I've made some statistical features and developed RNN-model with these 1-D multi-array inputs. But, I couldn't get the ~0.7 single model.
\n\nSo, I made a 2D plot using spectogram or recur plot as input, and developed 2D-CNN model.\nBut it didn't work.
\n\nSo Difficult !!
\n\nI really want to know how to tackle this my situation.
\n\nI'm not begging the real solution. Just a piece of advice from kagglers.
\n\nPlease, help me. I really want to improve my ability at this signal processing competition. Thanks.!
", "messages": [ { "id": "472321", "postDate": "02/15/2019 17:30:42", "content": "I've made some statistical features and developed RNN-model with these 1-D multi-array inputs. But, I couldn't get the ~0.7 single model.
\n\nSo, I made a 2D plot using spectogram or recur plot as input, and developed 2D-CNN model.\nBut it didn't work.
\n\nSo Difficult !!
\n\nI really want to know how to tackle this my situation.
\n\nI'm not begging the real solution. Just a piece of advice from kagglers.
\n\nPlease, help me. I really want to improve my ability at this signal processing competition. Thanks.!
", "rawMarkdown": "I've made some statistical features and developed RNN-model with these 1-D multi-array inputs. But, I couldn't get the ~0.7 single model.\n\nSo, I made a 2D plot using spectogram or recur plot as input, and developed 2D-CNN model.\nBut it didn't work.\n\nSo Difficult !!\n\nI really want to know how to tackle this my situation. \n\nI'm not begging the real solution. Just a piece of advice from kagglers.\n\nPlease, help me. I really want to improve my ability at this signal processing competition. Thanks.!", "votes": null }, { "id": "472594", "postDate": "02/16/2019 08:39:02", "content": "Avoiding too much degrading Signal-to-Noise Ratio seems important here , though I'm not using NN so far.
", "rawMarkdown": "Avoiding too much degrading Signal-to-Noise Ratio seems important here , though I'm not using NN so far.", "votes": null }, { "id": "472682", "postDate": "02/16/2019 13:17:42", "content": "Thanks for sharing. If signal-to-noise ratio is important, statistical features and algorithm could be powerful as you said. Very helpful advice to me!
", "rawMarkdown": "Thanks for sharing. If signal-to-noise ratio is important, statistical features and algorithm could be powerful as you said. Very helpful advice to me!", "votes": null }, { "id": "472804", "postDate": "02/16/2019 17:34:29", "content": "Could you please explain it further?
", "rawMarkdown": "Could you please explain it further?", "votes": null }, { "id": "472825", "postDate": "02/16/2019 17:52:54", "content": "Signal-to-Noise Ratio is a ratio of the signal values to baseline signal value(simply, mean).\nIf partial charge exists, the signal value will be higher or lower than the baseline.
\n\nLet say, baseline is 10 and signal value with partial discharge is 20, then signal-to-noise ratio = 2.
\n\nIf we denoise or degrade signal too much, the signal value with partial discharge will be less, so, the signal-to-noise ratio will be less than 2.(1.5 or ~1)
\n\nIf the ratio is ~1, we can't distinguish the partial discharge from baseline.
\n\nIn summary, Too much flattening the signal would be harmful for this competition.\nThis is my understanding after reading @Tom advice.
\n\nIf something is wrong, Please correct me, Thanks!
", "rawMarkdown": "Signal-to-Noise Ratio is a ratio of the signal values to baseline signal value(simply, mean).\nIf partial charge exists, the signal value will be higher or lower than the baseline. \n\nLet say, baseline is 10 and signal value with partial discharge is 20, then signal-to-noise ratio = 2.\n\nIf we denoise or degrade signal too much, the signal value with partial discharge will be less, so, the signal-to-noise ratio will be less than 2.(1.5 or ~1) \n\nIf the ratio is ~1, we can't distinguish the partial discharge from baseline. \n\nIn summary, Too much flattening the signal would be harmful for this competition.\nThis is my understanding after reading @Tom advice.\n\nIf something is wrong, Please correct me, Thanks!", "votes": null }, { "id": "472834", "postDate": "02/16/2019 18:20:46", "content": "I was aware of SNR calculated as (mean / std) (analogous to the coefficient of variation).
\n\nThen, denoising or not, the presence of partial discharge patterns would increase more the std with respect to the mean and, therefore, degrade SNR.
\n\nI just cannot see why too much denoising would affect it
", "rawMarkdown": "I was aware of SNR calculated as (mean / std) (analogous to the coefficient of variation).\n\nThen, denoising or not, the presence of partial discharge patterns would increase more the std with respect to the mean and, therefore, degrade SNR.\n\nI just cannot see why too much denoising would affect it", "votes": null }, { "id": "472848", "postDate": "02/16/2019 18:55:45", "content": "@YouHan Lee, Is your RNN model stable ?
", "rawMarkdown": "YouHan Lee, Is your RNN model stable ?", "votes": null }, { "id": "472984", "postDate": "02/17/2019 03:04:36", "content": "I think it's not stable. I'm using keras for now. I've used RNN model from 5-fold LSTM Attention (fully commented).
", "rawMarkdown": "I think it's not stable. I'm using keras for now. I've used RNN model from 5-fold LSTM Attention (fully commented).", "votes": null }, { "id": "477950", "postDate": "02/25/2019 14:44:55", "content": "@harshit92, where you able to make you CuDNNLSTM stable on GPU by setting seeds ?
", "rawMarkdown": "harshit92, where you able to make you CuDNNLSTM stable on GPU by setting seeds ?", "votes": null }, { "id": "477987", "postDate": "02/25/2019 15:33:31", "content": "Antione, No it is giving different results even after fixing all the seeds. When GPU is off, then i am getting the same result.
", "rawMarkdown": "Antione, No it is giving different results even after fixing all the seeds. When GPU is off, then i am getting the same result.", "votes": null }, { "id": "478041", "postDate": "02/25/2019 16:37:15", "content": "I see Thanks ! But with CPU you can only use keras.layers.LSTM or keras.layers.GRU since CuDNNLSTM only works on GPU ?
", "rawMarkdown": "I see Thanks ! But with CPU you can only use keras.layers.LSTM or keras.layers.GRU since CuDNNLSTM only works on GPU ?", "votes": null }, { "id": "478042", "postDate": "02/25/2019 16:38:54", "content": "Yes, That is correct! And it will be slow for obvious reason.
", "rawMarkdown": "Yes, That is correct! And it will be slow for obvious reason.", "votes": null }, { "id": "478058", "postDate": "02/25/2019 17:11:12", "content": "So we have to choose between slowness or non-reproducibility :)
", "rawMarkdown": "So we have to choose between slowness or non-reproducibility :)", "votes": null }, { "id": "478077", "postDate": "02/25/2019 17:39:57", "content": "Yep! I am using randomness so far.
", "rawMarkdown": "Yep! I am using randomness so far.", "votes": null } ], "comments": [ { "id": 472594, "author_name": "tikutiku", "author_url": "", "post_date": "02/16/2019 08:39:02", "content": "Avoiding too much degrading Signal-to-Noise Ratio seems important here , though I'm not using NN so far.
", "votes": null, "replies": [ { "id": 472682, "author_name": "youhanlee", "author_url": "", "post_date": "02/16/2019 13:17:42", "content": "Thanks for sharing. If signal-to-noise ratio is important, statistical features and algorithm could be powerful as you said. Very helpful advice to me!
", "votes": null, "replies": [] }, { "id": 472804, "author_name": "fernandoramacciotti", "author_url": "", "post_date": "02/16/2019 17:34:29", "content": "Could you please explain it further?
", "votes": null, "replies": [] }, { "id": 472825, "author_name": "youhanlee", "author_url": "", "post_date": "02/16/2019 17:52:54", "content": "Signal-to-Noise Ratio is a ratio of the signal values to baseline signal value(simply, mean).\nIf partial charge exists, the signal value will be higher or lower than the baseline.
\n\nLet say, baseline is 10 and signal value with partial discharge is 20, then signal-to-noise ratio = 2.
\n\nIf we denoise or degrade signal too much, the signal value with partial discharge will be less, so, the signal-to-noise ratio will be less than 2.(1.5 or ~1)
\n\nIf the ratio is ~1, we can't distinguish the partial discharge from baseline.
\n\nIn summary, Too much flattening the signal would be harmful for this competition.\nThis is my understanding after reading @Tom advice.
\n\nIf something is wrong, Please correct me, Thanks!
", "votes": null, "replies": [] }, { "id": 472834, "author_name": "fernandoramacciotti", "author_url": "", "post_date": "02/16/2019 18:20:46", "content": "I was aware of SNR calculated as (mean / std) (analogous to the coefficient of variation).
\n\nThen, denoising or not, the presence of partial discharge patterns would increase more the std with respect to the mean and, therefore, degrade SNR.
\n\nI just cannot see why too much denoising would affect it
", "votes": null, "replies": [] } ] }, { "id": 472848, "author_name": "harshit92", "author_url": "", "post_date": "02/16/2019 18:55:45", "content": "@YouHan Lee, Is your RNN model stable ?
", "votes": null, "replies": [ { "id": 472984, "author_name": "youhanlee", "author_url": "", "post_date": "02/17/2019 03:04:36", "content": "I think it's not stable. I'm using keras for now. I've used RNN model from 5-fold LSTM Attention (fully commented).
", "votes": null, "replies": [] }, { "id": 477950, "author_name": "areveillon", "author_url": "", "post_date": "02/25/2019 14:44:55", "content": "@harshit92, where you able to make you CuDNNLSTM stable on GPU by setting seeds ?
", "votes": null, "replies": [] }, { "id": 477987, "author_name": "harshit92", "author_url": "", "post_date": "02/25/2019 15:33:31", "content": "Antione, No it is giving different results even after fixing all the seeds. When GPU is off, then i am getting the same result.
", "votes": null, "replies": [] }, { "id": 478041, "author_name": "areveillon", "author_url": "", "post_date": "02/25/2019 16:37:15", "content": "I see Thanks ! But with CPU you can only use keras.layers.LSTM or keras.layers.GRU since CuDNNLSTM only works on GPU ?
", "votes": null, "replies": [] }, { "id": 478042, "author_name": "harshit92", "author_url": "", "post_date": "02/25/2019 16:38:54", "content": "Yes, That is correct! And it will be slow for obvious reason.
", "votes": null, "replies": [] }, { "id": 478058, "author_name": "areveillon", "author_url": "", "post_date": "02/25/2019 17:11:12", "content": "So we have to choose between slowness or non-reproducibility :)
", "votes": null, "replies": [] }, { "id": 478077, "author_name": "harshit92", "author_url": "", "post_date": "02/25/2019 17:39:57", "content": "Yep! I am using randomness so far.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "472321": "I've made some statistical features and developed RNN-model with these 1-D multi-array inputs. But, I couldn't get the ~0.7 single model.\n\nSo, I made a 2D plot using spectogram or recur plot as input, and developed 2D-CNN model.\nBut it didn't work.\n\nSo Difficult !!\n\nI really want to know how to tackle this my situation. \n\nI'm not begging the real solution. Just a piece of advice from kagglers.\n\nPlease, help me. I really want to improve my ability at this signal processing competition. Thanks.!", "472594": "Avoiding too much degrading Signal-to-Noise Ratio seems important here , though I'm not using NN so far.", "472682": "Thanks for sharing. If signal-to-noise ratio is important, statistical features and algorithm could be powerful as you said. Very helpful advice to me!", "472804": "Could you please explain it further?", "472825": "Signal-to-Noise Ratio is a ratio of the signal values to baseline signal value(simply, mean).\nIf partial charge exists, the signal value will be higher or lower than the baseline. \n\nLet say, baseline is 10 and signal value with partial discharge is 20, then signal-to-noise ratio = 2.\n\nIf we denoise or degrade signal too much, the signal value with partial discharge will be less, so, the signal-to-noise ratio will be less than 2.(1.5 or ~1) \n\nIf the ratio is ~1, we can't distinguish the partial discharge from baseline. \n\nIn summary, Too much flattening the signal would be harmful for this competition.\nThis is my understanding after reading @Tom advice.\n\nIf something is wrong, Please correct me, Thanks!", "472834": "I was aware of SNR calculated as (mean / std) (analogous to the coefficient of variation).\n\nThen, denoising or not, the presence of partial discharge patterns would increase more the std with respect to the mean and, therefore, degrade SNR.\n\nI just cannot see why too much denoising would affect it", "472848": "YouHan Lee, Is your RNN model stable ?", "472984": "I think it's not stable. I'm using keras for now. I've used RNN model from 5-fold LSTM Attention (fully commented).", "477950": "harshit92, where you able to make you CuDNNLSTM stable on GPU by setting seeds ?", "477987": "Antione, No it is giving different results even after fixing all the seeds. When GPU is off, then i am getting the same result.", "478041": "I see Thanks ! But with CPU you can only use keras.layers.LSTM or keras.layers.GRU since CuDNNLSTM only works on GPU ?", "478042": "Yes, That is correct! And it will be slow for obvious reason.", "478058": "So we have to choose between slowness or non-reproducibility :)", "478077": "Yep! I am using randomness so far." }, "source": "meta" }