We have to process 33 hour long test recordings in 2 hours.
\nWould not it make sense to process the recordings real-time on edge devices? If you move them then cheap GPUs would be probably more cost efficient.
", "messages": [ { "id": "2172834", "postDate": "03/07/2023 20:54:02", "content": "We have to process 33 hour long test recordings in 2 hours.
\nWould not it make sense to process the recordings real-time on edge devices? If you move them then cheap GPUs would be probably more cost efficient.
", "rawMarkdown": "We have to process 33 hour long test recordings in 2 hours.\n\nWould not it make sense to process the recordings real-time on edge devices? If you move them then cheap GPUs would be probably more cost efficient.", "votes": null }, { "id": "2172840", "postDate": "03/07/2023 21:04:05", "content": "We want this year's entry to be applicable to actual on-the-ground field work, which means that we want to avoid large ensembles. That's why we limited the total notebook runtime to 120min. That should still be plenty of time to process all soundscapes.
", "rawMarkdown": "We want this year's entry to be applicable to actual on-the-ground field work, which means that we want to avoid large ensembles. That's why we limited the total notebook runtime to 120min. That should still be plenty of time to process all soundscapes.", "votes": null }, { "id": "2172848", "postDate": "03/07/2023 21:13:07", "content": "Yeah it makes sense to punish crazy ensembles a bit. I am curious to see how it works.
\nKagglers can be quite creative how to build overly complex solutions under different constraints :)
We do have a need for speed in conservation. The faster we can analyze data, the better, and I would have been even more restrictive with the runtime ;) Yes, edge computing is another way how to tackle this. However, it has its challenges. Computing requires power, and providing sufficient power for long deployments in remote areas is difficult. In addition, real-time detection only makes sense if you can transmit the data. This increases the complexity and costs of the system, especially in areas where no cell service is available. All that being said, I think edge computing solutions will become a more important tool in the future, and I know a bunch of folks working on such systems at the moment.
", "rawMarkdown": "We do have a need for speed in conservation. The faster we can analyze data, the better, and I would have been even more restrictive with the runtime ;) Yes, edge computing is another way how to tackle this. However, it has its challenges. Computing requires power, and providing sufficient power for long deployments in remote areas is difficult. In addition, real-time detection only makes sense if you can transmit the data. This increases the complexity and costs of the system, especially in areas where no cell service is available. All that being said, I think edge computing solutions will become a more important tool in the future, and I know a bunch of folks working on such systems at the moment.", "votes": null }, { "id": "2176840", "postDate": "03/11/2023 00:40:47", "content": "I've been working on a very similar space for the Department of Conservation, here in New Zealand. These models are really a lot more useful to people if they don't need fancy GPUs, or any GPU at all. Field staff don't have this kind of option, mostly they have some combination of skinny/old laptops. And the budget for building and maintaining things is small. I suspect this is probably true for conservation work the world over.
\nAnd if the models can run on edge devices like a Raspberry Pi or a phone, that's next level useful. But the question of power supply & consumption becomes really important. Nobody seems to be particularly concerned about that last couple of percent that might come from a crazy complex ensemble.
\nSo I think this time limit and CPU only constraint has made this competition much more relevant. I'm really interested to see how it affects the kind of solutions we see from it.
", "rawMarkdown": "I've been working on a very similar space for the Department of Conservation, here in New Zealand. These models are really a lot more useful to people if they don't need fancy GPUs, or any GPU at all. Field staff don't have this kind of option, mostly they have some combination of skinny/old laptops. And the budget for building and maintaining things is small. I suspect this is probably true for conservation work the world over. \n\nAnd if the models can run on edge devices like a Raspberry Pi or a phone, that's next level useful. But the question of power supply & consumption becomes really important. Nobody seems to be particularly concerned about that last couple of percent that might come from a crazy complex ensemble.\n\nSo I think this time limit and CPU only constraint has made this competition much more relevant. I'm really interested to see how it affects the kind of solutions we see from it.", "votes": null } ], "comments": [ { "id": 2172840, "author_name": "stefankahl", "author_url": "", "post_date": "03/07/2023 21:04:05", "content": "We want this year's entry to be applicable to actual on-the-ground field work, which means that we want to avoid large ensembles. That's why we limited the total notebook runtime to 120min. That should still be plenty of time to process all soundscapes.
", "votes": null, "replies": [] }, { "id": 2172848, "author_name": "gaborfodor", "author_url": "", "post_date": "03/07/2023 21:13:07", "content": "Yeah it makes sense to punish crazy ensembles a bit. I am curious to see how it works.
\nKagglers can be quite creative how to build overly complex solutions under different constraints :)
We do have a need for speed in conservation. The faster we can analyze data, the better, and I would have been even more restrictive with the runtime ;) Yes, edge computing is another way how to tackle this. However, it has its challenges. Computing requires power, and providing sufficient power for long deployments in remote areas is difficult. In addition, real-time detection only makes sense if you can transmit the data. This increases the complexity and costs of the system, especially in areas where no cell service is available. All that being said, I think edge computing solutions will become a more important tool in the future, and I know a bunch of folks working on such systems at the moment.
", "votes": null, "replies": [] }, { "id": 2176840, "author_name": "ollypowell", "author_url": "", "post_date": "03/11/2023 00:40:47", "content": "I've been working on a very similar space for the Department of Conservation, here in New Zealand. These models are really a lot more useful to people if they don't need fancy GPUs, or any GPU at all. Field staff don't have this kind of option, mostly they have some combination of skinny/old laptops. And the budget for building and maintaining things is small. I suspect this is probably true for conservation work the world over.
\nAnd if the models can run on edge devices like a Raspberry Pi or a phone, that's next level useful. But the question of power supply & consumption becomes really important. Nobody seems to be particularly concerned about that last couple of percent that might come from a crazy complex ensemble.
\nSo I think this time limit and CPU only constraint has made this competition much more relevant. I'm really interested to see how it affects the kind of solutions we see from it.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2172834": "We have to process 33 hour long test recordings in 2 hours.\n\nWould not it make sense to process the recordings real-time on edge devices? If you move them then cheap GPUs would be probably more cost efficient.", "2172840": "We want this year's entry to be applicable to actual on-the-ground field work, which means that we want to avoid large ensembles. That's why we limited the total notebook runtime to 120min. That should still be plenty of time to process all soundscapes.", "2172848": "Yeah it makes sense to punish crazy ensembles a bit. I am curious to see how it works.\nKagglers can be quite creative how to build overly complex solutions under different constraints :)", "2172894": "We do have a need for speed in conservation. The faster we can analyze data, the better, and I would have been even more restrictive with the runtime ;) Yes, edge computing is another way how to tackle this. However, it has its challenges. Computing requires power, and providing sufficient power for long deployments in remote areas is difficult. In addition, real-time detection only makes sense if you can transmit the data. This increases the complexity and costs of the system, especially in areas where no cell service is available. All that being said, I think edge computing solutions will become a more important tool in the future, and I know a bunch of folks working on such systems at the moment.", "2176840": "I've been working on a very similar space for the Department of Conservation, here in New Zealand. These models are really a lot more useful to people if they don't need fancy GPUs, or any GPU at all. Field staff don't have this kind of option, mostly they have some combination of skinny/old laptops. And the budget for building and maintaining things is small. I suspect this is probably true for conservation work the world over. \n\nAnd if the models can run on edge devices like a Raspberry Pi or a phone, that's next level useful. But the question of power supply & consumption becomes really important. Nobody seems to be particularly concerned about that last couple of percent that might come from a crazy complex ensemble.\n\nSo I think this time limit and CPU only constraint has made this competition much more relevant. I'm really interested to see how it affects the kind of solutions we see from it." }, "source": "meta" }