In this competition, we don't have a separate private test set. In the test.zarr we have all the samples, so we can predict locally. For me, it is much faster: instead of ~2 hours, I submitted my last result in less than 15 minutes.
\nimport pandas as pd\n\n# Change these to your dataset/submission.csv\nSUBMISSION_FOLDER = '/kaggle/input/lyft-submissions-private'\nSUBMISSION_FILE = 'best_valid__submission.csv'\n\n\nsubmissions = pd.read_csv(f\"{SUBMISSION_FOLDER}/{SUBMISSION_FILE}\")\nsubmissions.to_csv(\"submission.csv\", index=False)\nIn this competition, we don't have a separate private test set. In the test.zarr we have all the samples, so we can predict locally. For me, it is much faster: instead of ~2 hours, I submitted my last result in less than 15 minutes.
\nimport pandas as pd\n\n# Change these to your dataset/submission.csv\nSUBMISSION_FOLDER = '/kaggle/input/lyft-submissions-private'\nSUBMISSION_FILE = 'best_valid__submission.csv'\n\n\nsubmissions = pd.read_csv(f\"{SUBMISSION_FOLDER}/{SUBMISSION_FILE}\")\nsubmissions.to_csv(\"submission.csv\", index=False)\nThank you for the post @pestipeti , I demonstrated it in the kernel Save your time, submit without kernel inference using your kernel output, it worked very well. Thanks again!
", "rawMarkdown": "Thank you for the post @pestipeti , I demonstrated it in the kernel [Save your time, submit without kernel inference](https://www.kaggle.com/corochann/save-your-time-submit-without-kernel-inference) using your kernel output, it worked very well. Thanks again!", "votes": 5 }, { "id": 1046757, "postDate": "2020-10-12T01:29:31.073Z", "content": "Hmm… If you can create a private dataset which contains the predictions offline, then what is the point of being a Code Competition? Wouldn't it be easier for Kaggle just provide the regular submit csv as prediction?
", "rawMarkdown": "Hmm... If you can create a private dataset which contains the predictions offline, then what is the point of being a Code Competition? Wouldn't it be easier for Kaggle just provide the regular submit csv as prediction?", "votes": 1 }, { "id": 1074840, "postDate": "2020-11-11T06:00:45.007Z", "content": "Or simply
\n!cp ../input/lyft-submissions-private/best_valid__submission.csv submission.csv
\ninstead of reading in by pandas since read then write csv don't will introduce precision error.
it can be ever faster ~4 min https://www.kaggle.com/lao777/fast-submission-valid-for-public-private-lb
", "rawMarkdown": "it can be ever faster ~4 min https://www.kaggle.com/lao777/fast-submission-valid-for-public-private-lb", "votes": 2 }, { "id": 1090877, "postDate": "2020-11-25T16:30:02.793Z", "content": "I am new to Kaggle and I have already stored submission.csv as a private dataset and it's in the input folder. I get an error saying \"your notebook tried to allocate more memory than is available\". Is there any work around it?
", "rawMarkdown": "I am new to Kaggle and I have already stored submission.csv as a private dataset and it's in the input folder. I get an error saying \"your notebook tried to allocate more memory than is available\". Is there any work around it?" }, { "id": 1074971, "postDate": "2020-11-11T09:23:47.257Z", "content": "Hi @pestipeti, thanks for pointing out this technique. I'm wondering if we need to generate a local submission file from all test.zarr or the chopped_dataset of test.zarr? I think we should do the first option because Kaggle kernel will select the \"100th frames\" and show a metric score? Am I right?
Thank you!
", "rawMarkdown": "Hi @pestipeti, thanks for pointing out this technique. I'm wondering if we need to generate a local submission file from all test.zarr or the `chopped_dataset of test.zarr`? I think we should do the first option because Kaggle kernel will select the \"100th frames\" and show a metric score? Am I right?\n\nThank you!", "replies": [ { "id": 1075026, "postDate": "2020-11-11T10:30:38.597Z", "content": "You should predict the samples from test.zarr, it is already chopped.
", "rawMarkdown": "You should predict the samples from test.zarr, it is already chopped." }, { "id": 1075101, "postDate": "2020-11-11T11:40:41.287Z", "content": "Hi Peter, thank you for your confirmation. Now I get it! Thank you!
", "rawMarkdown": "Hi Peter, thank you for your confirmation. Now I get it! Thank you!" } ] }, { "id": 990750, "postDate": "2020-08-29T20:19:29.740Z", "content": "Good share
", "rawMarkdown": "Good share" }, { "id": 990158, "postDate": "2020-08-29T11:33:55.297Z", "content": "Thanks you it's very useful @pestipeti ,It's saves time …
", "rawMarkdown": "Thanks you it's very useful @pestipeti ,It's saves time ..." }, { "id": 988091, "postDate": "2020-08-27T19:11:05.890Z", "content": "Very useful indeed. I have noticed that that was the case for few other competitions as well, i.e. you can predict offline by building a private dataset. Thanks for sharing!
", "rawMarkdown": "Very useful indeed. I have noticed that that was the case for few other competitions as well, i.e. you can predict offline by building a private dataset. Thanks for sharing!" }, { "id": 987391, "postDate": "2020-08-27T08:12:12.360Z", "content": "This is a great idea! The test dataset is SO big it takes a ridiculous amount of time to create a submission from a kernel that can only run on CPU (since there's a bug in the l5kit script you have to include that means you can't run in GPU).
", "rawMarkdown": "This is a great idea! The test dataset is SO big it takes a ridiculous amount of time to create a submission from a kernel that can only run on CPU (since there's a bug in the l5kit script you have to include that means you can't run in GPU). ", "replies": [ { "id": 987399, "postDate": "2020-08-27T08:20:40.813Z", "content": "Even if you are able to use the GPU the process is slow, because of the rasterization. The speedup on my local machine is because I have 12 CPU cores (instead of the 4 on the kernels). The rasterization is faster, but still, it is the bottleneck.
", "rawMarkdown": "Even if you are able to use the GPU the process is slow, because of the rasterization. The speedup on my local machine is because I have 12 CPU cores (instead of the 4 on the kernels). The rasterization is faster, but still, it is the bottleneck. ", "votes": 3 } ] }, { "id": 1052824, "postDate": "2020-10-18T10:18:38.723Z", "rawMarkdown": "", "isDeleted": true }, { "id": 1047071, "postDate": "2020-10-12T08:29:54.567Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 988276, "author_name": "corochann", "author_url": "", "post_date": "2020-08-28T00:04:51.027000", "content": "Thank you for the post @pestipeti , I demonstrated it in the kernel Save your time, submit without kernel inference using your kernel output, it worked very well. Thanks again!
", "votes": 5, "replies": [] }, { "id": 1046757, "author_name": "Louis Yang", "author_url": "", "post_date": "2020-10-12T01:29:31.073000", "content": "Hmm… If you can create a private dataset which contains the predictions offline, then what is the point of being a Code Competition? Wouldn't it be easier for Kaggle just provide the regular submit csv as prediction?
", "votes": 1, "replies": [] }, { "id": 1074840, "author_name": "Louis Yang", "author_url": "", "post_date": "2020-11-11T06:00:45.007000", "content": "Or simply
\n!cp ../input/lyft-submissions-private/best_valid__submission.csv submission.csv
\ninstead of reading in by pandas since read then write csv don't will introduce precision error.
it can be ever faster ~4 min https://www.kaggle.com/lao777/fast-submission-valid-for-public-private-lb
", "votes": 2, "replies": [] }, { "id": 1090877, "author_name": "Praveen Phatate", "author_url": "", "post_date": "2020-11-25T16:30:02.793000", "content": "I am new to Kaggle and I have already stored submission.csv as a private dataset and it's in the input folder. I get an error saying \"your notebook tried to allocate more memory than is available\". Is there any work around it?
", "votes": 0, "replies": [] }, { "id": 1074971, "author_name": "豆柴金鯱", "author_url": "", "post_date": "2020-11-11T09:23:47.257000", "content": "Hi @pestipeti, thanks for pointing out this technique. I'm wondering if we need to generate a local submission file from all test.zarr or the chopped_dataset of test.zarr? I think we should do the first option because Kaggle kernel will select the \"100th frames\" and show a metric score? Am I right?
Thank you!
", "votes": 0, "replies": [ { "id": 1075026, "author_name": "Peter", "author_url": "", "post_date": "2020-11-11T10:30:38.597000", "content": "You should predict the samples from test.zarr, it is already chopped.
", "votes": 0, "replies": [] }, { "id": 1075101, "author_name": "豆柴金鯱", "author_url": "", "post_date": "2020-11-11T11:40:41.287000", "content": "Hi Peter, thank you for your confirmation. Now I get it! Thank you!
", "votes": 0, "replies": [] } ] }, { "id": 990750, "author_name": "Alin Cijov", "author_url": "", "post_date": "2020-08-29T20:19:29.740000", "content": "Good share
", "votes": 0, "replies": [] }, { "id": 990158, "author_name": "Surekha Ramireddy", "author_url": "", "post_date": "2020-08-29T11:33:55.297000", "content": "Thanks you it's very useful @pestipeti ,It's saves time …
", "votes": 0, "replies": [] }, { "id": 988091, "author_name": "Yassine Alouini", "author_url": "", "post_date": "2020-08-27T19:11:05.890000", "content": "Very useful indeed. I have noticed that that was the case for few other competitions as well, i.e. you can predict offline by building a private dataset. Thanks for sharing!
", "votes": 0, "replies": [] }, { "id": 987391, "author_name": "Ian Ormesher", "author_url": "", "post_date": "2020-08-27T08:12:12.360000", "content": "This is a great idea! The test dataset is SO big it takes a ridiculous amount of time to create a submission from a kernel that can only run on CPU (since there's a bug in the l5kit script you have to include that means you can't run in GPU).
", "votes": 0, "replies": [ { "id": 987399, "author_name": "Peter", "author_url": "", "post_date": "2020-08-27T08:20:40.813000", "content": "Even if you are able to use the GPU the process is slow, because of the rasterization. The speedup on my local machine is because I have 12 CPU cores (instead of the 4 on the kernels). The rasterization is faster, but still, it is the bottleneck.
", "votes": 3, "replies": [] } ] }, { "id": 1052824, "author_name": "", "author_url": "", "post_date": "2020-10-18T10:18:38.723000", "content": "", "votes": 0, "replies": [] }, { "id": 1047071, "author_name": "", "author_url": "", "post_date": "2020-10-12T08:29:54.567000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "986830": "In this competition, we don't have a separate private test set. In the test.zarr we have all the samples, so we can predict locally. For me, it is much faster: instead of ~2 hours, I submitted my last result in less than 15 minutes.\n\n- Predict locally (using the test.zarr and the mask.npz)\n- Store the result csv (I used the official method from l5kit)\n- Create a private dataset (you only have to do it once)\n- Upload the locally saved csv\n- Create a submission notebook (see the script below)\n- Commit, submit.\n\n```\nimport pandas as pd\n\n# Change these to your dataset/submission.csv\nSUBMISSION_FOLDER = '/kaggle/input/lyft-submissions-private'\nSUBMISSION_FILE = 'best_valid__submission.csv'\n\n\nsubmissions = pd.read_csv(f\"{SUBMISSION_FOLDER}/{SUBMISSION_FILE}\")\nsubmissions.to_csv(\"submission.csv\", index=False)\n```\n", "988276": "Thank you for the post @pestipeti , I demonstrated it in the kernel [Save your time, submit without kernel inference](https://www.kaggle.com/corochann/save-your-time-submit-without-kernel-inference) using your kernel output, it worked very well. Thanks again!", "1046757": "Hmm... If you can create a private dataset which contains the predictions offline, then what is the point of being a Code Competition? Wouldn't it be easier for Kaggle just provide the regular submit csv as prediction?", "1074840": "Or simply \n`!cp ../input/lyft-submissions-private/best_valid__submission.csv submission.csv`\ninstead of reading in by pandas since read then write csv don't will introduce precision error.", "1025396": "it can be ever faster ~4 min https://www.kaggle.com/lao777/fast-submission-valid-for-public-private-lb", "1090877": "I am new to Kaggle and I have already stored submission.csv as a private dataset and it's in the input folder. I get an error saying \"your notebook tried to allocate more memory than is available\". Is there any work around it?", "1074971": "Hi @pestipeti, thanks for pointing out this technique. I'm wondering if we need to generate a local submission file from all test.zarr or the `chopped_dataset of test.zarr`? I think we should do the first option because Kaggle kernel will select the \"100th frames\" and show a metric score? Am I right?\n\nThank you!", "990750": "Good share", "990158": "Thanks you it's very useful @pestipeti ,It's saves time ...", "988091": "Very useful indeed. I have noticed that that was the case for few other competitions as well, i.e. you can predict offline by building a private dataset. Thanks for sharing!", "987391": "This is a great idea! The test dataset is SO big it takes a ridiculous amount of time to create a submission from a kernel that can only run on CPU (since there's a bug in the l5kit script you have to include that means you can't run in GPU). ", "1052824": "", "1047071": "" } }