You are limited to 9 hours of GPU compute time. This constraint is also imposed on the Public Test Set re-run. You should anticipate the Public Test set to be 10 times the size of the Public Validation Set and budget accordingly.
This should be interesting
", "messages": [ { "id": "693076", "postDate": "12/12/2019 02:29:17", "content": "You are limited to 9 hours of GPU compute time. This constraint is also imposed on the Public Test Set re-run. You should anticipate the Public Test set to be 10 times the size of the Public Validation Set and budget accordingly.
This should be interesting
", "rawMarkdown": "```You are limited to 9 hours of GPU compute time. This constraint is also imposed on the Public Test Set re-run. You should anticipate the Public Test set to be 10 times the size of the Public Validation Set and budget accordingly.```\n\nThis should be interesting", "votes": null }, { "id": "693192", "postDate": "12/12/2019 05:59:51", "content": "I foresee many timeout errors happening in the private leaderboard due to this
", "rawMarkdown": "I foresee many timeout errors happening in the private leaderboard due to this", "votes": null }, { "id": "693252", "postDate": "12/12/2019 07:24:55", "content": "I think that a bigger problem is in 30 hours of GPU per week
", "rawMarkdown": "I think that a bigger problem is in 30 hours of GPU per week", "votes": null }, { "id": "693372", "postDate": "12/12/2019 10:07:43", "content": "Could this be telling us that we can only train this on our own HW/Cloud and upload trained model? No way training could be done in kaggle GPU quota, unless someone has enough patience to save/load models again and again...
", "rawMarkdown": "Could this be telling us that we can only train this on our own HW/Cloud and upload trained model? No way training could be done in kaggle GPU quota, unless someone has enough patience to save/load models again and again...", "votes": null }, { "id": "693417", "postDate": "12/12/2019 10:50:19", "content": "\n\n\nAndrew Lukyanenko wrote:
\n \nI think that a bigger problem is in 30 hours of GPU per week
\n
And sometimes it's less than 30 hours
", "rawMarkdown": "> **Andrew Lukyanenko wrote:**\n> \n> I think that a bigger problem is in 30 hours of GPU per week\n\nAnd sometimes it's less than 30 hours", "votes": null }, { "id": "693419", "postDate": "12/12/2019 10:54:39", "content": "As much as I love Kaggle, the kernels for me are still very flaky-I try to avoid using Kernels as much as possible.
", "rawMarkdown": "As much as I love Kaggle, the kernels for me are still very flaky-I try to avoid using Kernels as much as possible.", "votes": null }, { "id": "693567", "postDate": "12/12/2019 14:14:19", "content": "It is similar to Severstal competition, size of full data was 3x time of public data, we had 1 hour kernel time so we tried to fit public data in about 20 minutes.
", "rawMarkdown": "It is similar to Severstal competition, size of full data was 3x time of public data, we had 1 hour kernel time so we tried to fit public data in about 20 minutes.", "votes": null }, { "id": "693574", "postDate": "12/12/2019 14:22:10", "content": "Yes Off-line Training is allowed, please check the requirements section. There are a few inference constraints that have been introduced.
", "rawMarkdown": "Yes Off-line Training is allowed, please check the requirements section. There are a few inference constraints that have been introduced.", "votes": null }, { "id": "693879", "postDate": "12/12/2019 22:04:30", "content": "The 9 hour compute constraint will be applied to the submission in Kaggle, when evaluated against the (hidden) public test set (see Getting Started) For the private test set evaluation, the time constraint will scale with that private test set's size. So if you are able to make a successful submission that posts a score to the public leaderboard, then it is successful on the (hidden) public test set and has met the 9 hour constraint. Because this is done synchronously with your submission, you will receive (nearly) immediate feedback if you've suffered a timeout.
", "rawMarkdown": "The 9 hour compute constraint will be applied to the submission in Kaggle, when evaluated against the (hidden) public test set (see [Getting Started](https://www.kaggle.com/c/deepfake-detection-challenge/overview/getting-started)) For the private test set evaluation, the time constraint will scale with that private test set's size. So if you are able to make a successful submission that posts a score to the public leaderboard, then it is successful on the (hidden) public test set and has met the 9 hour constraint. Because this is done synchronously with your submission, you will receive (nearly) immediate feedback if you've suffered a timeout.", "votes": null }, { "id": "693996", "postDate": "12/13/2019 02:50:23", "content": "@juliaelliott Thanks for the clarifications! Please add these details to the comp section-many people might miss out on this.
", "rawMarkdown": "juliaelliott Thanks for the clarifications! Please add these details to the comp section-many people might miss out on this.", "votes": null }, { "id": "694208", "postDate": "12/13/2019 09:55:02", "content": "@init27 I'm a little confused. Don't the requirements say that any external data uploaded into the Kernel environment (including pretrained models) must be made public? If so, how does offline training make any sense? Is it just to figure out the hyperparameter values?\nWhat am I missing?
", "rawMarkdown": "init27 I'm a little confused. Don't the requirements say that any external data uploaded into the Kernel environment (including pretrained models) must be made public? If so, how does offline training make any sense? Is it just to figure out the hyperparameter values?\nWhat am I missing?", "votes": null }, { "id": "694282", "postDate": "12/13/2019 11:41:56", "content": "Does anyone know the numbers of GPU and CPU available during inference ?
", "rawMarkdown": "Does anyone know the numbers of GPU and CPU available during inference ?", "votes": null }, { "id": "694284", "postDate": "12/13/2019 11:43:19", "content": "Hi , how many GPUs and CPUs are available during one inference ?
", "rawMarkdown": "Hi , how many GPUs and CPUs are available during one inference ?", "votes": null }, { "id": "694496", "postDate": "12/13/2019 17:45:12", "content": "Yes, offline training is both allowed and encouraged, due to the train set’s size. Please read the Getting Started information which details this.
\n\nAlso, datasets and pretrained models obtained externally need to be declared, but your full self-created/original work offline-trained models do not. See the clarification I’ve made on the official external data thread.
", "rawMarkdown": "Yes, offline training is both allowed and encouraged, due to the train set’s size. Please read the Getting Started information which details this.\n\nAlso, datasets and pretrained models obtained externally need to be declared, but your full self-created/original work offline-trained models do not. See the clarification I’ve made on the official external data thread.", "votes": null }, { "id": "695159", "postDate": "12/14/2019 17:23:52", "content": "Thank you for your response @juliaelliott!
", "rawMarkdown": "Thank you for your response @juliaelliott!", "votes": null }, { "id": "712717", "postDate": "01/07/2020 14:28:50", "content": "@juliaelliott , do i understood correctly, that code requierements are the following: 9 hours for 400 videos (public leaderbord), and 90 hours for 4000 videos (private leaderbord)?
", "rawMarkdown": "juliaelliott , do i understood correctly, that code requierements are the following: 9 hours for 400 videos (public leaderbord), and 90 hours for 4000 videos (private leaderbord)?", "votes": null }, { "id": "712942", "postDate": "01/07/2020 18:40:36", "content": "@alekseykachalov No, for both the validation set (400 videos when you \"Commit\" your notebook) and the public test set (4000 videos when you \"Submit to Competition\" from your notebook), you are required to stay within 9 hours. The private test set evaluation at the end of the competition done outside of Kaggle will scale from 9 hours based on the video count volume, accordingly.
", "rawMarkdown": "alekseykachalov No, for both the validation set (400 videos when you \"Commit\" your notebook) and the public test set (4000 videos when you \"Submit to Competition\" from your notebook), you are required to stay within 9 hours. The private test set evaluation at the end of the competition done outside of Kaggle will scale from 9 hours based on the video count volume, accordingly.", "votes": null } ], "comments": [ { "id": 693192, "author_name": "roryoreilly", "author_url": "", "post_date": "12/12/2019 05:59:51", "content": "I foresee many timeout errors happening in the private leaderboard due to this
", "votes": null, "replies": [ { "id": 693879, "author_name": "juliaelliott", "author_url": "", "post_date": "12/12/2019 22:04:30", "content": "The 9 hour compute constraint will be applied to the submission in Kaggle, when evaluated against the (hidden) public test set (see Getting Started) For the private test set evaluation, the time constraint will scale with that private test set's size. So if you are able to make a successful submission that posts a score to the public leaderboard, then it is successful on the (hidden) public test set and has met the 9 hour constraint. Because this is done synchronously with your submission, you will receive (nearly) immediate feedback if you've suffered a timeout.
", "votes": null, "replies": [] }, { "id": 693996, "author_name": "init27", "author_url": "", "post_date": "12/13/2019 02:50:23", "content": "@juliaelliott Thanks for the clarifications! Please add these details to the comp section-many people might miss out on this.
", "votes": null, "replies": [] }, { "id": 694284, "author_name": "fionalxd", "author_url": "", "post_date": "12/13/2019 11:43:19", "content": "Hi , how many GPUs and CPUs are available during one inference ?
", "votes": null, "replies": [] }, { "id": 712717, "author_name": "alekseykachalov", "author_url": "", "post_date": "01/07/2020 14:28:50", "content": "@juliaelliott , do i understood correctly, that code requierements are the following: 9 hours for 400 videos (public leaderbord), and 90 hours for 4000 videos (private leaderbord)?
", "votes": null, "replies": [] }, { "id": 712942, "author_name": "juliaelliott", "author_url": "", "post_date": "01/07/2020 18:40:36", "content": "@alekseykachalov No, for both the validation set (400 videos when you \"Commit\" your notebook) and the public test set (4000 videos when you \"Submit to Competition\" from your notebook), you are required to stay within 9 hours. The private test set evaluation at the end of the competition done outside of Kaggle will scale from 9 hours based on the video count volume, accordingly.
", "votes": null, "replies": [] } ] }, { "id": 693252, "author_name": "artgor", "author_url": "", "post_date": "12/12/2019 07:24:55", "content": "I think that a bigger problem is in 30 hours of GPU per week
", "votes": null, "replies": [ { "id": 693417, "author_name": "serigne", "author_url": "", "post_date": "12/12/2019 10:50:19", "content": "\n\n\nAndrew Lukyanenko wrote:
\n \nI think that a bigger problem is in 30 hours of GPU per week
\n
And sometimes it's less than 30 hours
", "votes": null, "replies": [] }, { "id": 693419, "author_name": "init27", "author_url": "", "post_date": "12/12/2019 10:54:39", "content": "As much as I love Kaggle, the kernels for me are still very flaky-I try to avoid using Kernels as much as possible.
", "votes": null, "replies": [] } ] }, { "id": 693372, "author_name": "bags8040", "author_url": "", "post_date": "12/12/2019 10:07:43", "content": "Could this be telling us that we can only train this on our own HW/Cloud and upload trained model? No way training could be done in kaggle GPU quota, unless someone has enough patience to save/load models again and again...
", "votes": null, "replies": [ { "id": 693574, "author_name": "init27", "author_url": "", "post_date": "12/12/2019 14:22:10", "content": "Yes Off-line Training is allowed, please check the requirements section. There are a few inference constraints that have been introduced.
", "votes": null, "replies": [] }, { "id": 694208, "author_name": "dranzer", "author_url": "", "post_date": "12/13/2019 09:55:02", "content": "@init27 I'm a little confused. Don't the requirements say that any external data uploaded into the Kernel environment (including pretrained models) must be made public? If so, how does offline training make any sense? Is it just to figure out the hyperparameter values?\nWhat am I missing?
", "votes": null, "replies": [] }, { "id": 694496, "author_name": "juliaelliott", "author_url": "", "post_date": "12/13/2019 17:45:12", "content": "Yes, offline training is both allowed and encouraged, due to the train set’s size. Please read the Getting Started information which details this.
\n\nAlso, datasets and pretrained models obtained externally need to be declared, but your full self-created/original work offline-trained models do not. See the clarification I’ve made on the official external data thread.
", "votes": null, "replies": [] }, { "id": 695159, "author_name": "dranzer", "author_url": "", "post_date": "12/14/2019 17:23:52", "content": "Thank you for your response @juliaelliott!
", "votes": null, "replies": [] } ] }, { "id": 693567, "author_name": "jacekpoplawski", "author_url": "", "post_date": "12/12/2019 14:14:19", "content": "It is similar to Severstal competition, size of full data was 3x time of public data, we had 1 hour kernel time so we tried to fit public data in about 20 minutes.
", "votes": null, "replies": [] }, { "id": 694282, "author_name": "fionalxd", "author_url": "", "post_date": "12/13/2019 11:41:56", "content": "Does anyone know the numbers of GPU and CPU available during inference ?
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "693076": "```You are limited to 9 hours of GPU compute time. This constraint is also imposed on the Public Test Set re-run. You should anticipate the Public Test set to be 10 times the size of the Public Validation Set and budget accordingly.```\n\nThis should be interesting", "693192": "I foresee many timeout errors happening in the private leaderboard due to this", "693252": "I think that a bigger problem is in 30 hours of GPU per week", "693372": "Could this be telling us that we can only train this on our own HW/Cloud and upload trained model? No way training could be done in kaggle GPU quota, unless someone has enough patience to save/load models again and again...", "693417": "> **Andrew Lukyanenko wrote:**\n> \n> I think that a bigger problem is in 30 hours of GPU per week\n\nAnd sometimes it's less than 30 hours", "693419": "As much as I love Kaggle, the kernels for me are still very flaky-I try to avoid using Kernels as much as possible.", "693567": "It is similar to Severstal competition, size of full data was 3x time of public data, we had 1 hour kernel time so we tried to fit public data in about 20 minutes.", "693574": "Yes Off-line Training is allowed, please check the requirements section. There are a few inference constraints that have been introduced.", "693879": "The 9 hour compute constraint will be applied to the submission in Kaggle, when evaluated against the (hidden) public test set (see [Getting Started](https://www.kaggle.com/c/deepfake-detection-challenge/overview/getting-started)) For the private test set evaluation, the time constraint will scale with that private test set's size. So if you are able to make a successful submission that posts a score to the public leaderboard, then it is successful on the (hidden) public test set and has met the 9 hour constraint. Because this is done synchronously with your submission, you will receive (nearly) immediate feedback if you've suffered a timeout.", "693996": "juliaelliott Thanks for the clarifications! Please add these details to the comp section-many people might miss out on this.", "694208": "init27 I'm a little confused. Don't the requirements say that any external data uploaded into the Kernel environment (including pretrained models) must be made public? If so, how does offline training make any sense? Is it just to figure out the hyperparameter values?\nWhat am I missing?", "694282": "Does anyone know the numbers of GPU and CPU available during inference ?", "694284": "Hi , how many GPUs and CPUs are available during one inference ?", "694496": "Yes, offline training is both allowed and encouraged, due to the train set’s size. Please read the Getting Started information which details this.\n\nAlso, datasets and pretrained models obtained externally need to be declared, but your full self-created/original work offline-trained models do not. See the clarification I’ve made on the official external data thread.", "695159": "Thank you for your response @juliaelliott!", "712717": "juliaelliott , do i understood correctly, that code requierements are the following: 9 hours for 400 videos (public leaderbord), and 90 hours for 4000 videos (private leaderbord)?", "712942": "alekseykachalov No, for both the validation set (400 videos when you \"Commit\" your notebook) and the public test set (4000 videos when you \"Submit to Competition\" from your notebook), you are required to stay within 9 hours. The private test set evaluation at the end of the competition done outside of Kaggle will scale from 9 hours based on the video count volume, accordingly." }, "source": "meta" }