I think it is worthy to have a post focusing on the NN model and its performance, welcome to comment your best NN model so far and the performance.
", "messages": [ { "id": "412957", "postDate": "10/31/2018 03:34:11", "content": "I think it is worthy to have a post focusing on the NN model and its performance, welcome to comment your best NN model so far and the performance.
", "rawMarkdown": "I think it is worthy to have a post focusing on the NN model and its performance, welcome to comment your best NN model so far and the performance.", "votes": null }, { "id": "412977", "postDate": "10/31/2018 04:33:55", "content": "love that we're starting to explore neural nets; i wanted to try some NNs since the beginning but haven't had the time. so far i've stuck to 'single model' 5-f lgbm. i think i'm reaching my limit for what i can engineer there (lb 1.311). i'm starting to engineer features that take a day to generate and then only provide a minor boost. i'll be sure to post my performance as soon as i get my nn up and running.
", "rawMarkdown": "love that we're starting to explore neural nets; i wanted to try some NNs since the beginning but haven't had the time. so far i've stuck to 'single model' 5-f lgbm. i think i'm reaching my limit for what i can engineer there (lb 1.311). i'm starting to engineer features that take a day to generate and then only provide a minor boost. i'll be sure to post my performance as soon as i get my nn up and running.", "votes": null }, { "id": "413031", "postDate": "10/31/2018 07:13:04", "content": "I'm very interested in implementing the model used in this paper. The great thing about this model is that it trains an autoencoder for feature extraction, so the entire test set can be used in training as well. I am still having some trouble trying to figure out how to adapt the single-band model in the papar to the multi-band case we have in our dataset, as both imputation and sample pooling either lose too much information or leave too many NAs. Maybe I should try to directly apply to a single band first and see how it goes.
", "rawMarkdown": "I'm very interested in implementing the model used in [this paper](https://arxiv.org/pdf/1711.10609.pdf). The great thing about this model is that it trains an autoencoder for feature extraction, so the entire test set can be used in training as well. I am still having some trouble trying to figure out how to adapt the single-band model in the papar to the multi-band case we have in our dataset, as both imputation and sample pooling either lose too much information or leave too many NAs. Maybe I should try to directly apply to a single band first and see how it goes.", "votes": null }, { "id": "415785", "postDate": "11/05/2018 17:27:36", "content": "So far I found out, that very simple NN architectures work best for me.\nTried LSTM on light curve (with time relative from the first observation), but no luck so far.\nI plan to do the same with phase curves.
", "rawMarkdown": "So far I found out, that very simple NN architectures work best for me.\nTried LSTM on light curve (with time relative from the first observation), but no luck so far.\nI plan to do the same with phase curves.", "votes": null }, { "id": "416113", "postDate": "11/06/2018 07:59:57", "content": "Do you use keras?
", "rawMarkdown": "Do you use keras?", "votes": null }, { "id": "416608", "postDate": "11/07/2018 01:03:28", "content": "My code was based on the public kernel
", "rawMarkdown": "My code was based on the public kernel", "votes": null }, { "id": "416976", "postDate": "11/07/2018 14:57:42", "content": "it means keras?
", "rawMarkdown": "it means keras?", "votes": null }, { "id": "416980", "postDate": "11/07/2018 15:05:25", "content": "Yes
", "rawMarkdown": "Yes", "votes": null }, { "id": "418425", "postDate": "11/09/2018 22:06:19", "content": "How is your implementation?
", "rawMarkdown": "How is your implementation?", "votes": null }, { "id": "434923", "postDate": "12/07/2018 06:54:30", "content": "Did you hand craft the features and added them to the same model from the public kernel ?
", "rawMarkdown": "Did you hand craft the features and added them to the same model from the public kernel ?", "votes": null } ], "comments": [ { "id": 412977, "author_name": "yvanscher", "author_url": "", "post_date": "10/31/2018 04:33:55", "content": "love that we're starting to explore neural nets; i wanted to try some NNs since the beginning but haven't had the time. so far i've stuck to 'single model' 5-f lgbm. i think i'm reaching my limit for what i can engineer there (lb 1.311). i'm starting to engineer features that take a day to generate and then only provide a minor boost. i'll be sure to post my performance as soon as i get my nn up and running.
", "votes": null, "replies": [] }, { "id": 413031, "author_name": "mithrillion", "author_url": "", "post_date": "10/31/2018 07:13:04", "content": "I'm very interested in implementing the model used in this paper. The great thing about this model is that it trains an autoencoder for feature extraction, so the entire test set can be used in training as well. I am still having some trouble trying to figure out how to adapt the single-band model in the papar to the multi-band case we have in our dataset, as both imputation and sample pooling either lose too much information or leave too many NAs. Maybe I should try to directly apply to a single band first and see how it goes.
", "votes": null, "replies": [ { "id": 418425, "author_name": "strideradu", "author_url": "", "post_date": "11/09/2018 22:06:19", "content": "How is your implementation?
", "votes": null, "replies": [] } ] }, { "id": 415785, "author_name": "rejpalcz", "author_url": "", "post_date": "11/05/2018 17:27:36", "content": "So far I found out, that very simple NN architectures work best for me.\nTried LSTM on light curve (with time relative from the first observation), but no luck so far.\nI plan to do the same with phase curves.
", "votes": null, "replies": [] }, { "id": 416113, "author_name": "suvalex", "author_url": "", "post_date": "11/06/2018 07:59:57", "content": "Do you use keras?
", "votes": null, "replies": [ { "id": 416608, "author_name": "strideradu", "author_url": "", "post_date": "11/07/2018 01:03:28", "content": "My code was based on the public kernel
", "votes": null, "replies": [] }, { "id": 416976, "author_name": "suvalex", "author_url": "", "post_date": "11/07/2018 14:57:42", "content": "it means keras?
", "votes": null, "replies": [] }, { "id": 416980, "author_name": "strideradu", "author_url": "", "post_date": "11/07/2018 15:05:25", "content": "Yes
", "votes": null, "replies": [] } ] }, { "id": 434923, "author_name": "vignesh1905", "author_url": "", "post_date": "12/07/2018 06:54:30", "content": "Did you hand craft the features and added them to the same model from the public kernel ?
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "412957": "I think it is worthy to have a post focusing on the NN model and its performance, welcome to comment your best NN model so far and the performance.", "412977": "love that we're starting to explore neural nets; i wanted to try some NNs since the beginning but haven't had the time. so far i've stuck to 'single model' 5-f lgbm. i think i'm reaching my limit for what i can engineer there (lb 1.311). i'm starting to engineer features that take a day to generate and then only provide a minor boost. i'll be sure to post my performance as soon as i get my nn up and running.", "413031": "I'm very interested in implementing the model used in [this paper](https://arxiv.org/pdf/1711.10609.pdf). The great thing about this model is that it trains an autoencoder for feature extraction, so the entire test set can be used in training as well. I am still having some trouble trying to figure out how to adapt the single-band model in the papar to the multi-band case we have in our dataset, as both imputation and sample pooling either lose too much information or leave too many NAs. Maybe I should try to directly apply to a single band first and see how it goes.", "415785": "So far I found out, that very simple NN architectures work best for me.\nTried LSTM on light curve (with time relative from the first observation), but no luck so far.\nI plan to do the same with phase curves.", "416113": "Do you use keras?", "416608": "My code was based on the public kernel", "416976": "it means keras?", "416980": "Yes", "418425": "How is your implementation?", "434923": "Did you hand craft the features and added them to the same model from the public kernel ?" }, "source": "meta" }