Since the plates leak, I have been training four 277 class models and joining their predictions. In a couple recent posts I have seen high scoring people talking about training 1108 class models.
\n\nCan anyone share their experience in the two different methods?
", "messages": [ { "id": "633449", "postDate": "09/25/2019 00:24:23", "content": "Since the plates leak, I have been training four 277 class models and joining their predictions. In a couple recent posts I have seen high scoring people talking about training 1108 class models.
\n\nCan anyone share their experience in the two different methods?
", "rawMarkdown": "Since the plates leak, I have been training four 277 class models and joining their predictions. In a couple recent posts I have seen high scoring people talking about training 1108 class models.\n\nCan anyone share their experience in the two different methods?", "votes": null }, { "id": "633468", "postDate": "09/25/2019 01:36:52", "content": "In my experiment, 1108 class model is better than 277 class model. This may be because 1108 class model can use more information for training.
", "rawMarkdown": "In my experiment, 1108 class model is better than 277 class model. This may be because 1108 class model can use more information for training.", "votes": null }, { "id": "633695", "postDate": "09/25/2019 09:45:06", "content": "Thanks @shimacos for the input\nWe never got the chance to check it, but my plan was initially to train a strong model with all data, and then fine tune it's predictions on a per cell type and plate basis. That would make it 4*4 (16) models.\nBut we're out of time and GPU quota now hehe\nIn fact, our current best submission (LB:0.94) is from a very single model, no blending
", "rawMarkdown": "Thanks @shimacos for the input\nWe never got the chance to check it, but my plan was initially to train a strong model with all data, and then fine tune it's predictions on a per cell type and plate basis. That would make it 4*4 (16) models.\nBut we're out of time and GPU quota now hehe\nIn fact, our current best submission (LB:0.94) is from a very single model, no blending", "votes": null }, { "id": "633707", "postDate": "09/25/2019 09:56:12", "content": "Thanks for sharing :) Looks like I spent way too long trying and failing to get metric learning working.
\n\nThat is an awesome score for a single model (I think, I can't wait to see everyone's solutions 😄 ), is that a metric learning model?
", "rawMarkdown": "Thanks for sharing :) Looks like I spent way too long trying and failing to get metric learning working.\n\nThat is an awesome score for a single model (I think, I can't wait to see everyone's solutions 😄 ), is that a metric learning model?", "votes": null }, { "id": "633785", "postDate": "09/25/2019 12:18:14", "content": "@hmendonca We tried to do that and it didn't work well​ :(
", "rawMarkdown": "hmendonca We tried to do that and it didn't work well​ :(", "votes": null }, { "id": "633840", "postDate": "09/25/2019 13:32:37", "content": "I am doing a single model so far, I added the group as an additional input to be concatnated with the CNN features to be passed to the final output layer.
", "rawMarkdown": "I am doing a single model so far, I added the group as an additional input to be concatnated with the CNN features to be passed to the final output layer.", "votes": null }, { "id": "633893", "postDate": "09/25/2019 14:46:38", "content": "@igorkrashenyi Same here
", "rawMarkdown": "igorkrashenyi Same here", "votes": null }, { "id": "635310", "postDate": "09/27/2019 11:10:35", "content": "@cherring sorry for the delay\nYes, it uses ArcNet: https://www.kaggle.com/hmendonca/fold1h4r3-arcenetb4-2-256px-rcic-lb-0-9759\nThe implementation I found from the Humpback comp worked really well for me, but others showed that it wasn't necessary though hehe
\n\nedit: I've added some more details and how I got to that solution here: https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110346
", "rawMarkdown": "cherring sorry for the delay\nYes, it uses ArcNet: https://www.kaggle.com/hmendonca/fold1h4r3-arcenetb4-2-256px-rcic-lb-0-9759\nThe implementation I found from the Humpback comp worked really well for me, but others showed that it wasn't necessary though hehe\n\nedit: I've added some more details and how I got to that solution here: https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110346", "votes": null } ], "comments": [ { "id": 633468, "author_name": "shimacos", "author_url": "", "post_date": "09/25/2019 01:36:52", "content": "In my experiment, 1108 class model is better than 277 class model. This may be because 1108 class model can use more information for training.
", "votes": null, "replies": [] }, { "id": 633695, "author_name": "hmendonca", "author_url": "", "post_date": "09/25/2019 09:45:06", "content": "Thanks @shimacos for the input\nWe never got the chance to check it, but my plan was initially to train a strong model with all data, and then fine tune it's predictions on a per cell type and plate basis. That would make it 4*4 (16) models.\nBut we're out of time and GPU quota now hehe\nIn fact, our current best submission (LB:0.94) is from a very single model, no blending
", "votes": null, "replies": [ { "id": 633707, "author_name": "cherring", "author_url": "", "post_date": "09/25/2019 09:56:12", "content": "Thanks for sharing :) Looks like I spent way too long trying and failing to get metric learning working.
\n\nThat is an awesome score for a single model (I think, I can't wait to see everyone's solutions 😄 ), is that a metric learning model?
", "votes": null, "replies": [] }, { "id": 633785, "author_name": "igorkrashenyi", "author_url": "", "post_date": "09/25/2019 12:18:14", "content": "@hmendonca We tried to do that and it didn't work well​ :(
", "votes": null, "replies": [] }, { "id": 633893, "author_name": "narsil", "author_url": "", "post_date": "09/25/2019 14:46:38", "content": "@igorkrashenyi Same here
", "votes": null, "replies": [] }, { "id": 635310, "author_name": "hmendonca", "author_url": "", "post_date": "09/27/2019 11:10:35", "content": "@cherring sorry for the delay\nYes, it uses ArcNet: https://www.kaggle.com/hmendonca/fold1h4r3-arcenetb4-2-256px-rcic-lb-0-9759\nThe implementation I found from the Humpback comp worked really well for me, but others showed that it wasn't necessary though hehe
\n\nedit: I've added some more details and how I got to that solution here: https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110346
", "votes": null, "replies": [] } ] }, { "id": 633840, "author_name": "ryanzhang", "author_url": "", "post_date": "09/25/2019 13:32:37", "content": "I am doing a single model so far, I added the group as an additional input to be concatnated with the CNN features to be passed to the final output layer.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "633449": "Since the plates leak, I have been training four 277 class models and joining their predictions. In a couple recent posts I have seen high scoring people talking about training 1108 class models.\n\nCan anyone share their experience in the two different methods?", "633468": "In my experiment, 1108 class model is better than 277 class model. This may be because 1108 class model can use more information for training.", "633695": "Thanks @shimacos for the input\nWe never got the chance to check it, but my plan was initially to train a strong model with all data, and then fine tune it's predictions on a per cell type and plate basis. That would make it 4*4 (16) models.\nBut we're out of time and GPU quota now hehe\nIn fact, our current best submission (LB:0.94) is from a very single model, no blending", "633707": "Thanks for sharing :) Looks like I spent way too long trying and failing to get metric learning working.\n\nThat is an awesome score for a single model (I think, I can't wait to see everyone's solutions 😄 ), is that a metric learning model?", "633785": "hmendonca We tried to do that and it didn't work well​ :(", "633840": "I am doing a single model so far, I added the group as an additional input to be concatnated with the CNN features to be passed to the final output layer.", "633893": "igorkrashenyi Same here", "635310": "cherring sorry for the delay\nYes, it uses ArcNet: https://www.kaggle.com/hmendonca/fold1h4r3-arcenetb4-2-256px-rcic-lb-0-9759\nThe implementation I found from the Humpback comp worked really well for me, but others showed that it wasn't necessary though hehe\n\nedit: I've added some more details and how I got to that solution here: https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110346" }, "source": "meta" }