As we can see from different EDA kernels it comes natural to display the signals as an image, which enables us to use various techniques from computer vision (e.g. augmentation). Anyone used pretrained image models already?
", "messages": [ { "id": "447192", "postDate": "12/29/2018 10:35:26", "content": "As we can see from different EDA kernels it comes natural to display the signals as an image, which enables us to use various techniques from computer vision (e.g. augmentation). Anyone used pretrained image models already?
", "rawMarkdown": "As we can see from different EDA kernels it comes natural to display the signals as an image, which enables us to use various techniques from computer vision (e.g. augmentation). Anyone used pretrained image models already?", "votes": null }, { "id": "448177", "postDate": "12/31/2018 11:44:35", "content": "pretrained image models won't work, because the nature of the problems is so different. You can keep the first two or three layers maybe.
", "rawMarkdown": "pretrained image models won't work, because the nature of the problems is so different. You can keep the first two or three layers maybe.", "votes": null }, { "id": "448239", "postDate": "12/31/2018 15:01:48", "content": "I am not so sure. I experienced in multiple competitions that pertained networks also work for very different problems. I might give it a go with saving the matplotlib pictures of the eda kernels and run a simple reset with them
", "rawMarkdown": "I am not so sure. I experienced in multiple competitions that pertained networks also work for very different problems. I might give it a go with saving the matplotlib pictures of the eda kernels and run a simple reset with them", "votes": null }, { "id": "454000", "postDate": "01/11/2019 02:37:58", "content": "I think is a good idea to convert the data into images, the experts probably look at images for fault detection. In other fields tracing were converted to images and computer vision techniques were used. This is an example not for fault but for fraud detection:\nhttps://www.splunk.com/blog/2017/04/18/deep-learning-with-splunk-and-tensorflow-for-security-catching-the-fraudster-in-neural-networks-with-behavioral-biometrics.html
", "rawMarkdown": "I think is a good idea to convert the data into images, the experts probably look at images for fault detection. In other fields tracing were converted to images and computer vision techniques were used. This is an example not for fault but for fraud detection:\nhttps://www.splunk.com/blog/2017/04/18/deep-learning-with-splunk-and-tensorflow-for-security-catching-the-fraudster-in-neural-networks-with-behavioral-biometrics.html", "votes": null }, { "id": "455740", "postDate": "01/14/2019 13:23:38", "content": "i tried this with plots and using mobilenetv2 and couldnt get any results not sure if i was doing something wrong as ive never done image classification before
", "rawMarkdown": "i tried this with plots and using mobilenetv2 and couldnt get any results not sure if i was doing something wrong as ive never done image classification before", "votes": null }, { "id": "457813", "postDate": "01/18/2019 05:47:19", "content": "Hi. I had no luck with pictures so far. I tried an rfft/freq (hist) plot, a picture of the smoothed or filtered curves or a waterfall diagramm showing the rfft/freq hist over time. It seems stat values over time are the way to go. Be it with LSTM or Conv1D. Cheers.
", "rawMarkdown": "Hi. I had no luck with pictures so far. I tried an rfft/freq (hist) plot, a picture of the smoothed or filtered curves or a waterfall diagramm showing the rfft/freq hist over time. It seems stat values over time are the way to go. Be it with LSTM or Conv1D. Cheers.", "votes": null }, { "id": "458957", "postDate": "01/20/2019 22:14:14", "content": "Image processing isn't the approach I would take but I'm genuinely interested to see if you make any headway with it.
", "rawMarkdown": "Image processing isn't the approach I would take but I'm genuinely interested to see if you make any headway with it.", "votes": null } ], "comments": [ { "id": 448177, "author_name": "zoujie", "author_url": "", "post_date": "12/31/2018 11:44:35", "content": "pretrained image models won't work, because the nature of the problems is so different. You can keep the first two or three layers maybe.
", "votes": null, "replies": [ { "id": 448239, "author_name": "christofhenkel", "author_url": "", "post_date": "12/31/2018 15:01:48", "content": "I am not so sure. I experienced in multiple competitions that pertained networks also work for very different problems. I might give it a go with saving the matplotlib pictures of the eda kernels and run a simple reset with them
", "votes": null, "replies": [] } ] }, { "id": 454000, "author_name": "agentili", "author_url": "", "post_date": "01/11/2019 02:37:58", "content": "I think is a good idea to convert the data into images, the experts probably look at images for fault detection. In other fields tracing were converted to images and computer vision techniques were used. This is an example not for fault but for fraud detection:\nhttps://www.splunk.com/blog/2017/04/18/deep-learning-with-splunk-and-tensorflow-for-security-catching-the-fraudster-in-neural-networks-with-behavioral-biometrics.html
", "votes": null, "replies": [] }, { "id": 455740, "author_name": "mab270", "author_url": "", "post_date": "01/14/2019 13:23:38", "content": "i tried this with plots and using mobilenetv2 and couldnt get any results not sure if i was doing something wrong as ive never done image classification before
", "votes": null, "replies": [] }, { "id": 457813, "author_name": "mightybird", "author_url": "", "post_date": "01/18/2019 05:47:19", "content": "Hi. I had no luck with pictures so far. I tried an rfft/freq (hist) plot, a picture of the smoothed or filtered curves or a waterfall diagramm showing the rfft/freq hist over time. It seems stat values over time are the way to go. Be it with LSTM or Conv1D. Cheers.
", "votes": null, "replies": [] }, { "id": 458957, "author_name": "jeffreyegan", "author_url": "", "post_date": "01/20/2019 22:14:14", "content": "Image processing isn't the approach I would take but I'm genuinely interested to see if you make any headway with it.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "447192": "As we can see from different EDA kernels it comes natural to display the signals as an image, which enables us to use various techniques from computer vision (e.g. augmentation). Anyone used pretrained image models already?", "448177": "pretrained image models won't work, because the nature of the problems is so different. You can keep the first two or three layers maybe.", "448239": "I am not so sure. I experienced in multiple competitions that pertained networks also work for very different problems. I might give it a go with saving the matplotlib pictures of the eda kernels and run a simple reset with them", "454000": "I think is a good idea to convert the data into images, the experts probably look at images for fault detection. In other fields tracing were converted to images and computer vision techniques were used. This is an example not for fault but for fraud detection:\nhttps://www.splunk.com/blog/2017/04/18/deep-learning-with-splunk-and-tensorflow-for-security-catching-the-fraudster-in-neural-networks-with-behavioral-biometrics.html", "455740": "i tried this with plots and using mobilenetv2 and couldnt get any results not sure if i was doing something wrong as ive never done image classification before", "457813": "Hi. I had no luck with pictures so far. I tried an rfft/freq (hist) plot, a picture of the smoothed or filtered curves or a waterfall diagramm showing the rfft/freq hist over time. It seems stat values over time are the way to go. Be it with LSTM or Conv1D. Cheers.", "458957": "Image processing isn't the approach I would take but I'm genuinely interested to see if you make any headway with it." }, "source": "meta" }