns should be predicted in the output, but I have no idea to deal with it.", "votes": null }, { "id": "1015964", "postDate": "09/18/2020 15:10:08", "content": "agents_mask is a 1D boolean array with the same length of the agent array. If a value is True, that agent will be considered as an element of the dataset, otherwise it will be skipped.
I see. Thanks a lot!
", "rawMarkdown": "I see. Thanks a lot!", "votes": null }, { "id": "1025165", "postDate": "09/24/2020 11:31:39", "content": "@lucabergamini
\nIsn't agents_mask a 2D array? In function load_agents_mask in agent.py file, it returns a numpy.ndarray with shape (1893736, 2) for sample.zarr dataset:
def load_agents_mask(self) -> np.ndarray:\n \"\"\"\n Loads a boolean mask of the agent availability stored into the zarr. Performs some sanity check against cfg.\n Returns: a boolean mask of the same length of the dataset agents\n \"\"\"\n\n ...\n ...\n agents_mask = convenience.load(str(agents_mask_path)) # This returns `(1893736, 2)` for sample.zarr\n return agents_mask\nLater on in the __init__ function of class AgentDataset, the agents_mask is compared against min_frame_history and min_frame_future as follows:
past_mask = agents_mask[:, 0] >= min_frame_history\nfuture_mask = agents_mask[:, 1] >= min_frame_future\nagents_mask = past_mask * future_mask\nI wonder what those two columns of agents_mask correspond to?
You are correct that it is a little bit confusing, agents_mask is different things depending on its shape. If it is 1D boolean vector then what @lucabergamini said. If it is a 2D integer vector then it specifies in how many frames agent is relevant before and after the sample in question (\"relevant\" here is my word). Note that relevant agent frames is not the same as frames where agent is available, read this discussion for explanation of the difference.
Yeah, there's two different stored masks. The agents_mask array in the zarr datasets which is the 2D format and is used internally in AgentDataset to filer frames (and create an internal 1D mask, affected by certain config options). You don't need to manually load this or generally deal with it.
\nThen there's the mask.npz file you load (with numpy.load) and pass in to AgentDataset which is 1D boolean mask with a pre-specified set of frames for which you must make test predictions.
Yes, sorry for the confusion. The mask stored on the disk has future and history availabilities (hence 2D) while the one used at runtime is a 1D boolean array. The first is converted into the second one at runtime here
", "rawMarkdown": "Yes, sorry for the confusion. The mask stored on the disk has future and history availabilities (hence 2D) while the one used at runtime is a 1D boolean array. The first is converted into the second one at runtime [here](https://github.com/lyft/l5kit/blob/90a6109754c8a75199219188dff2e25d1b4489ab/l5kit/l5kit/dataset/agent.py#L38)", "votes": null }, { "id": "1025388", "postDate": "09/24/2020 14:20:44", "content": "@lucabergamini , don't you introduce another confusion here by calling it \"availabilities\"? Isn't it availability + the additional checks?
", "rawMarkdown": "lucabergamini , don't you introduce another confusion here by calling it \"availabilities\"? Isn't it availability + the additional checks?", "votes": null }, { "id": "1025392", "postDate": "09/24/2020 14:26:34", "content": "depends what you mean with \"availabilities\" I guess :) Originally the array stored on disk was also 1D and there were no checks there..
", "rawMarkdown": "depends what you mean with \"availabilities\" I guess :) Originally the array stored on disk was also 1D and there were no checks there..", "votes": null }, { "id": "1025436", "postDate": "09/24/2020 14:48:45", "content": "The fields target_availabilities and history_availabilities which are generated by _create_targets_for_deep_prediction and based solely on presence in a frame in zarr arrays is what naturally we call \"availability\".
On the other hand, mask values are subset of those. To avoid confusion I wouldn't call masks by name \"availability\", it is a reserved word already. Does it make sense? :)
", "rawMarkdown": "The fields `target_availabilities` and `history_availabilities` which are generated by `_create_targets_for_deep_prediction` and based solely on presence in a frame in zarr arrays is what naturally we call \"availability\". \n\nOn the other hand, mask values are subset of those. To avoid confusion I wouldn't call masks by name \"availability\", it is a reserved word already. Does it make sense? :)", "votes": null } ], "comments": [ { "id": 1015964, "author_name": "lucabergamini", "author_url": "", "post_date": "09/18/2020 15:10:08", "content": "agents_mask is a 1D boolean array with the same length of the agent array. If a value is True, that agent will be considered as an element of the dataset, otherwise it will be skipped.
I see. Thanks a lot!
", "votes": null, "replies": [] }, { "id": 1025165, "author_name": "walkpast", "author_url": "", "post_date": "09/24/2020 11:31:39", "content": "@lucabergamini
\nIsn't agents_mask a 2D array? In function load_agents_mask in agent.py file, it returns a numpy.ndarray with shape (1893736, 2) for sample.zarr dataset:
def load_agents_mask(self) -> np.ndarray:\n \"\"\"\n Loads a boolean mask of the agent availability stored into the zarr. Performs some sanity check against cfg.\n Returns: a boolean mask of the same length of the dataset agents\n \"\"\"\n\n ...\n ...\n agents_mask = convenience.load(str(agents_mask_path)) # This returns `(1893736, 2)` for sample.zarr\n return agents_mask\nLater on in the __init__ function of class AgentDataset, the agents_mask is compared against min_frame_history and min_frame_future as follows:
past_mask = agents_mask[:, 0] >= min_frame_history\nfuture_mask = agents_mask[:, 1] >= min_frame_future\nagents_mask = past_mask * future_mask\nI wonder what those two columns of agents_mask correspond to?
You are correct that it is a little bit confusing, agents_mask is different things depending on its shape. If it is 1D boolean vector then what @lucabergamini said. If it is a 2D integer vector then it specifies in how many frames agent is relevant before and after the sample in question (\"relevant\" here is my word). Note that relevant agent frames is not the same as frames where agent is available, read this discussion for explanation of the difference.
Yeah, there's two different stored masks. The agents_mask array in the zarr datasets which is the 2D format and is used internally in AgentDataset to filer frames (and create an internal 1D mask, affected by certain config options). You don't need to manually load this or generally deal with it.
\nThen there's the mask.npz file you load (with numpy.load) and pass in to AgentDataset which is 1D boolean mask with a pre-specified set of frames for which you must make test predictions.
Yes, sorry for the confusion. The mask stored on the disk has future and history availabilities (hence 2D) while the one used at runtime is a 1D boolean array. The first is converted into the second one at runtime here
", "votes": null, "replies": [] }, { "id": 1025388, "author_name": "zaharch", "author_url": "", "post_date": "09/24/2020 14:20:44", "content": "@lucabergamini , don't you introduce another confusion here by calling it \"availabilities\"? Isn't it availability + the additional checks?
", "votes": null, "replies": [] }, { "id": 1025392, "author_name": "lucabergamini", "author_url": "", "post_date": "09/24/2020 14:26:34", "content": "depends what you mean with \"availabilities\" I guess :) Originally the array stored on disk was also 1D and there were no checks there..
", "votes": null, "replies": [] }, { "id": 1025436, "author_name": "zaharch", "author_url": "", "post_date": "09/24/2020 14:48:45", "content": "The fields target_availabilities and history_availabilities which are generated by _create_targets_for_deep_prediction and based solely on presence in a frame in zarr arrays is what naturally we call \"availability\".
On the other hand, mask values are subset of those. To avoid confusion I wouldn't call masks by name \"availability\", it is a reserved word already. Does it make sense? :)
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1015930": "I think this file can help me to understand which columns should be predicted in the output, but I have no idea to deal with it.", "1015964": "`agents_mask` is a 1D boolean array with the same length of the agent array. If a value is True, that agent will be considered as an element of the dataset, otherwise it will be skipped.", "1020584": "I see. Thanks a lot!", "1025165": "lucabergamini\nIsn't `agents_mask` a 2D array? In function `load_agents_mask` in `agent.py` file, it returns a `numpy.ndarray` with shape `(1893736, 2)` for `sample.zarr` dataset:\n\n```python\ndef load_agents_mask(self) -> np.ndarray:\n \"\"\"\n Loads a boolean mask of the agent availability stored into the zarr. Performs some sanity check against cfg.\n Returns: a boolean mask of the same length of the dataset agents\n \"\"\"\n \n ...\n ...\n agents_mask = convenience.load(str(agents_mask_path)) # This returns `(1893736, 2)` for sample.zarr\n return agents_mask\n```\n\nLater on in the `__init__` function of `class AgentDataset`, the `agents_mask` is compared against `min_frame_history` and `min_frame_future` as follows:\n\n```python\npast_mask = agents_mask[:, 0] >= min_frame_history\nfuture_mask = agents_mask[:, 1] >= min_frame_future\nagents_mask = past_mask * future_mask\n```\n\nI wonder what those two columns of `agents_mask` correspond to?", "1025240": "You are correct that it is a little bit confusing, `agents_mask` is different things depending on its shape. If it is 1D boolean vector then what @lucabergamini said. If it is a 2D integer vector then it specifies in how many frames agent is relevant before and after the sample in question (\"relevant\" here is my word). Note that relevant agent frames is not the same as frames where agent is available, read [this discussion](https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/185762#1022114) for explanation of the difference.", "1025303": "Yeah, there's two different stored masks. The `agents_mask` array in the zarr datasets which is the 2D format and is used internally in `AgentDataset` to filer frames (and create an internal 1D mask, affected by certain config options). You don't need to manually load this or generally deal with it.\nThen there's the `mask.npz` file you load (with `numpy.load`) and pass in to `AgentDataset` which is 1D boolean mask with a pre-specified set of frames for which you must make test predictions.", "1025331": "Yes, sorry for the confusion. The mask stored on the disk has future and history availabilities (hence 2D) while the one used at runtime is a 1D boolean array. The first is converted into the second one at runtime [here](https://github.com/lyft/l5kit/blob/90a6109754c8a75199219188dff2e25d1b4489ab/l5kit/l5kit/dataset/agent.py#L38)", "1025388": "lucabergamini , don't you introduce another confusion here by calling it \"availabilities\"? Isn't it availability + the additional checks?", "1025392": "depends what you mean with \"availabilities\" I guess :) Originally the array stored on disk was also 1D and there were no checks there..", "1025436": "The fields `target_availabilities` and `history_availabilities` which are generated by `_create_targets_for_deep_prediction` and based solely on presence in a frame in zarr arrays is what naturally we call \"availability\". \n\nOn the other hand, mask values are subset of those. To avoid confusion I wouldn't call masks by name \"availability\", it is a reserved word already. Does it make sense? :)" }, "source": "meta" }