{ "id": 47083, "title": "silence/non-silence classifier and 10-class classifier", "url": "/competitions/tensorflow-speech-recognition-challenge/discussion/47083", "author_name": "", "post_date": "2018-01-08T10:06:26.526686100Z", "votes": 6, "comment_count": 7, "views": 0, "content": "

This is an idea i haven't try. Anyone has results something like this? \n \"enter

", "messages": [ { "id": "266275", "postDate": "01/08/2018 10:06:26", "content": "

This is an idea i haven't try. Anyone has results something like this? \n \"enter

", "rawMarkdown": "This is an idea i haven't try. Anyone has results something like this? \n ![enter image description here][1]\n\n\n [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/266275/8225/net.png", "votes": null }, { "id": "266585", "postDate": "01/09/2018 05:09:59", "content": "

tried multi-task loss like this (binary classifier for silence or not, 30 class classifier for command word), very hard to converge on 1-d conv model (maybe lack of silence data), haven't tried on 2-d ones.

", "rawMarkdown": "tried multi-task loss like this (binary classifier for silence or not, 30 class classifier for command word), very hard to converge on 1-d conv model (maybe lack of silence data), haven't tried on 2-d ones.", "votes": null }, { "id": "266616", "postDate": "01/09/2018 07:09:21", "content": "

thanks for the infor!

", "rawMarkdown": "thanks for the infor!", "votes": null }, { "id": "267462", "postDate": "01/11/2018 12:33:30", "content": "

I'm also thinking of using similar idea. Had anyone had experience with this idea?

", "rawMarkdown": "I'm also thinking of using similar idea. Had anyone had experience with this idea?", "votes": null }, { "id": "267463", "postDate": "01/11/2018 12:38:14", "content": "

the data used to train the 2-class and 10-class can be different. You need to balance the multi-task loss also.\n(e.g. rpn loss and rcnn loss in faster rcnn)

", "rawMarkdown": "the data used to train the 2-class and 10-class can be different. You need to balance the multi-task loss also.\n(e.g. rpn loss and rcnn loss in faster rcnn)", "votes": null }, { "id": "267621", "postDate": "01/11/2018 22:15:24", "content": "

My first Kaggle, learning so much but restricted on time. Everyone is so nice on here, every time I get an idea on improving the model someone posts it up here before I can try it :) At least it shows I am thinking on the right track. This was one of my next things to try. Going to be a long week-end I think. Thanks.

", "rawMarkdown": "My first Kaggle, learning so much but restricted on time. Everyone is so nice on here, every time I get an idea on improving the model someone posts it up here before I can try it :) At least it shows I am thinking on the right track. This was one of my next things to try. Going to be a long week-end I think. Thanks.", "votes": null }, { "id": "268665", "postDate": "01/15/2018 07:00:07", "content": "

Thanks for your advice, I added weight to two kinds of loss, but after tries I just cannot manage to converge within acceptable training time so I just gave up.

", "rawMarkdown": "Thanks for your advice, I added weight to two kinds of loss, but after tries I just cannot manage to converge within acceptable training time so I just gave up.", "votes": null }, { "id": "268685", "postDate": "01/15/2018 07:49:03", "content": "

I used a similar approach to tackle unknown unknowns, trying to classify unknown vs. known instead of silence/non-silence as described here. I used a two-stage model (known vs unknown then classify 11 knowns including silence) and Contrastive Loss for the first stage, hoping that the latter would get me an edge with unknown-unknowns. Unfortunately, the first-stage model seems to overfit the unknowns, even though I could get a 94% accuracy on the training set. My overall model stuck at 84% and I moved on to better single-models.

\n\n

Inspiration from Medium article 1 and Research paper 2.

", "rawMarkdown": "I used a similar approach to tackle unknown unknowns, trying to classify unknown vs. known instead of silence/non-silence as described here. I used a two-stage model (known vs unknown then classify 11 knowns including silence) and Contrastive Loss for the first stage, hoping that the latter would get me an edge with unknown-unknowns. Unfortunately, the first-stage model seems to overfit the unknowns, even though I could get a 94% accuracy on the training set. My overall model stuck at 84% and I moved on to better single-models.\n\nInspiration from Medium article [1] and Research paper [2]. \n\n[1]: https://hackernoon.com/one-shot-learning-with-siamese-networks-in-pytorch-8ddaab10340e \"One Shot Learning with Siamese Networks in PyTorch\" \n[2]: https://ac.els-cdn.com/S1877050917318343/1-s2.0-S1877050917318343-main.pdf?_tid=9899f076-d45d-11e7-9f5f-00000aab0f02&acdnat=1511888591_b6d20f6714c43864f286b0ac735987a8 \"The Application of One-Class Classifier Based on CNN in Image Defect Detection\"", "votes": null } ], "comments": [ { "id": 266585, "author_name": "philipxue", "author_url": "", "post_date": "01/09/2018 05:09:59", "content": "

tried multi-task loss like this (binary classifier for silence or not, 30 class classifier for command word), very hard to converge on 1-d conv model (maybe lack of silence data), haven't tried on 2-d ones.

", "votes": null, "replies": [ { "id": 266616, "author_name": "hengck23", "author_url": "", "post_date": "01/09/2018 07:09:21", "content": "

thanks for the infor!

", "votes": null, "replies": [] }, { "id": 267463, "author_name": "hengck23", "author_url": "", "post_date": "01/11/2018 12:38:14", "content": "

the data used to train the 2-class and 10-class can be different. You need to balance the multi-task loss also.\n(e.g. rpn loss and rcnn loss in faster rcnn)

", "votes": null, "replies": [] }, { "id": 268665, "author_name": "philipxue", "author_url": "", "post_date": "01/15/2018 07:00:07", "content": "

Thanks for your advice, I added weight to two kinds of loss, but after tries I just cannot manage to converge within acceptable training time so I just gave up.

", "votes": null, "replies": [] } ] }, { "id": 267462, "author_name": "ori226", "author_url": "", "post_date": "01/11/2018 12:33:30", "content": "

I'm also thinking of using similar idea. Had anyone had experience with this idea?

", "votes": null, "replies": [] }, { "id": 267621, "author_name": "boltz0", "author_url": "", "post_date": "01/11/2018 22:15:24", "content": "

My first Kaggle, learning so much but restricted on time. Everyone is so nice on here, every time I get an idea on improving the model someone posts it up here before I can try it :) At least it shows I am thinking on the right track. This was one of my next things to try. Going to be a long week-end I think. Thanks.

", "votes": null, "replies": [] }, { "id": 268685, "author_name": "timothyman", "author_url": "", "post_date": "01/15/2018 07:49:03", "content": "

I used a similar approach to tackle unknown unknowns, trying to classify unknown vs. known instead of silence/non-silence as described here. I used a two-stage model (known vs unknown then classify 11 knowns including silence) and Contrastive Loss for the first stage, hoping that the latter would get me an edge with unknown-unknowns. Unfortunately, the first-stage model seems to overfit the unknowns, even though I could get a 94% accuracy on the training set. My overall model stuck at 84% and I moved on to better single-models.

\n\n

Inspiration from Medium article 1 and Research paper 2.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "266275": "This is an idea i haven't try. Anyone has results something like this? \n ![enter image description here][1]\n\n\n [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/266275/8225/net.png", "266585": "tried multi-task loss like this (binary classifier for silence or not, 30 class classifier for command word), very hard to converge on 1-d conv model (maybe lack of silence data), haven't tried on 2-d ones.", "266616": "thanks for the infor!", "267462": "I'm also thinking of using similar idea. Had anyone had experience with this idea?", "267463": "the data used to train the 2-class and 10-class can be different. You need to balance the multi-task loss also.\n(e.g. rpn loss and rcnn loss in faster rcnn)", "267621": "My first Kaggle, learning so much but restricted on time. Everyone is so nice on here, every time I get an idea on improving the model someone posts it up here before I can try it :) At least it shows I am thinking on the right track. This was one of my next things to try. Going to be a long week-end I think. Thanks.", "268665": "Thanks for your advice, I added weight to two kinds of loss, but after tries I just cannot manage to converge within acceptable training time so I just gave up.", "268685": "I used a similar approach to tackle unknown unknowns, trying to classify unknown vs. known instead of silence/non-silence as described here. I used a two-stage model (known vs unknown then classify 11 knowns including silence) and Contrastive Loss for the first stage, hoping that the latter would get me an edge with unknown-unknowns. Unfortunately, the first-stage model seems to overfit the unknowns, even though I could get a 94% accuracy on the training set. My overall model stuck at 84% and I moved on to better single-models.\n\nInspiration from Medium article [1] and Research paper [2]. \n\n[1]: https://hackernoon.com/one-shot-learning-with-siamese-networks-in-pytorch-8ddaab10340e \"One Shot Learning with Siamese Networks in PyTorch\" \n[2]: https://ac.els-cdn.com/S1877050917318343/1-s2.0-S1877050917318343-main.pdf?_tid=9899f076-d45d-11e7-9f5f-00000aab0f02&acdnat=1511888591_b6d20f6714c43864f286b0ac735987a8 \"The Application of One-Class Classifier Based on CNN in Image Defect Detection\"" }, "source": "meta" }