This competition has more of a academic/research setup compared with other featured Kaggle competitions. Looks like the Open Images dataset will be a benchmark, just like MS COCO or Pascal VOC. With that said, I am sure a lot of the research labs will be in the run.
\n\nFor the newcomers to the field of Computer Vision (including myself), could the more experienced researcher/competitor advise on the hardware setup for data of this magnitude? What kind of training time (full training, not transfer learning) are we talking about, say, on a single 1080 Ti or Titan V?
", "messages": [ { "id": "353046", "postDate": "07/05/2018 18:21:04", "content": "This competition has more of a academic/research setup compared with other featured Kaggle competitions. Looks like the Open Images dataset will be a benchmark, just like MS COCO or Pascal VOC. With that said, I am sure a lot of the research labs will be in the run.
\n\nFor the newcomers to the field of Computer Vision (including myself), could the more experienced researcher/competitor advise on the hardware setup for data of this magnitude? What kind of training time (full training, not transfer learning) are we talking about, say, on a single 1080 Ti or Titan V?
", "rawMarkdown": "This competition has more of a academic/research setup compared with other featured Kaggle competitions. Looks like the Open Images dataset will be a benchmark, just like MS COCO or Pascal VOC. With that said, I am sure a lot of the research labs will be in the run.\n\nFor the newcomers to the field of Computer Vision (including myself), could the more experienced researcher/competitor advise on the hardware setup for data of this magnitude? What kind of training time (full training, not transfer learning) are we talking about, say, on a single 1080 Ti or Titan V?", "votes": null }, { "id": "353082", "postDate": "07/05/2018 21:17:35", "content": "I don't think it is feasible to train from scratch on a single GPU, even for object recognition (a much simpler task) and even on a smaller dataset. DawnBench results provide a good idea of what can be achieved with what sort of hardware. Mind you, some of the teams there are using novel and advanced techniques to speed up training, including leveraging fp16, unusual learning rate schedules, very specific data augmentation, etc. I suspect training on imagenet to good results on a single GPU with some of the newer techniques should be doable but I am not sure if it is worth the effort that would be involved to make this happen. And for how long your GPU would be blocked as you try to iron out all the issues.
\n\nIn many ways, starting with transfer learning does not take anything away from the challenge, especially with limited hardware. There are so many interesting things to be figured out here and I am already quite looking forward to reading the write ups.
", "rawMarkdown": "I don't think it is feasible to train from scratch on a single GPU, even for object recognition (a much simpler task) and even on a smaller dataset. [DawnBench results][1] provide a good idea of what can be achieved with what sort of hardware. Mind you, some of the teams there are using novel and advanced techniques to speed up training, including leveraging fp16, unusual learning rate schedules, very specific data augmentation, etc. I suspect training on imagenet to good results on a single GPU with some of the newer techniques should be doable but I am not sure if it is worth the effort that would be involved to make this happen. And for how long your GPU would be blocked as you try to iron out all the issues.\n\nIn many ways, starting with transfer learning does not take anything away from the challenge, especially with limited hardware. There are so many interesting things to be figured out here and I am already quite looking forward to reading the write ups. \n\n [1]: https://dawn.cs.stanford.edu/benchmark/ImageNet/train.html", "votes": null }, { "id": "353087", "postDate": "07/05/2018 21:52:02", "content": "Thanks @radek. This was the response I was looking for.
\n\nNothing against transfer learning. The reason that I specify \"training from scratch\" was because I feel it is required in order to be competitive in this competition.
", "rawMarkdown": "Thanks @radek. This was the response I was looking for.\n\nNothing against transfer learning. The reason that I specify \"training from scratch\" was because I feel it is required in order to be competitive in this competition.", "votes": null }, { "id": "354650", "postDate": "07/10/2018 01:54:11", "content": "where did you get all this transfered learning
", "rawMarkdown": "where did you get all this transfered learning", "votes": null }, { "id": "356096", "postDate": "07/12/2018 21:31:04", "content": "Relatedly, I'd like to see how far some cloud credits would go. If I were to store all the data in a GCP bucket and run training in a decently powerful VM (say a using a TPU resource), I'm trying to figure out if I could submit a good effort to this using the $500 GCP credits, or if I'd burn through the credits before I could iterate to a good result. How many hours is it taking others to train something like YOLO-v2 on this whole dataset using a powerful cloud service?
", "rawMarkdown": "Relatedly, I'd like to see how far some cloud credits would go. If I were to store all the data in a GCP bucket and run training in a decently powerful VM (say a using a TPU resource), I'm trying to figure out if I could submit a good effort to this using the $500 GCP credits, or if I'd burn through the credits before I could iterate to a good result. How many hours is it taking others to train something like YOLO-v2 on this whole dataset using a powerful cloud service?", "votes": null } ], "comments": [ { "id": 353082, "author_name": "radek1", "author_url": "", "post_date": "07/05/2018 21:17:35", "content": "I don't think it is feasible to train from scratch on a single GPU, even for object recognition (a much simpler task) and even on a smaller dataset. DawnBench results provide a good idea of what can be achieved with what sort of hardware. Mind you, some of the teams there are using novel and advanced techniques to speed up training, including leveraging fp16, unusual learning rate schedules, very specific data augmentation, etc. I suspect training on imagenet to good results on a single GPU with some of the newer techniques should be doable but I am not sure if it is worth the effort that would be involved to make this happen. And for how long your GPU would be blocked as you try to iron out all the issues.
\n\nIn many ways, starting with transfer learning does not take anything away from the challenge, especially with limited hardware. There are so many interesting things to be figured out here and I am already quite looking forward to reading the write ups.
", "votes": null, "replies": [ { "id": 353087, "author_name": "sijunhe9248", "author_url": "", "post_date": "07/05/2018 21:52:02", "content": "Thanks @radek. This was the response I was looking for.
\n\nNothing against transfer learning. The reason that I specify \"training from scratch\" was because I feel it is required in order to be competitive in this competition.
", "votes": null, "replies": [] }, { "id": 354650, "author_name": "remrem08007", "author_url": "", "post_date": "07/10/2018 01:54:11", "content": "where did you get all this transfered learning
", "votes": null, "replies": [] } ] }, { "id": 356096, "author_name": "thenathanielw", "author_url": "", "post_date": "07/12/2018 21:31:04", "content": "Relatedly, I'd like to see how far some cloud credits would go. If I were to store all the data in a GCP bucket and run training in a decently powerful VM (say a using a TPU resource), I'm trying to figure out if I could submit a good effort to this using the $500 GCP credits, or if I'd burn through the credits before I could iterate to a good result. How many hours is it taking others to train something like YOLO-v2 on this whole dataset using a powerful cloud service?
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "353046": "This competition has more of a academic/research setup compared with other featured Kaggle competitions. Looks like the Open Images dataset will be a benchmark, just like MS COCO or Pascal VOC. With that said, I am sure a lot of the research labs will be in the run.\n\nFor the newcomers to the field of Computer Vision (including myself), could the more experienced researcher/competitor advise on the hardware setup for data of this magnitude? What kind of training time (full training, not transfer learning) are we talking about, say, on a single 1080 Ti or Titan V?", "353082": "I don't think it is feasible to train from scratch on a single GPU, even for object recognition (a much simpler task) and even on a smaller dataset. [DawnBench results][1] provide a good idea of what can be achieved with what sort of hardware. Mind you, some of the teams there are using novel and advanced techniques to speed up training, including leveraging fp16, unusual learning rate schedules, very specific data augmentation, etc. I suspect training on imagenet to good results on a single GPU with some of the newer techniques should be doable but I am not sure if it is worth the effort that would be involved to make this happen. And for how long your GPU would be blocked as you try to iron out all the issues.\n\nIn many ways, starting with transfer learning does not take anything away from the challenge, especially with limited hardware. There are so many interesting things to be figured out here and I am already quite looking forward to reading the write ups. \n\n [1]: https://dawn.cs.stanford.edu/benchmark/ImageNet/train.html", "353087": "Thanks @radek. This was the response I was looking for.\n\nNothing against transfer learning. The reason that I specify \"training from scratch\" was because I feel it is required in order to be competitive in this competition.", "354650": "where did you get all this transfered learning", "356096": "Relatedly, I'd like to see how far some cloud credits would go. If I were to store all the data in a GCP bucket and run training in a decently powerful VM (say a using a TPU resource), I'm trying to figure out if I could submit a good effort to this using the $500 GCP credits, or if I'd burn through the credits before I could iterate to a good result. How many hours is it taking others to train something like YOLO-v2 on this whole dataset using a powerful cloud service?" }, "source": "meta" }