I have a question regarding the usage of pretrained models and weights in this competition. According to the rules, publicly accessible external data and pretrained models can be used as long as they are available to all participants.
\nWould it be allowed to train a model or obtain model weights using a local or external GPU environment and then upload them to Kaggle for use in submissions? Technically, if the model or weights are uploaded as a dataset, they would be publicly accessible, but I am concerned that this might contradict the CPU-only restriction of the competition.
\nCan anyone clarify whether this approach is permitted under the competition rules? I’d appreciate any official response or insights from the organizers.
\nThank you!
", "messages": [ { "id": "3154586", "postDate": "03/20/2025 05:51:37", "content": "I have a question regarding the usage of pretrained models and weights in this competition. According to the rules, publicly accessible external data and pretrained models can be used as long as they are available to all participants.
\nWould it be allowed to train a model or obtain model weights using a local or external GPU environment and then upload them to Kaggle for use in submissions? Technically, if the model or weights are uploaded as a dataset, they would be publicly accessible, but I am concerned that this might contradict the CPU-only restriction of the competition.
\nCan anyone clarify whether this approach is permitted under the competition rules? I’d appreciate any official response or insights from the organizers.
\nThank you!
", "rawMarkdown": "I have a question regarding the usage of pretrained models and weights in this competition. According to the rules, publicly accessible external data and pretrained models can be used as long as they are available to all participants.\n\nWould it be allowed to train a model or obtain model weights using a local or external GPU environment and then upload them to Kaggle for use in submissions? Technically, if the model or weights are uploaded as a dataset, they would be publicly accessible, but I am concerned that this might contradict the CPU-only restriction of the competition.\n\nCan anyone clarify whether this approach is permitted under the competition rules? I’d appreciate any official response or insights from the organizers.\n\nThank you!", "votes": null }, { "id": "3154915", "postDate": "03/20/2025 14:15:50", "content": "You do not have to share your model weights or share any external data. If you win prizes you will be asked to show external data at that time and private data might lose you a prize. The same with a pre-trained model - lets say you work at Cornell and have access to a model that is not available to the public - you win a prize and will likely lose it if you used a model created by others, not available publicly. Or your a student at Sejong and you have access to data that was used in a friends Doctoral - the data is not shared in the public, University and Chinese government provided money to fund the study, it was massive and contains lots of data. In a few months it might be made public - that's the external data that is not to be used.
\nI am training model locally right now and in about an hour I will put it in a private data set and make a submission. Like many of the shared models it uses B0 as its starting point. We are not required to make public any weights that we generated.
\nHost might reply but this question has been asked and answered in almost every competition I have been in for the past 7 years. They keep trying little tweaks to the wording to make it easier to understand, but.. …
", "rawMarkdown": "You do not have to share your model weights or share any external data. If you win prizes you will be asked to show external data at that time and private data might lose you a prize. The same with a pre-trained model - lets say you work at Cornell and have access to a model that is not available to the public - you win a prize and will likely lose it if you used a model created by others, not available publicly. Or your a student at Sejong and you have access to data that was used in a friends Doctoral - the data is not shared in the public, University and Chinese government provided money to fund the study, it was massive and contains lots of data. In a few months it might be made public - that's the external data that is not to be used.\n\nI am training model locally right now and in about an hour I will put it in a private data set and make a submission. Like many of the shared models it uses B0 as its starting point. We are not required to make public any weights that we generated.\n\nHost might reply but this question has been asked and answered in almost every competition I have been in for the past 7 years. They keep trying little tweaks to the wording to make it easier to understand, but.. ...", "votes": null }, { "id": "3155420", "postDate": "03/21/2025 02:56:52", "content": "Thank you for the kind explanation!
\nIf I understood correctly, it's acceptable to train a model locally using GPU and then upload the trained model and its weights as a private dataset for use in the competition.
In that case, I’m wondering — why is there a 90-minute CPU time limit enforced in the competition notebooks?
\nIf everyone is allowed to use their own GPUs locally, doesn't that make the GPU usage restriction in the competition environment somewhat meaningless?
Part of the objective seems to be a fast and light model rather than a monster or a very big ensemble of models. Not sure I understand the logic :)
", "rawMarkdown": "Part of the objective seems to be a fast and light model rather than a monster or a very big ensemble of models. Not sure I understand the logic :)", "votes": null }, { "id": "3156205", "postDate": "03/21/2025 21:54:34", "content": "Expensive models are expensive to run, and it is fairly easy to accumulate vast amounts of audio data once a passive acoustic monitoring system is in place…
\nGroups we work with are often under-funded conservation organizations, and spending money on field work is generally a better use of resources than spending it on cloud credits for larger models. Thus, we ultimately want to find lighter-weight methods for handling domain transfer problems.
\nAdditionally, some aplications are moving towards on-device inference, which brings really extreme inference limitations. These limitations are driven by power restrictions, rather than raw compute, but, again, targeting smaller models is ultimately better for these applications.
", "rawMarkdown": "Expensive models are expensive to run, and it is fairly easy to accumulate vast amounts of audio data once a passive acoustic monitoring system is in place... \n\nGroups we work with are often under-funded conservation organizations, and spending money on field work is generally a better use of resources than spending it on cloud credits for larger models. Thus, we ultimately want to find lighter-weight methods for handling domain transfer problems.\n\nAdditionally, some aplications are moving towards on-device inference, which brings really extreme inference limitations. These limitations are driven by power restrictions, rather than raw compute, but, again, targeting smaller models is ultimately better for these applications.", "votes": null }, { "id": "3156424", "postDate": "03/22/2025 05:56:32", "content": "I have trained my model . Now i am getting error of Timeout. How much time for test data prediction we have to in submission. after 29 minutes I got error of Timeout for my notebook.should i change my training startegy to decrease preprocessing and prediction time.
", "rawMarkdown": "I have trained my model . Now i am getting error of Timeout. How much time for test data prediction we have to in submission. after 29 minutes I got error of Timeout for my notebook.should i change my training startegy to decrease preprocessing and prediction time.", "votes": null }, { "id": "3156743", "postDate": "03/22/2025 13:33:31", "content": "hmm per the Overview you should have 90 minutes.
\nSubmissions to this competition must be made through Notebooks. For the \"Submit\" button to be active after a commit, the following conditions must be met:
CPU Notebook <= 90 minutes run-time
\nGPU Notebook submissions are disabled. You can technically submit but will only have 1 minute of runtime.
If you have timeout after 29 that sounds strange? Where are you seeing the 29 minutes stated? For my timeout error I don't see where the run time was stated.
", "rawMarkdown": "hmm per the Overview you should have 90 minutes.\nSubmissions to this competition must be made through Notebooks. For the \"Submit\" button to be active after a commit, the following conditions must be met:\n\nCPU Notebook <= 90 minutes run-time\nGPU Notebook submissions are disabled. You can technically submit but will only have 1 minute of runtime.\n\nIf you have timeout after 29 that sounds strange? Where are you seeing the 29 minutes stated? For my timeout error I don't see where the run time was stated.", "votes": null }, { "id": "3156753", "postDate": "03/22/2025 13:54:26", "content": "Your are right. I submit using GPU. We can not use GPU. But i think that 29 minutes was time of notebook which i mistook.If preporcessing and prediction on cpu in is within one 90 minutes will i able to submit it then.
", "rawMarkdown": "Your are right. I submit using GPU. We can not use GPU. But i think that 29 minutes was time of notebook which i mistook.If preporcessing and prediction on cpu in is within one 90 minutes will i able to submit it then.", "votes": null }, { "id": "3156798", "postDate": "03/22/2025 14:54:37", "content": "Yes - 90 minutes is not much time. When I try to add too much TTA to the prediction I get the time out error. You can add some code to do a bit of estimation. In my case i have it run 10 of the train soundscapes to get a time estimate.
", "rawMarkdown": "Yes - 90 minutes is not much time. When I try to add too much TTA to the prediction I get the time out error. You can add some code to do a bit of estimation. In my case i have it run 10 of the train soundscapes to get a time estimate.", "votes": null } ], "comments": [ { "id": 3154915, "author_name": "pcjimmmy", "author_url": "", "post_date": "03/20/2025 14:15:50", "content": "You do not have to share your model weights or share any external data. If you win prizes you will be asked to show external data at that time and private data might lose you a prize. The same with a pre-trained model - lets say you work at Cornell and have access to a model that is not available to the public - you win a prize and will likely lose it if you used a model created by others, not available publicly. Or your a student at Sejong and you have access to data that was used in a friends Doctoral - the data is not shared in the public, University and Chinese government provided money to fund the study, it was massive and contains lots of data. In a few months it might be made public - that's the external data that is not to be used.
\nI am training model locally right now and in about an hour I will put it in a private data set and make a submission. Like many of the shared models it uses B0 as its starting point. We are not required to make public any weights that we generated.
\nHost might reply but this question has been asked and answered in almost every competition I have been in for the past 7 years. They keep trying little tweaks to the wording to make it easier to understand, but.. …
", "votes": null, "replies": [ { "id": 3155420, "author_name": "kimbumju", "author_url": "", "post_date": "03/21/2025 02:56:52", "content": "Thank you for the kind explanation!
\nIf I understood correctly, it's acceptable to train a model locally using GPU and then upload the trained model and its weights as a private dataset for use in the competition.
In that case, I’m wondering — why is there a 90-minute CPU time limit enforced in the competition notebooks?
\nIf everyone is allowed to use their own GPUs locally, doesn't that make the GPU usage restriction in the competition environment somewhat meaningless?
Part of the objective seems to be a fast and light model rather than a monster or a very big ensemble of models. Not sure I understand the logic :)
", "votes": null, "replies": [ { "id": 3156205, "author_name": "tomdenton", "author_url": "", "post_date": "03/21/2025 21:54:34", "content": "Expensive models are expensive to run, and it is fairly easy to accumulate vast amounts of audio data once a passive acoustic monitoring system is in place…
\nGroups we work with are often under-funded conservation organizations, and spending money on field work is generally a better use of resources than spending it on cloud credits for larger models. Thus, we ultimately want to find lighter-weight methods for handling domain transfer problems.
\nAdditionally, some aplications are moving towards on-device inference, which brings really extreme inference limitations. These limitations are driven by power restrictions, rather than raw compute, but, again, targeting smaller models is ultimately better for these applications.
", "votes": null, "replies": [ { "id": 3156424, "author_name": "", "author_url": "", "post_date": "03/22/2025 05:56:32", "content": "I have trained my model . Now i am getting error of Timeout. How much time for test data prediction we have to in submission. after 29 minutes I got error of Timeout for my notebook.should i change my training startegy to decrease preprocessing and prediction time.
", "votes": null, "replies": [ { "id": 3156743, "author_name": "pcjimmmy", "author_url": "", "post_date": "03/22/2025 13:33:31", "content": "hmm per the Overview you should have 90 minutes.
\nSubmissions to this competition must be made through Notebooks. For the \"Submit\" button to be active after a commit, the following conditions must be met:
CPU Notebook <= 90 minutes run-time
\nGPU Notebook submissions are disabled. You can technically submit but will only have 1 minute of runtime.
If you have timeout after 29 that sounds strange? Where are you seeing the 29 minutes stated? For my timeout error I don't see where the run time was stated.
", "votes": null, "replies": [ { "id": 3156753, "author_name": "", "author_url": "", "post_date": "03/22/2025 13:54:26", "content": "Your are right. I submit using GPU. We can not use GPU. But i think that 29 minutes was time of notebook which i mistook.If preporcessing and prediction on cpu in is within one 90 minutes will i able to submit it then.
", "votes": null, "replies": [ { "id": 3156798, "author_name": "pcjimmmy", "author_url": "", "post_date": "03/22/2025 14:54:37", "content": "Yes - 90 minutes is not much time. When I try to add too much TTA to the prediction I get the time out error. You can add some code to do a bit of estimation. In my case i have it run 10 of the train soundscapes to get a time estimate.
", "votes": null, "replies": [] } ] } ] } ] } ] } ] } ] } ] } ], "raw_markdown_by_id": { "3154586": "I have a question regarding the usage of pretrained models and weights in this competition. According to the rules, publicly accessible external data and pretrained models can be used as long as they are available to all participants.\n\nWould it be allowed to train a model or obtain model weights using a local or external GPU environment and then upload them to Kaggle for use in submissions? Technically, if the model or weights are uploaded as a dataset, they would be publicly accessible, but I am concerned that this might contradict the CPU-only restriction of the competition.\n\nCan anyone clarify whether this approach is permitted under the competition rules? I’d appreciate any official response or insights from the organizers.\n\nThank you!", "3154915": "You do not have to share your model weights or share any external data. If you win prizes you will be asked to show external data at that time and private data might lose you a prize. The same with a pre-trained model - lets say you work at Cornell and have access to a model that is not available to the public - you win a prize and will likely lose it if you used a model created by others, not available publicly. Or your a student at Sejong and you have access to data that was used in a friends Doctoral - the data is not shared in the public, University and Chinese government provided money to fund the study, it was massive and contains lots of data. In a few months it might be made public - that's the external data that is not to be used.\n\nI am training model locally right now and in about an hour I will put it in a private data set and make a submission. Like many of the shared models it uses B0 as its starting point. We are not required to make public any weights that we generated.\n\nHost might reply but this question has been asked and answered in almost every competition I have been in for the past 7 years. They keep trying little tweaks to the wording to make it easier to understand, but.. ...", "3155420": "Thank you for the kind explanation!\nIf I understood correctly, it's acceptable to train a model locally using GPU and then upload the trained model and its weights as a private dataset for use in the competition.\n\nIn that case, I’m wondering — why is there a 90-minute CPU time limit enforced in the competition notebooks?\nIf everyone is allowed to use their own GPUs locally, doesn't that make the GPU usage restriction in the competition environment somewhat meaningless?", "3155442": "Part of the objective seems to be a fast and light model rather than a monster or a very big ensemble of models. Not sure I understand the logic :)", "3156205": "Expensive models are expensive to run, and it is fairly easy to accumulate vast amounts of audio data once a passive acoustic monitoring system is in place... \n\nGroups we work with are often under-funded conservation organizations, and spending money on field work is generally a better use of resources than spending it on cloud credits for larger models. Thus, we ultimately want to find lighter-weight methods for handling domain transfer problems.\n\nAdditionally, some aplications are moving towards on-device inference, which brings really extreme inference limitations. These limitations are driven by power restrictions, rather than raw compute, but, again, targeting smaller models is ultimately better for these applications.", "3156424": "I have trained my model . Now i am getting error of Timeout. How much time for test data prediction we have to in submission. after 29 minutes I got error of Timeout for my notebook.should i change my training startegy to decrease preprocessing and prediction time.", "3156743": "hmm per the Overview you should have 90 minutes.\nSubmissions to this competition must be made through Notebooks. For the \"Submit\" button to be active after a commit, the following conditions must be met:\n\nCPU Notebook <= 90 minutes run-time\nGPU Notebook submissions are disabled. You can technically submit but will only have 1 minute of runtime.\n\nIf you have timeout after 29 that sounds strange? Where are you seeing the 29 minutes stated? For my timeout error I don't see where the run time was stated.", "3156753": "Your are right. I submit using GPU. We can not use GPU. But i think that 29 minutes was time of notebook which i mistook.If preporcessing and prediction on cpu in is within one 90 minutes will i able to submit it then.", "3156798": "Yes - 90 minutes is not much time. When I try to add too much TTA to the prediction I get the time out error. You can add some code to do a bit of estimation. In my case i have it run 10 of the train soundscapes to get a time estimate." }, "source": "meta" }