6th place solution is a neural network consisting of 3 components:
\n\nMeta Encoder - taking as input meta features hostgal photoz, hostgal photoz err, is galactic and summary stats for flux and flux_err as min/max/mean etc.
Light Curve Encoder - bidirectional GRU taking as input grouped by day flux, flux_err, detected and time difference
Spectroscopic Redshift Predictor – two fully-connected layers for predicting hostgal specz
Outputs of these 3 components are fed into two last fully-connected layers for predicting class probabilities.
\n\nLight curve encoder was pre-trained as autoencoder on test data. Spectroscopic redshift predictor was also pre-trained on test data objects for which hostgal specz is available.
\n\nAugmentations. Three types of augmentations are performed:
\n\nTop submission is an average of 5 cross-validation runs of 3 neural network variations.
", "messages": [ { "id": "441059", "postDate": "12/18/2018 08:15:12", "content": "6th place solution is a neural network consisting of 3 components:
\n\nMeta Encoder - taking as input meta features hostgal photoz, hostgal photoz err, is galactic and summary stats for flux and flux_err as min/max/mean etc.
Light Curve Encoder - bidirectional GRU taking as input grouped by day flux, flux_err, detected and time difference
Spectroscopic Redshift Predictor – two fully-connected layers for predicting hostgal specz
Outputs of these 3 components are fed into two last fully-connected layers for predicting class probabilities.
\n\nLight curve encoder was pre-trained as autoencoder on test data. Spectroscopic redshift predictor was also pre-trained on test data objects for which hostgal specz is available.
\n\nAugmentations. Three types of augmentations are performed:
\n\nTop submission is an average of 5 cross-validation runs of 3 neural network variations.
", "rawMarkdown": "6th place solution is a neural network consisting of 3 components:\n\n1. Meta Encoder - taking as input meta features hostgal photoz, hostgal photoz err, is galactic and summary stats for flux and flux_err as min/max/mean etc.\n\n2. Light Curve Encoder - bidirectional GRU taking as input grouped by day flux, flux_err, detected and time difference\n\n3. Spectroscopic Redshift Predictor – two fully-connected layers for predicting hostgal specz\n\nOutputs of these 3 components are fed into two last fully-connected layers for predicting class probabilities.\n\nLight curve encoder was pre-trained as autoencoder on test data. Spectroscopic redshift predictor was also pre-trained on test data objects for which hostgal specz is available.\n\nAugmentations. Three types of augmentations are performed:\n\n1. flux as a Gaussian with standard deviation flux_err\n2. hostgal photoz as a Gaussian with standard deviation hostgal photoz err\n3. randomly dropping observations\n\nTop submission is an average of 5 cross-validation runs of 3 neural network variations.", "votes": null }, { "id": "441067", "postDate": "12/18/2018 08:20:56", "content": "Very interesting, thanks for sharing, and congrats on the result. Third good solution with specz predictor, we clearly missed that one. Do you intend to share your code?
", "rawMarkdown": "Very interesting, thanks for sharing, and congrats on the result. Third good solution with specz predictor, we clearly missed that one. Do you intend to share your code?", "votes": null }, { "id": "441068", "postDate": "12/18/2018 08:22:25", "content": "Congrats! Your solution is pretty efficient! Haven't tried adding a spectroscopic redshift predictor in training time, good idea!
", "rawMarkdown": "Congrats! Your solution is pretty efficient! Haven't tried adding a spectroscopic redshift predictor in training time, good idea!", "votes": null }, { "id": "441070", "postDate": "12/18/2018 08:23:11", "content": "Seems we noticed the same thing at almost the same time :)
", "rawMarkdown": "Seems we noticed the same thing at almost the same time :)", "votes": null }, { "id": "441072", "postDate": "12/18/2018 08:26:26", "content": "Awesome solution. Congratulations for the 6th position. \nEvery time I saw leader board I wondered about your low number of submissions and very high score. It's clear now, you have a solid approach.
", "rawMarkdown": "Awesome solution. Congratulations for the 6th position. \nEvery time I saw leader board I wondered about your low number of submissions and very high score. It's clear now, you have a solid approach.", "votes": null }, { "id": "441093", "postDate": "12/18/2018 08:54:39", "content": "Very interesting work! Kudos
", "rawMarkdown": "Very interesting work! Kudos", "votes": null }, { "id": "441148", "postDate": "12/18/2018 10:28:25", "content": "Cool! I tried to use NN as predictor and encoder,but I was failed :( \nDo you use Pytorch? Keras? TensorFlow?\nWill you show your code?
", "rawMarkdown": "Cool! I tried to use NN as predictor and encoder,but I was failed :( \nDo you use Pytorch? Keras? TensorFlow?\nWill you show your code?", "votes": null }, { "id": "441189", "postDate": "12/18/2018 11:39:28", "content": "Congrats Stefan, clean and nice solution.
", "rawMarkdown": "Congrats Stefan, clean and nice solution.", "votes": null }, { "id": "441375", "postDate": "12/18/2018 15:43:38", "content": "Congratulations and Thanks for sharing your solution. \nIt's amazing that experts like you need only few submissions to reach up to the top. I tried to squeeze even the last bit of what was left in my already-beaten-to-death model by submitting as much as possible, and I've been lingering below the bronze medalists all along.
", "rawMarkdown": "Congratulations and Thanks for sharing your solution. \nIt's amazing that experts like you need only few submissions to reach up to the top. I tried to squeeze even the last bit of what was left in my already-beaten-to-death model by submitting as much as possible, and I've been lingering below the bronze medalists all along.", "votes": null }, { "id": "441434", "postDate": "12/18/2018 16:53:27", "content": "Thank you all! I used PyTorch and skorch for this competition. The code needs some serious efforts before becoming publishable and usable for others so not sure whether and when I will be able to do this.
", "rawMarkdown": "Thank you all! I used PyTorch and skorch for this competition. The code needs some serious efforts before becoming publishable and usable for others so not sure whether and when I will be able to do this.", "votes": null }, { "id": "441850", "postDate": "12/19/2018 06:50:58", "content": "Congratulations Stefan ! Interesting solution, thanks for sharing :)
", "rawMarkdown": "Congratulations Stefan ! Interesting solution, thanks for sharing :)", "votes": null } ], "comments": [ { "id": 441067, "author_name": "cpmpml", "author_url": "", "post_date": "12/18/2018 08:20:56", "content": "Very interesting, thanks for sharing, and congrats on the result. Third good solution with specz predictor, we clearly missed that one. Do you intend to share your code?
", "votes": null, "replies": [ { "id": 441070, "author_name": "khyeh0719", "author_url": "", "post_date": "12/18/2018 08:23:11", "content": "Seems we noticed the same thing at almost the same time :)
", "votes": null, "replies": [] } ] }, { "id": 441068, "author_name": "khyeh0719", "author_url": "", "post_date": "12/18/2018 08:22:25", "content": "Congrats! Your solution is pretty efficient! Haven't tried adding a spectroscopic redshift predictor in training time, good idea!
", "votes": null, "replies": [] }, { "id": 441072, "author_name": "subrahmanyamv", "author_url": "", "post_date": "12/18/2018 08:26:26", "content": "Awesome solution. Congratulations for the 6th position. \nEvery time I saw leader board I wondered about your low number of submissions and very high score. It's clear now, you have a solid approach.
", "votes": null, "replies": [] }, { "id": 441093, "author_name": "malyaban", "author_url": "", "post_date": "12/18/2018 08:54:39", "content": "Very interesting work! Kudos
", "votes": null, "replies": [] }, { "id": 441148, "author_name": "suvalex", "author_url": "", "post_date": "12/18/2018 10:28:25", "content": "Cool! I tried to use NN as predictor and encoder,but I was failed :( \nDo you use Pytorch? Keras? TensorFlow?\nWill you show your code?
", "votes": null, "replies": [] }, { "id": 441189, "author_name": "titericz", "author_url": "", "post_date": "12/18/2018 11:39:28", "content": "Congrats Stefan, clean and nice solution.
", "votes": null, "replies": [] }, { "id": 441375, "author_name": "vignam", "author_url": "", "post_date": "12/18/2018 15:43:38", "content": "Congratulations and Thanks for sharing your solution. \nIt's amazing that experts like you need only few submissions to reach up to the top. I tried to squeeze even the last bit of what was left in my already-beaten-to-death model by submitting as much as possible, and I've been lingering below the bronze medalists all along.
", "votes": null, "replies": [] }, { "id": 441434, "author_name": "stefanstefanov", "author_url": "", "post_date": "12/18/2018 16:53:27", "content": "Thank you all! I used PyTorch and skorch for this competition. The code needs some serious efforts before becoming publishable and usable for others so not sure whether and when I will be able to do this.
", "votes": null, "replies": [] }, { "id": 441850, "author_name": "ogrellier", "author_url": "", "post_date": "12/19/2018 06:50:58", "content": "Congratulations Stefan ! Interesting solution, thanks for sharing :)
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "441059": "6th place solution is a neural network consisting of 3 components:\n\n1. Meta Encoder - taking as input meta features hostgal photoz, hostgal photoz err, is galactic and summary stats for flux and flux_err as min/max/mean etc.\n\n2. Light Curve Encoder - bidirectional GRU taking as input grouped by day flux, flux_err, detected and time difference\n\n3. Spectroscopic Redshift Predictor – two fully-connected layers for predicting hostgal specz\n\nOutputs of these 3 components are fed into two last fully-connected layers for predicting class probabilities.\n\nLight curve encoder was pre-trained as autoencoder on test data. Spectroscopic redshift predictor was also pre-trained on test data objects for which hostgal specz is available.\n\nAugmentations. Three types of augmentations are performed:\n\n1. flux as a Gaussian with standard deviation flux_err\n2. hostgal photoz as a Gaussian with standard deviation hostgal photoz err\n3. randomly dropping observations\n\nTop submission is an average of 5 cross-validation runs of 3 neural network variations.", "441067": "Very interesting, thanks for sharing, and congrats on the result. Third good solution with specz predictor, we clearly missed that one. Do you intend to share your code?", "441068": "Congrats! Your solution is pretty efficient! Haven't tried adding a spectroscopic redshift predictor in training time, good idea!", "441070": "Seems we noticed the same thing at almost the same time :)", "441072": "Awesome solution. Congratulations for the 6th position. \nEvery time I saw leader board I wondered about your low number of submissions and very high score. It's clear now, you have a solid approach.", "441093": "Very interesting work! Kudos", "441148": "Cool! I tried to use NN as predictor and encoder,but I was failed :( \nDo you use Pytorch? Keras? TensorFlow?\nWill you show your code?", "441189": "Congrats Stefan, clean and nice solution.", "441375": "Congratulations and Thanks for sharing your solution. \nIt's amazing that experts like you need only few submissions to reach up to the top. I tried to squeeze even the last bit of what was left in my already-beaten-to-death model by submitting as much as possible, and I've been lingering below the bronze medalists all along.", "441434": "Thank you all! I used PyTorch and skorch for this competition. The code needs some serious efforts before becoming publishable and usable for others so not sure whether and when I will be able to do this.", "441850": "Congratulations Stefan ! Interesting solution, thanks for sharing :)" }, "source": "meta" }