Congratulations to each of the kagglers, it was a very interesting game, and thanks to each of the selfless kagglers on the discussion.\nIn the meantime, I really appreciate the hard work of each of my teammates, and without their experiments and their GPUs, I don't think we can get the gold medal.
\n\n[Update]
\n\n\n\n> We used the external HPA data in Gray format(512 size)
\n\nModels
\n\nData Augumentation
\n\ntrain\n> Add/Multiply/Crop/Affine/Filplr/Filpub/
12 TTA
Optimizer
\n\nLoss Function
\n\nThreshold\n> We tried the search threshold，but it was not work, so we finally had no idea and chose 0.205 as threshold.
\n\nResult
\n\n> we changed the format external HPA data (you can find this in discussion)
\n\nEnsemble\nwe used the first and second stage model to ensemble, and it should get 0.62+ score(public)
\n\nCongratulations to each of the kagglers, it was a very interesting game, and thanks to each of the selfless kagglers on the discussion.\nIn the meantime, I really appreciate the hard work of each of my teammates, and without their experiments and their GPUs, I don't think we can get the gold medal.
\n\n[Update]
\n\n\n\n> We used the external HPA data in Gray format(512 size)
\n\nModels
\n\nData Augumentation
\n\ntrain\n> Add/Multiply/Crop/Affine/Filplr/Filpub/
12 TTA
Optimizer
\n\nLoss Function
\n\nThreshold\n> We tried the search threshold，but it was not work, so we finally had no idea and chose 0.205 as threshold.
\n\nResult
\n\n> we changed the format external HPA data (you can find this in discussion)
\n\nEnsemble\nwe used the first and second stage model to ensemble, and it should get 0.62+ score(public)
\n\nThanks for sharing. I did the same experiments as your first and second stages. Look forward to sharing your third \nstage experiments.
", "rawMarkdown": "Thanks for sharing. I did the same experiments as your first and second stages. Look forward to sharing your third \nstage experiments.", "votes": null }, { "id": "454100", "postDate": "01/11/2019 06:08:39", "content": "Yep, Thank you. The third stage was funny.
", "rawMarkdown": "Yep, Thank you. The third stage was funny.", "votes": null }, { "id": "454111", "postDate": "01/11/2019 06:23:43", "content": "teammate's solution
\n\nhttps://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/77289
", "rawMarkdown": "teammate's solution\n\nhttps://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/77289", "votes": null }, { "id": "454149", "postDate": "01/11/2019 07:09:34", "content": "Congratulations and thanks for sharing!
", "rawMarkdown": "Congratulations and thanks for sharing!", "votes": null }, { "id": "454206", "postDate": "01/11/2019 08:41:15", "content": "Congratulations too and you are welcome.
", "rawMarkdown": "Congratulations too and you are welcome.", "votes": null }, { "id": "454840", "postDate": "01/12/2019 10:57:30", "content": "I've uploaded our team's code to github，thanks for all kagglers.
", "rawMarkdown": "I've uploaded our team's code to github，thanks for all kagglers.", "votes": null }, { "id": "456525", "postDate": "01/16/2019 01:23:49", "content": "学弟，你也是东莞理工的？
", "rawMarkdown": "学弟，你也是东莞理工的？", "votes": null }, { "id": "456531", "postDate": "01/16/2019 01:58:48", "content": "Yep
", "rawMarkdown": "Yep", "votes": null }, { "id": "457365", "postDate": "01/17/2019 09:35:50", "content": "hi,thanks for posting in solution could you please tell me where i can find in download script for external data hpa
", "rawMarkdown": "hi,thanks for posting in solution could you please tell me where i can find in download script for external data hpa", "votes": null } ], "comments": [ { "id": 454095, "author_name": "lbyg1994", "author_url": "", "post_date": "01/11/2019 05:56:46", "content": "Thanks for sharing. I did the same experiments as your first and second stages. Look forward to sharing your third \nstage experiments.
", "votes": null, "replies": [ { "id": 454100, "author_name": "garybios", "author_url": "", "post_date": "01/11/2019 06:08:39", "content": "Yep, Thank you. The third stage was funny.
", "votes": null, "replies": [] } ] }, { "id": 454111, "author_name": "shisususu", "author_url": "", "post_date": "01/11/2019 06:23:43", "content": "teammate's solution
\n\nhttps://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/77289
", "votes": null, "replies": [] }, { "id": 454149, "author_name": "sgalib", "author_url": "", "post_date": "01/11/2019 07:09:34", "content": "Congratulations and thanks for sharing!
", "votes": null, "replies": [ { "id": 454206, "author_name": "garybios", "author_url": "", "post_date": "01/11/2019 08:41:15", "content": "Congratulations too and you are welcome.
", "votes": null, "replies": [] } ] }, { "id": 454840, "author_name": "garybios", "author_url": "", "post_date": "01/12/2019 10:57:30", "content": "I've uploaded our team's code to github，thanks for all kagglers.
", "votes": null, "replies": [] }, { "id": 456525, "author_name": "willor", "author_url": "", "post_date": "01/16/2019 01:23:49", "content": "学弟，你也是东莞理工的？
", "votes": null, "replies": [ { "id": 456531, "author_name": "garybios", "author_url": "", "post_date": "01/16/2019 01:58:48", "content": "Yep
", "votes": null, "replies": [] } ] }, { "id": 457365, "author_name": "jaideepvalani", "author_url": "", "post_date": "01/17/2019 09:35:50", "content": "hi,thanks for posting in solution could you please tell me where i can find in download script for external data hpa
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "454067": "Congratulations to each of the kagglers, it was a very interesting game, and thanks to each of the selfless kagglers on the discussion.\nIn the meantime, I really appreciate the hard work of each of my teammates, and without their experiments and their GPUs, I don't think we can get the gold medal.\n\n##The following is our experiment:\n**[Update]**\n\n[The code of our solution ][2]\n\n### The first stage experiments \n> We used the external HPA data in Gray format(512 size)\n\n**Models**\n\n - res18 (batchsize=64)\n - res34 (batchsize=32)\n - bninception (batchsize=32)\n - inceptionv3 (batchsize=32)\n - xception (batchsize=24, P40-24G) \n - Se-resnext50(batchsize=24, P40-24G)\n\n**Data Augumentation**\n\n - train\n> Add/Multiply/Crop/Affine/Filplr/Filpub/\n\n - 12 TTA\n\n**Optimizer**\n\n - NAdam with different LR for different layers \n\n**Loss Function**\n\n- bce\n\n**Threshold**\n> We tried the search threshold，but it was not work, so we finally had no idea and chose 0.205 as threshold.\n\n**Result**\n\n- The best score from single model with 5 fold was 0.597(public)\n\n### The second stage experiments \n> we changed the format external HPA data (you can find this in discussion)\n\n\n**Ensemble**\nwe used the first and second stage model to ensemble, and it should get 0.62+ score(public)\n\n### The third stage experiments\n> [My teammate, shisu's method][1]\n\n\n [1]: https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/77289\n [2]: https://github.com/Gary-Deeplearning/Human_Protein", "454095": "Thanks for sharing. I did the same experiments as your first and second stages. Look forward to sharing your third \nstage experiments.", "454100": "Yep, Thank you. The third stage was funny.", "454111": "teammate's solution\n\nhttps://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/77289", "454149": "Congratulations and thanks for sharing!", "454206": "Congratulations too and you are welcome.", "454840": "I've uploaded our team's code to github，thanks for all kagglers.", "456525": "学弟，你也是东莞理工的？", "456531": "Yep", "457365": "hi,thanks for posting in solution could you please tell me where i can find in download script for external data hpa" }, "source": "meta" }