{ "id": 384350, "title": "Are the constraints for training or inference?", "url": "/competitions/predict-student-performance-from-game-play/discussion/384350", "author_name": "Mohamed Eltayeb", "post_date": "2023-02-07T15:53:10.730000", "votes": 9, "comment_count": 6, "views": 0, "content": "

Hello
\nJust want to know whether we can train our model locally, save them, upload them to Kaggle then just use them for inference? Or we should train our model with the given constraints to count this as a valid solution at the end?

", "messages": [ { "id": 2133794, "postDate": "2023-02-07T15:53:10.730Z", "content": "

Hello
\nJust want to know whether we can train our model locally, save them, upload them to Kaggle then just use them for inference? Or we should train our model with the given constraints to count this as a valid solution at the end?

", "rawMarkdown": "Hello\nJust want to know whether we can train our model locally, save them, upload them to Kaggle then just use them for inference? Or we should train our model with the given constraints to count this as a valid solution at the end?", "votes": 9 }, { "id": 2140426, "postDate": "2023-02-11T18:28:26.740Z", "content": "

No there are no constraints. Furthermore, i suggest that you train/validate models using GPU with cuDF-GPU (for feature engineer) and GBT-GPU (for model train). Then you can run faster experiments. Then save your best model with pickle to Kaggle dataset. Then load into a CPU inference submit notebook.

", "rawMarkdown": "No there are no constraints. Furthermore, i suggest that you train/validate models using GPU with cuDF-GPU (for feature engineer) and GBT-GPU (for model train). Then you can run faster experiments. Then save your best model with pickle to Kaggle dataset. Then load into a CPU inference submit notebook.", "votes": 3, "replies": [ { "id": 2140562, "postDate": "2023-02-11T20:44:57.627Z", "content": "

do u think a GPU inference speed up the process?

", "rawMarkdown": "do u think a GPU inference speed up the process?", "replies": [ { "id": 2140574, "postDate": "2023-02-11T21:10:43.723Z", "content": "

We cannot use GPU for infer. Only CPU notebook can be submitted to the competition. But we can use GPU notebook to train and save our models to a Kaggle dataset. Then load these models into a CPU notebook for inference.

\n

Using GPU for train/validate will help us run faster experiments to discover what makes our model have better CV score and resultantly better LB score.

", "rawMarkdown": "We cannot use GPU for infer. Only CPU notebook can be submitted to the competition. But we can use GPU notebook to train and save our models to a Kaggle dataset. Then load these models into a CPU notebook for inference.\n\nUsing GPU for train/validate will help us run faster experiments to discover what makes our model have better CV score and resultantly better LB score.", "votes": 3 }, { "id": 2140586, "postDate": "2023-02-11T21:29:38.127Z", "rawMarkdown": "", "isDeleted": true } ] } ] }, { "id": 2139940, "postDate": "2023-02-11T09:43:52.603Z", "content": "

Hello Mohamed,
\nDid find the answer for your question?
\nI am also concerned,
\nThanks!

", "rawMarkdown": "Hello Mohamed,\nDid find the answer for your question?\nI am also concerned,\nThanks!", "votes": 1, "replies": [ { "id": 2139959, "postDate": "2023-02-11T10:15:43.273Z", "content": "

No. Let me ask the host.
\nCould you answer my question please @alexmlfranklin ?

", "rawMarkdown": "No. Let me ask the host.\nCould you answer my question please @alexmlfranklin ?", "votes": 2 } ] } ], "comments": [ { "id": 2140426, "author_name": "Chris Deotte", "author_url": "", "post_date": "2023-02-11T18:28:26.740000", "content": "

No there are no constraints. Furthermore, i suggest that you train/validate models using GPU with cuDF-GPU (for feature engineer) and GBT-GPU (for model train). Then you can run faster experiments. Then save your best model with pickle to Kaggle dataset. Then load into a CPU inference submit notebook.

", "votes": 3, "replies": [ { "id": 2140562, "author_name": "Reacher", "author_url": "", "post_date": "2023-02-11T20:44:57.627000", "content": "

do u think a GPU inference speed up the process?

", "votes": 0, "replies": [ { "id": 2140574, "author_name": "Chris Deotte", "author_url": "", "post_date": "2023-02-11T21:10:43.723000", "content": "

We cannot use GPU for infer. Only CPU notebook can be submitted to the competition. But we can use GPU notebook to train and save our models to a Kaggle dataset. Then load these models into a CPU notebook for inference.

\n

Using GPU for train/validate will help us run faster experiments to discover what makes our model have better CV score and resultantly better LB score.

", "votes": 3, "replies": [] }, { "id": 2140586, "author_name": "", "author_url": "", "post_date": "2023-02-11T21:29:38.127000", "content": "", "votes": 0, "replies": [] } ] } ] }, { "id": 2139940, "author_name": "Tesnim Hadhri", "author_url": "", "post_date": "2023-02-11T09:43:52.603000", "content": "

Hello Mohamed,
\nDid find the answer for your question?
\nI am also concerned,
\nThanks!

", "votes": 1, "replies": [ { "id": 2139959, "author_name": "Mohamed Eltayeb", "author_url": "", "post_date": "2023-02-11T10:15:43.273000", "content": "

No. Let me ask the host.
\nCould you answer my question please @alexmlfranklin ?

", "votes": 2, "replies": [] } ] } ], "raw_markdown_by_id": { "2133794": "Hello\nJust want to know whether we can train our model locally, save them, upload them to Kaggle then just use them for inference? Or we should train our model with the given constraints to count this as a valid solution at the end?", "2140426": "No there are no constraints. Furthermore, i suggest that you train/validate models using GPU with cuDF-GPU (for feature engineer) and GBT-GPU (for model train). Then you can run faster experiments. Then save your best model with pickle to Kaggle dataset. Then load into a CPU inference submit notebook.", "2139940": "Hello Mohamed,\nDid find the answer for your question?\nI am also concerned,\nThanks!" } }