hi,i see few good kernels written using kaggle kernels are not getting good score so i would like to know who got impressive lb score using kaggle kernels only?please comment your lb score using kaggle kernels only,thank you in advance
\n\nhi,i see few good kernels written using kaggle kernels are not getting good score so i would like to know who got impressive lb score using kaggle kernels only?please comment your lb score using kaggle kernels only,thank you in advance
\n\nAs far as I know, our 0.929 is essentially all from Kaggle kernels. There are a lot of kernels involved, taking advantage of the relatively liberal GPU availability prior to the introduction of quotas. And our attempts to use other platforms have mostly failed. (The rtx2060 on my laptop keeps crashing with \"cuDNN failed to get convolution algorithm.\" My desktop is running unbearably slow because only 6GB of GPU memory means I have to run batchsize=8, and it hasn't produced anything useful yet. I haven't managed to get the TPU code working on Google Colab. And we haven't managed to set up GCP to run our models yet. I guess if the quotas had been around since the beginning of the competition, we would have put more effort into these alternatives.)
", "rawMarkdown": "As far as I know, our 0.929 is essentially all from Kaggle kernels. There are a _lot_ of kernels involved, taking advantage of the relatively liberal GPU availability prior to the introduction of quotas. And our attempts to use other platforms have mostly failed. (The rtx2060 on my laptop keeps crashing with \"cuDNN failed to get convolution algorithm.\" My desktop is running unbearably slow because only 6GB of GPU memory means I have to run batchsize=8, and it hasn't produced anything useful yet. I haven't managed to get the TPU code working on Google Colab. And we haven't managed to set up GCP to run our models yet. I guess if the quotas had been around since the beginning of the competition, we would have put more effort into these alternatives.)", "votes": null }, { "id": "631528", "postDate": "09/22/2019 07:23:52", "content": "hey,did you get 0.929 by computing your model in kaggle kernel? this is really very impressive and seems very difficult with 6 channels,thanks for letting me know,wish you luck @aharless
", "rawMarkdown": "hey,did you get 0.929 by computing your model in kaggle kernel? this is really very impressive and seems very difficult with 6 channels,thanks for letting me know,wish you luck @aharless", "votes": null }, { "id": "631680", "postDate": "09/22/2019 13:14:54", "content": "There are models that run on different folds, each run in a separate kernel, and then saving weights and continuing in a new kernel with the saved weights, sometimes re-randomizing a subset of the weights. And there are a number of post-processing steps that run in separate kernels. And some feedback from final predictions back into new versions of the models. So a lot of kernels that each take output from several other kernels as input. And the submissions are directly from kernels.
", "rawMarkdown": "There are models that run on different folds, each run in a separate kernel, and then saving weights and continuing in a new kernel with the saved weights, sometimes re-randomizing a subset of the weights. And there are a number of post-processing steps that run in separate kernels. And some feedback from final predictions back into new versions of the models. So a lot of kernels that each take output from several other kernels as input. And the submissions are directly from kernels.", "votes": null }, { "id": "631729", "postDate": "09/22/2019 15:02:47", "content": "It’s a long way from just one or two kernels (as in the public ones, which don’t carry the chain very far) to having a whole web of kernels that use one another’s output and apply multiple tricks (whereas with public kernels that demonstrated particular tricks, most didn’t combine multiple tricks).
", "rawMarkdown": "It’s a long way from just one or two kernels (as in the public ones, which don’t carry the chain very far) to having a whole web of kernels that use one another’s output and apply multiple tricks (whereas with public kernels that demonstrated particular tricks, most didn’t combine multiple tricks).", "votes": null }, { "id": "631784", "postDate": "09/22/2019 16:51:03", "content": "wow!!! really very interesting,i am desperately waiting to see your solution as soon as this competition ends
", "rawMarkdown": "wow!!! really very interesting,i am desperately waiting to see your solution as soon as this competition ends", "votes": null }, { "id": "632066", "postDate": "09/23/2019 07:21:24", "content": "Yeah, Kaggle used to be a great platform to experiment ideas, before the weekly quota! 😿 \nLuckily we still had time to try a few... we'll make some of them public here once the comp is over\nCheers!
", "rawMarkdown": "Yeah, Kaggle used to be a great platform to experiment ideas, before the weekly quota! 😿 \nLuckily we still had time to try a few... we'll make some of them public here once the comp is over\nCheers!", "votes": null }, { "id": "632099", "postDate": "09/23/2019 08:19:59", "content": "thank you a lot @hmendonca
", "rawMarkdown": "thank you a lot @hmendonca", "votes": null } ], "comments": [ { "id": 631446, "author_name": "aharless", "author_url": "", "post_date": "09/22/2019 03:13:28", "content": "As far as I know, our 0.929 is essentially all from Kaggle kernels. There are a lot of kernels involved, taking advantage of the relatively liberal GPU availability prior to the introduction of quotas. And our attempts to use other platforms have mostly failed. (The rtx2060 on my laptop keeps crashing with \"cuDNN failed to get convolution algorithm.\" My desktop is running unbearably slow because only 6GB of GPU memory means I have to run batchsize=8, and it hasn't produced anything useful yet. I haven't managed to get the TPU code working on Google Colab. And we haven't managed to set up GCP to run our models yet. I guess if the quotas had been around since the beginning of the competition, we would have put more effort into these alternatives.)
", "votes": null, "replies": [ { "id": 631528, "author_name": "mobassir", "author_url": "", "post_date": "09/22/2019 07:23:52", "content": "hey,did you get 0.929 by computing your model in kaggle kernel? this is really very impressive and seems very difficult with 6 channels,thanks for letting me know,wish you luck @aharless
", "votes": null, "replies": [] }, { "id": 631680, "author_name": "aharless", "author_url": "", "post_date": "09/22/2019 13:14:54", "content": "There are models that run on different folds, each run in a separate kernel, and then saving weights and continuing in a new kernel with the saved weights, sometimes re-randomizing a subset of the weights. And there are a number of post-processing steps that run in separate kernels. And some feedback from final predictions back into new versions of the models. So a lot of kernels that each take output from several other kernels as input. And the submissions are directly from kernels.
", "votes": null, "replies": [] }, { "id": 631729, "author_name": "aharless", "author_url": "", "post_date": "09/22/2019 15:02:47", "content": "It’s a long way from just one or two kernels (as in the public ones, which don’t carry the chain very far) to having a whole web of kernels that use one another’s output and apply multiple tricks (whereas with public kernels that demonstrated particular tricks, most didn’t combine multiple tricks).
", "votes": null, "replies": [] }, { "id": 631784, "author_name": "mobassir", "author_url": "", "post_date": "09/22/2019 16:51:03", "content": "wow!!! really very interesting,i am desperately waiting to see your solution as soon as this competition ends
", "votes": null, "replies": [] }, { "id": 632066, "author_name": "hmendonca", "author_url": "", "post_date": "09/23/2019 07:21:24", "content": "Yeah, Kaggle used to be a great platform to experiment ideas, before the weekly quota! 😿 \nLuckily we still had time to try a few... we'll make some of them public here once the comp is over\nCheers!
", "votes": null, "replies": [] }, { "id": 632099, "author_name": "mobassir", "author_url": "", "post_date": "09/23/2019 08:19:59", "content": "thank you a lot @hmendonca
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "631279": "hi,i see few good kernels written using kaggle kernels are not getting good score so i would like to know who got impressive lb score using kaggle kernels only?please comment your lb score using kaggle kernels only,thank you in advance\n#happyKaggling", "631446": "As far as I know, our 0.929 is essentially all from Kaggle kernels. There are a _lot_ of kernels involved, taking advantage of the relatively liberal GPU availability prior to the introduction of quotas. And our attempts to use other platforms have mostly failed. (The rtx2060 on my laptop keeps crashing with \"cuDNN failed to get convolution algorithm.\" My desktop is running unbearably slow because only 6GB of GPU memory means I have to run batchsize=8, and it hasn't produced anything useful yet. I haven't managed to get the TPU code working on Google Colab. And we haven't managed to set up GCP to run our models yet. I guess if the quotas had been around since the beginning of the competition, we would have put more effort into these alternatives.)", "631528": "hey,did you get 0.929 by computing your model in kaggle kernel? this is really very impressive and seems very difficult with 6 channels,thanks for letting me know,wish you luck @aharless", "631680": "There are models that run on different folds, each run in a separate kernel, and then saving weights and continuing in a new kernel with the saved weights, sometimes re-randomizing a subset of the weights. And there are a number of post-processing steps that run in separate kernels. And some feedback from final predictions back into new versions of the models. So a lot of kernels that each take output from several other kernels as input. And the submissions are directly from kernels.", "631729": "It’s a long way from just one or two kernels (as in the public ones, which don’t carry the chain very far) to having a whole web of kernels that use one another’s output and apply multiple tricks (whereas with public kernels that demonstrated particular tricks, most didn’t combine multiple tricks).", "631784": "wow!!! really very interesting,i am desperately waiting to see your solution as soon as this competition ends", "632066": "Yeah, Kaggle used to be a great platform to experiment ideas, before the weekly quota! 😿 \nLuckily we still had time to try a few... we'll make some of them public here once the comp is over\nCheers!", "632099": "thank you a lot @hmendonca" }, "source": "meta" }