Hi, first time trying out a competition, so I've got a nooby Kaggle question. I'm trying to use torch-geometric in a submission notebook, but I need to have internet disabled so I cant install via !pip. I've tried creating two separate notebooks for training and inference (is this a common Kaggle competition paradigm?) and using torch.jit.save to save and import the .pt model file, but I get this error when trying to use torch script with a GATConv layer… I've tried saving the model as a .pth/pickling, but that requires redefining the model class in the inference notebook (which therefore needs torch-geometric)?
Thinking about writing/copying the GATconv layer, but wondering if there's an easier work around for utilizing torch geometric models in an offline nb?
", "messages": [ { "id": "3139819", "postDate": "03/03/2025 23:43:31", "content": "Hi, first time trying out a competition, so I've got a nooby Kaggle question. I'm trying to use torch-geometric in a submission notebook, but I need to have internet disabled so I cant install via !pip. I've tried creating two separate notebooks for training and inference (is this a common Kaggle competition paradigm?) and using torch.jit.save to save and import the .pt model file, but I get this error when trying to use torch script with a GATConv layer… I've tried saving the model as a .pth/pickling, but that requires redefining the model class in the inference notebook (which therefore needs torch-geometric)?
Thinking about writing/copying the GATconv layer, but wondering if there's an easier work around for utilizing torch geometric models in an offline nb?
", "rawMarkdown": "Hi, first time trying out a competition, so I've got a nooby Kaggle question. I'm trying to use torch-geometric in a submission notebook, but I need to have internet disabled so I cant install via `!pip`. I've tried creating two separate notebooks for training and inference (is this a common Kaggle competition paradigm?) and using `torch.jit.save` to save and import the .pt model file, but I get [this error](https://github.com/pyg-team/pytorch_geometric/issues/8867) when trying to use torch script with a GATConv layer... I've tried saving the model as a .pth/pickling, but that requires redefining the model class in the inference notebook (which therefore needs torch-geometric)?\n\nThinking about writing/copying the GATconv layer, but wondering if there's an easier work around for utilizing torch geometric models in an offline nb?", "votes": null }, { "id": "3139855", "postDate": "03/04/2025 00:35:24", "content": "It is normal to separate training and inference. In fact, you can do training on your local computer, create a dataset that will hold a model, and do the inference from that dataset. It seems you already tried that, which is the right track.
\nAs to the internet, again create a Kaggle dataset, upload the wheels for packages of interest, and use pip to install from those local wheels. Something like this:
\nYou can install with pip under the dependency manager (see Install Dependencies in the menu), it's run before the notebook and it works with the internet disabled
", "rawMarkdown": "You can install with pip under the dependency manager (see Install Dependencies in the menu), it's run before the notebook and it works with the internet disabled", "votes": null }, { "id": "3140303", "postDate": "03/04/2025 13:33:02", "content": "Thanks! That makes sense, and ty for the links 👀
", "rawMarkdown": "Thanks! That makes sense, and ty for the links 👀", "votes": null }, { "id": "3140304", "postDate": "03/04/2025 13:33:16", "content": "Thanks! Will try this out
", "rawMarkdown": "Thanks! Will try this out", "votes": null } ], "comments": [ { "id": 3139855, "author_name": "tilii7", "author_url": "", "post_date": "03/04/2025 00:35:24", "content": "It is normal to separate training and inference. In fact, you can do training on your local computer, create a dataset that will hold a model, and do the inference from that dataset. It seems you already tried that, which is the right track.
\nAs to the internet, again create a Kaggle dataset, upload the wheels for packages of interest, and use pip to install from those local wheels. Something like this:
\nThanks! That makes sense, and ty for the links 👀
", "votes": null, "replies": [] } ] }, { "id": 3140131, "author_name": "eugenebaulin", "author_url": "", "post_date": "03/04/2025 10:04:07", "content": "You can install with pip under the dependency manager (see Install Dependencies in the menu), it's run before the notebook and it works with the internet disabled
", "votes": null, "replies": [ { "id": 3140304, "author_name": "ubitquitin", "author_url": "", "post_date": "03/04/2025 13:33:16", "content": "Thanks! Will try this out
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3139819": "Hi, first time trying out a competition, so I've got a nooby Kaggle question. I'm trying to use torch-geometric in a submission notebook, but I need to have internet disabled so I cant install via `!pip`. I've tried creating two separate notebooks for training and inference (is this a common Kaggle competition paradigm?) and using `torch.jit.save` to save and import the .pt model file, but I get [this error](https://github.com/pyg-team/pytorch_geometric/issues/8867) when trying to use torch script with a GATConv layer... I've tried saving the model as a .pth/pickling, but that requires redefining the model class in the inference notebook (which therefore needs torch-geometric)?\n\nThinking about writing/copying the GATconv layer, but wondering if there's an easier work around for utilizing torch geometric models in an offline nb?", "3139855": "It is normal to separate training and inference. In fact, you can do training on your local computer, create a dataset that will hold a model, and do the inference from that dataset. It seems you already tried that, which is the right track.\n\nAs to the internet, again create a Kaggle dataset, upload the wheels for packages of interest, and use pip to install from those local wheels. Something like this:\n\n- https://www.kaggle.com/competitions/llm-detect-ai-generated-text/discussion/455681\n- https://www.kaggle.com/competitions/hpa-single-cell-image-classification/discussion/215172", "3140131": "You can install with pip under the dependency manager (see Install Dependencies in the menu), it's run before the notebook and it works with the internet disabled", "3140303": "Thanks! That makes sense, and ty for the links 👀", "3140304": "Thanks! Will try this out" }, "source": "meta" }