one can use UMAP or slower t-SNE to make embedding visualization, e.g. https://github.com/lmcinnes/umap
\n\nhttps://experiments.withgoogle.com/t-sne-map\nhttps://projector.tensorflow.org/
", "messages": [ { "id": 732486, "postDate": "2020-01-29T21:46:04.830Z", "content": "one can use UMAP or slower t-SNE to make embedding visualization, e.g. https://github.com/lmcinnes/umap
\n\nhttps://experiments.withgoogle.com/t-sne-map\nhttps://projector.tensorflow.org/
", "rawMarkdown": "one can use UMAP or slower t-SNE to make embedding visualization, e.g. https://github.com/lmcinnes/umap\n\nhttps://experiments.withgoogle.com/t-sne-map\nhttps://projector.tensorflow.org/", "votes": 2 }, { "id": 732703, "postDate": "2020-01-30T06:03:01.067Z", "content": "I was thinking about using PCA as well in addition to those two, not for visualization but for speeding up training by reducing the size of the dataset.
\n\nHas anybody tried that?
", "rawMarkdown": "I was thinking about using PCA as well in addition to those two, not for visualization but for speeding up training by reducing the size of the dataset.\n\nHas anybody tried that?", "replies": [ { "id": 732898, "postDate": "2020-01-30T12:45:32.333Z", "content": "I've tried Umap but it's too slow. :( Even though I set the image size to 64 * 64, it is slow and hard in the kaggle notebook.
", "rawMarkdown": "I've tried Umap but it's too slow. :( Even though I set the image size to 64 * 64, it is slow and hard in the kaggle notebook." }, { "id": 733488, "postDate": "2020-01-31T07:43:44.810Z", "content": "Yeah I would imagine it would be quite long on image datasets. Maybe it’s worth doing it locally if you have a better CPU?\nOr on a stratified subset of the data?
", "rawMarkdown": "Yeah I would imagine it would be quite long on image datasets. Maybe it’s worth doing it locally if you have a better CPU?\nOr on a stratified subset of the data? " }, { "id": 733499, "postDate": "2020-01-31T08:14:19.860Z", "content": "\"I've tried Umap but it's too slow. \"
\n\nyou can just subsample a couple of train images, instead of using all
", "rawMarkdown": "\"I've tried Umap but it's too slow. \"\n\nyou can just subsample a couple of train images, instead of using all" }, { "id": 736428, "postDate": "2020-02-04T07:04:19.910Z", "content": "Well, PCA should be fast, and some time ago it was used in CV competitions as preprocessing of inputs. And these images contain a lot of kinda useless information (like background).\nI guess I will try PCA, maybe Kernel PCA, time it to see if it's possible to run in kernel.\nAlso there is CUDA version of t-SNE tha should run faster than sklearn implementation, opentsne or fft-tsne.
", "rawMarkdown": "Well, PCA should be fast, and some time ago it was used in CV competitions as preprocessing of inputs. And these images contain a lot of kinda useless information (like background).\nI guess I will try PCA, maybe Kernel PCA, time it to see if it's possible to run in kernel.\nAlso there is CUDA version of t-SNE tha should run faster than sklearn implementation, opentsne or fft-tsne." } ] }, { "id": 738448, "postDate": "2020-02-06T14:33:41.950Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 732703, "author_name": "Maxime Lenormand", "author_url": "", "post_date": "2020-01-30T06:03:01.067000", "content": "I was thinking about using PCA as well in addition to those two, not for visualization but for speeding up training by reducing the size of the dataset.
\n\nHas anybody tried that?
", "votes": 0, "replies": [ { "id": 732898, "author_name": "Subin An", "author_url": "", "post_date": "2020-01-30T12:45:32.333000", "content": "I've tried Umap but it's too slow. :( Even though I set the image size to 64 * 64, it is slow and hard in the kaggle notebook.
", "votes": 0, "replies": [] }, { "id": 733488, "author_name": "Maxime Lenormand", "author_url": "", "post_date": "2020-01-31T07:43:44.810000", "content": "Yeah I would imagine it would be quite long on image datasets. Maybe it’s worth doing it locally if you have a better CPU?\nOr on a stratified subset of the data?
", "votes": 0, "replies": [] }, { "id": 733499, "author_name": "hengck23", "author_url": "", "post_date": "2020-01-31T08:14:19.860000", "content": "\"I've tried Umap but it's too slow. \"
\n\nyou can just subsample a couple of train images, instead of using all
", "votes": 0, "replies": [] }, { "id": 736428, "author_name": "Hleb Levitski", "author_url": "", "post_date": "2020-02-04T07:04:19.910000", "content": "Well, PCA should be fast, and some time ago it was used in CV competitions as preprocessing of inputs. And these images contain a lot of kinda useless information (like background).\nI guess I will try PCA, maybe Kernel PCA, time it to see if it's possible to run in kernel.\nAlso there is CUDA version of t-SNE tha should run faster than sklearn implementation, opentsne or fft-tsne.
", "votes": 0, "replies": [] } ] }, { "id": 738448, "author_name": "", "author_url": "", "post_date": "2020-02-06T14:33:41.950000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "732486": "one can use UMAP or slower t-SNE to make embedding visualization, e.g. https://github.com/lmcinnes/umap\n\nhttps://experiments.withgoogle.com/t-sne-map\nhttps://projector.tensorflow.org/", "732703": "I was thinking about using PCA as well in addition to those two, not for visualization but for speeding up training by reducing the size of the dataset.\n\nHas anybody tried that?", "738448": "" } }