I have access to external GPUs (google colab), and I’m wondering whether it’s allowed to use them for experimental training sessions for this competition. The 30h/week Kaggle GPU limit isn’t sufficient for my workflow, especially with a large dataset.
\nTo be clear, I understand that for fairness, the final model must be trained and submitted using Kaggle Notebooks only. I will retrain the model on Kaggle once the GPU quota resets. The externally trained models will not be submitted—just purely for experimentation and performance measurement.
\nIf external training is allowed for experimentation, are there any specific considerations I should keep in mind?
\nYes
\nyou can use external GPUs (like Google Colab) for experimentation and training during development, as long as you respect the key rule: the final model must be trained and submitted on Kaggle Notebooks. Using Colab or other external environments is fine for prototyping, debugging, and performance testing, but the official submission must be reproducible inside Kaggle.
Things to Keep in Mind
\nBottom Line
\nExternal GPUs are great for speeding up iteration and exploration. Just remember: all final models and submissions must be trained from scratch on Kaggle so that results are fair and reproducible.
I would like to ask if we can use a model pre-trained on ImageNet, rather than training one from scratch ?
", "rawMarkdown": "I would like to ask if we can use a model pre-trained on ImageNet, rather than training one from scratch ?", "replies": [ { "id": 3281915, "postDate": "2025-09-05T08:17:25.833Z", "content": "Yes, pre-trained models (such as those trained on ImageNet) are fully allowed. Transfer learning is standard practice in medical imaging competitions, and as long as the weights are publicly available and accessible to all participants, you can use them. The only requirement is that the final training and submission must happen within Kaggle Notebooks.
", "rawMarkdown": "Yes, pre-trained models (such as those trained on ImageNet) are fully allowed. Transfer learning is standard practice in medical imaging competitions, and as long as the weights are publicly available and accessible to all participants, you can use them. The only requirement is that the final training and submission must happen within Kaggle Notebooks." }, { "id": 3283946, "postDate": "2025-09-08T03:26:08.980Z", "content": "@guntasdhanjal Could I ask whether chest X-ray foundation models, such as EVA-X and CheXFound, can be used?
", "rawMarkdown": "@guntasdhanjal Could I ask whether chest X-ray foundation models, such as [EVA-X](https://github.com/hustvl/EVA-X) and [CheXFound](https://github.com/RPIDIAL/CheXFound/tree/main), can be used?" }, { "id": 3284401, "postDate": "2025-09-08T08:11:17.710Z", "content": "@sjtuwangshuo Yes, you can use chest X-ray foundation models like EVA-X or CheXFound, since they’re publicly available and accessible to everyone. Just make sure your final fine-tuning and submission are done inside Kaggle Notebooks to keep results fair and reproducible.
", "rawMarkdown": "@sjtuwangshuo Yes, you can use chest X-ray foundation models like EVA-X or CheXFound, since they’re publicly available and accessible to everyone. Just make sure your final fine-tuning and submission are done inside Kaggle Notebooks to keep results fair and reproducible." } ] } ] }, { "id": 3281293, "postDate": "2025-09-04T02:35:41.027Z", "content": "I have access to external GPUs (google colab), and I’m wondering whether it’s allowed to use them for experimental training sessions for this competition. The 30h/week Kaggle GPU limit isn’t sufficient for my workflow, especially with a large dataset.
\nTo be clear, I understand that for fairness, the final model must be trained and submitted using Kaggle Notebooks only. I will retrain the model on Kaggle once the GPU quota resets. The externally trained models will not be submitted—just purely for experimentation and performance measurement.
\nIf external training is allowed for experimentation, are there any specific considerations I should keep in mind?
\nSorry, is there anywhere stating that participants are only allowed to use the Kaggle notebook GPU?
", "rawMarkdown": "Sorry, is there anywhere stating that participants are only allowed to use the Kaggle notebook GPU?", "replies": [ { "id": 3281411, "postDate": "2025-09-04T08:35:11.390Z", "content": "Well not in the rules since it said anything that is publicly available can be used.
\nBut I was reading this comment https://www.kaggle.com/competitions/grand-xray-slam-division-a/discussion/603071#3280154 and that just got me thinking.
\nI didnt know if he meant only Kaggle GPU or any GPU but the same limitation of 12h/sessions.
", "rawMarkdown": "Well not in the rules since it said anything that is publicly available can be used.\n\nBut I was reading this comment [https://www.kaggle.com/competitions/grand-xray-slam-division-a/discussion/603071#3280154](https://www.kaggle.com/competitions/grand-xray-slam-division-a/discussion/603071#3280154) and that just got me thinking.\n\nI didnt know if he meant only Kaggle GPU or any GPU but the same limitation of 12h/sessions.", "replies": [ { "id": 3281916, "postDate": "2025-09-05T08:21:08.807Z", "content": "That’s correct, external GPUs can be used for experimentation. The only rule is that final training and submission must be reproducible in Kaggle Notebooks.
", "rawMarkdown": "That’s correct, external GPUs can be used for experimentation. The only rule is that final training and submission must be reproducible in Kaggle Notebooks." } ] } ] } ], "comments": [ { "id": 3281508, "author_name": "Guntas Dhanjal", "author_url": "", "post_date": "2025-09-04T13:09:37.230000", "content": "Yes
\nyou can use external GPUs (like Google Colab) for experimentation and training during development, as long as you respect the key rule: the final model must be trained and submitted on Kaggle Notebooks. Using Colab or other external environments is fine for prototyping, debugging, and performance testing, but the official submission must be reproducible inside Kaggle.
Things to Keep in Mind
\nBottom Line
\nExternal GPUs are great for speeding up iteration and exploration. Just remember: all final models and submissions must be trained from scratch on Kaggle so that results are fair and reproducible.
I would like to ask if we can use a model pre-trained on ImageNet, rather than training one from scratch ?
", "votes": 0, "replies": [ { "id": 3281915, "author_name": "Guntas Dhanjal", "author_url": "", "post_date": "2025-09-05T08:17:25.833000", "content": "Yes, pre-trained models (such as those trained on ImageNet) are fully allowed. Transfer learning is standard practice in medical imaging competitions, and as long as the weights are publicly available and accessible to all participants, you can use them. The only requirement is that the final training and submission must happen within Kaggle Notebooks.
", "votes": 0, "replies": [] }, { "id": 3283946, "author_name": "AnnieGo", "author_url": "", "post_date": "2025-09-08T03:26:08.980000", "content": "@guntasdhanjal Could I ask whether chest X-ray foundation models, such as EVA-X and CheXFound, can be used?
", "votes": 0, "replies": [] }, { "id": 3284401, "author_name": "Guntas Dhanjal", "author_url": "", "post_date": "2025-09-08T08:11:17.710000", "content": "@sjtuwangshuo Yes, you can use chest X-ray foundation models like EVA-X or CheXFound, since they’re publicly available and accessible to everyone. Just make sure your final fine-tuning and submission are done inside Kaggle Notebooks to keep results fair and reproducible.
", "votes": 0, "replies": [] } ] } ] }, { "id": 3281382, "author_name": "AnnieGo", "author_url": "", "post_date": "2025-09-04T07:36:43.540000", "content": "Sorry, is there anywhere stating that participants are only allowed to use the Kaggle notebook GPU?
", "votes": 0, "replies": [ { "id": 3281411, "author_name": "Hiếu Lê Ngọc", "author_url": "", "post_date": "2025-09-04T08:35:11.390000", "content": "Well not in the rules since it said anything that is publicly available can be used.
\nBut I was reading this comment https://www.kaggle.com/competitions/grand-xray-slam-division-a/discussion/603071#3280154 and that just got me thinking.
\nI didnt know if he meant only Kaggle GPU or any GPU but the same limitation of 12h/sessions.
", "votes": 0, "replies": [ { "id": 3281916, "author_name": "Guntas Dhanjal", "author_url": "", "post_date": "2025-09-05T08:21:08.807000", "content": "That’s correct, external GPUs can be used for experimentation. The only rule is that final training and submission must be reproducible in Kaggle Notebooks.
", "votes": 0, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "3281508": "Yes \nyou can use external GPUs (like Google Colab) for experimentation and training during development, as long as you respect the key rule: the final model must be trained and submitted on Kaggle Notebooks. Using Colab or other external environments is fine for prototyping, debugging, and performance testing, but the official submission must be reproducible inside Kaggle.\n\n**Things to Keep in Mind**\n\n* Environment Matching: You don’t need to perfectly replicate Kaggle’s GPU/TPU type or training duration when experimenting externally. What matters is that your final model can be trained end-to-end in the Kaggle environment.\n* Insights from External Runs: It’s totally acceptable to use learnings from external experiments (e.g., hyperparameter tuning, architecture choices, preprocessing tricks). The restriction is only on where the final training run happens, not on how you arrive at your ideas.\n* Fairness & Reproducibility: Don’t submit externally trained models or weights. Instead, re-train on Kaggle once your quota resets. This ensures that all submissions are fairly benchmarked within the same infrastructure limits.\n\n**Bottom Line**\nExternal GPUs are great for speeding up iteration and exploration. Just remember: all final models and submissions must be trained from scratch on Kaggle so that results are fair and reproducible.", "3281293": "I have access to external GPUs (google colab), and I’m wondering whether it’s allowed to use them for experimental training sessions for this competition. The 30h/week Kaggle GPU limit isn’t sufficient for my workflow, especially with a large dataset.\n\nTo be clear, I understand that for fairness, the final model must be trained and submitted using Kaggle Notebooks only. I will retrain the model on Kaggle once the GPU quota resets. The externally trained models will not be submitted—just purely for experimentation and performance measurement.\n\nIf external training is allowed for experimentation, are there any specific considerations I should keep in mind? \n\n- Should I match the GPU/TPU type and training time to Kaggle’s environment?\n- Are there any restrictions on using insights gained from external runs?\n- Is it acceptable to use external training to guide hyperparameter tuning or model architecture choices?\n\n@guntasdhanjal ", "3281382": "Sorry, is there anywhere stating that participants are only allowed to use the Kaggle notebook GPU?" } }