![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1267014%2Fa6ef9ed24832bf03851a8103670489cc%2Fsubmit_score.PNG?generation=1577012116482994&alt=media\")
This discussion is based on this discussion(2019 Data Science Bowl - Discussion - Fast submission, train and predict locally).
\n\nThis Kernel's submitBengali.AI Handwritten Grapheme Classification - Kernel - Bengali_PublicScore_check is based on [Grapheme] ResNet-15 Naive Learning Inference.
\n\nSo Publicscore should be 0.8566, but 0.0615.\nWhy?\n
This discussion is based on this discussion(2019 Data Science Bowl - Discussion - Fast submission, train and predict locally).
\n\nThis Kernel's submitBengali.AI Handwritten Grapheme Classification - Kernel - Bengali_PublicScore_check is based on [Grapheme] ResNet-15 Naive Learning Inference.
\n\nSo Publicscore should be 0.8566, but 0.0615.\nWhy?\n
Because the size of the locally available test set is 12. The public leaderboard is calculated based on ~100000 samples, so we have to predict during the evaluation process (private rerun). That solution won't work in this competition.
", "rawMarkdown": "Because the size of the locally available test set is 12. The public leaderboard is calculated based on ~100000 samples, so we have to predict during the evaluation process (private rerun). That solution won't work in this competition.", "votes": null }, { "id": "700665", "postDate": "12/22/2019 11:35:53", "content": "Hello Wakame, as stated in the Data section, the competitors have access to a few (12) test images through the Notebooks. Notice that the resnet submission file has 12*3 entries. This is not the full public test set.\nDuring testing, the partial test .parquet files are swapped by the full versions and the model is evaluated on that. The actual number of test images are roughly the same as the train images. \nThis is different from the DSB 2019, where I believe all of the public test set is downloadable.
", "rawMarkdown": "Hello Wakame, as stated in the Data section, the competitors have access to a few (12) test images through the Notebooks. Notice that the resnet submission file has 12*3 entries. This is not the full public test set.\nDuring testing, the partial test .parquet files are swapped by the full versions and the model is evaluated on that. The actual number of test images are roughly the same as the train images. \nThis is different from the DSB 2019, where I believe all of the public test set is downloadable.", "votes": null }, { "id": "701975", "postDate": "12/24/2019 05:54:21", "content": "Thank you for answering！
", "rawMarkdown": "Thank you for answering！", "votes": null }, { "id": "701976", "postDate": "12/24/2019 05:54:39", "content": "Thanks！
", "rawMarkdown": "Thanks！", "votes": null } ], "comments": [ { "id": 700662, "author_name": "pestipeti", "author_url": "", "post_date": "12/22/2019 11:30:57", "content": "Because the size of the locally available test set is 12. The public leaderboard is calculated based on ~100000 samples, so we have to predict during the evaluation process (private rerun). That solution won't work in this competition.
", "votes": null, "replies": [ { "id": 701976, "author_name": "wakamezake", "author_url": "", "post_date": "12/24/2019 05:54:39", "content": "Thanks！
", "votes": null, "replies": [] } ] }, { "id": 700665, "author_name": "reasat", "author_url": "", "post_date": "12/22/2019 11:35:53", "content": "Hello Wakame, as stated in the Data section, the competitors have access to a few (12) test images through the Notebooks. Notice that the resnet submission file has 12*3 entries. This is not the full public test set.\nDuring testing, the partial test .parquet files are swapped by the full versions and the model is evaluated on that. The actual number of test images are roughly the same as the train images. \nThis is different from the DSB 2019, where I believe all of the public test set is downloadable.
", "votes": null, "replies": [ { "id": 701975, "author_name": "wakamezake", "author_url": "", "post_date": "12/24/2019 05:54:21", "content": "Thank you for answering！
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "700637": "This discussion is based on this discussion([2019 Data Science Bowl - Discussion - Fast submission, train and predict locally](https://www.kaggle.com/c/data-science-bowl-2019/discussion/120840)).\n\nThis Kernel's submit[Bengali.AI Handwritten Grapheme Classification - Kernel - Bengali_PublicScore_check](https://www.kaggle.com/wakamezake/bengali-publicscore-check) is based on [[Grapheme] ResNet-15 Naive Learning Inference](https://www.kaggle.com/hanjoonchoe/grapheme-resnet-18-n-l-inference-lb-0-8566).\n\nSo Publicscore should be 0.8566, but 0.0615.\nWhy?\n", "700662": "Because the size of the locally available test set is 12. The public leaderboard is calculated based on ~100000 samples, so we have to predict during the evaluation process (private rerun). That solution won't work in this competition.", "700665": "Hello Wakame, as stated in the Data section, the competitors have access to a few (12) test images through the Notebooks. Notice that the resnet submission file has 12*3 entries. This is not the full public test set.\nDuring testing, the partial test .parquet files are swapped by the full versions and the model is evaluated on that. The actual number of test images are roughly the same as the train images. \nThis is different from the DSB 2019, where I believe all of the public test set is downloadable.", "701975": "Thank you for answering！", "701976": "Thanks！" }, "source": "meta" }