When i visited the repository of EfficientNet, i realized that many models have been trained with Noisy Student but they don't offer an example of how to do that. Then, i'm wondering the following, anyone had successfully implemented an Efficientnet model using Noisy Student? and if the answer it's yes, did improve your score?
\n\n", "messages": [ { "id": "950694", "postDate": "07/29/2020 14:53:53", "content": "When i visited the repository of EfficientNet, i realized that many models have been trained with Noisy Student but they don't offer an example of how to do that. Then, i'm wondering the following, anyone had successfully implemented an Efficientnet model using Noisy Student? and if the answer it's yes, did improve your score?
\n\n", "rawMarkdown": "When i visited the repository of EfficientNet, i realized that many models have been trained with **Noisy Student** but they don't offer an example of how to do that. Then, i'm wondering the following, anyone had successfully implemented an Efficientnet model using Noisy Student? and if the answer it's yes, did improve your score?\n\n- [The paper](https://arxiv.org/pdf/1911.04252.pdf)\n- [Video Explanation](https://www.youtube.com/watch?v=q7PjrmGNx5A)", "votes": null }, { "id": "950799", "postDate": "07/29/2020 16:20:07", "content": "The pre-trained weights are already available for noisy student.
\n\nIf you are using qubvel, you can call:\n\nimport efficientnet.keras as eff\nmodel = eff.EfficientNetB0(weights='noisy-student')\n
For Tensorflow, the pre-trained checkpoints can be downloaded from Tensorflow github.
", "rawMarkdown": "The pre-trained weights are already available for **noisy student**.\n\nIf you are using [qubvel](https://github.com/qubvel/efficientnet), you can call:\n```\nimport efficientnet.keras as eff\nmodel = eff.EfficientNetB0(weights='noisy-student')\n```\n\nFor Tensorflow, the pre-trained checkpoints can be downloaded from [Tensorflow github](https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet).", "votes": null }, { "id": "950805", "postDate": "07/29/2020 16:24:10", "content": "That's what i wanted thank you so much.
", "rawMarkdown": "That's what i wanted thank you so much.", "votes": null }, { "id": "951954", "postDate": "07/30/2020 14:16:30", "content": "Sirish gave the answer for how to use it, about your second question I'd say it increased my score, so yes...
", "rawMarkdown": "Sirish gave the answer for how to use it, about your second question I'd say it increased my score, so yes...", "votes": null }, { "id": "952007", "postDate": "07/30/2020 14:54:39", "content": "@datafan07 Abusing of your kindness i have one question, I used \"noisy-student\" but my score was lower than using \"imagenet\" weights, did my previous model overfit the data and therefore had a higher score?
", "rawMarkdown": "datafan07 Abusing of your kindness i have one question, I used \"noisy-student\" but my score was lower than using \"imagenet\" weights, did my previous model overfit the data and therefore had a higher score?", "votes": null }, { "id": "952055", "postDate": "07/30/2020 15:25:38", "content": "Maybe I wouldn't say if it overfits without seeing CV's but if you think it's overfitting try to reduce epochs?
", "rawMarkdown": "Maybe I wouldn't say if it overfits without seeing CV's but if you think it's overfitting try to reduce epochs?", "votes": null }, { "id": "954860", "postDate": "08/02/2020 06:37:40", "content": "They seem to have reported relatively higher accuracy with noisy-student weights rather than imagenet.
", "rawMarkdown": "They seem to have reported relatively higher accuracy with noisy-student weights rather than imagenet.", "votes": null }, { "id": "955204", "postDate": "08/02/2020 12:50:50", "content": "Not my case but probably it's because i'm still learning the \"properties\"(?) of EfficientNet.
", "rawMarkdown": "Not my case but probably it's because i'm still learning the \"properties\"(?) of EfficientNet.", "votes": null } ], "comments": [ { "id": 950799, "author_name": "sirishks", "author_url": "", "post_date": "07/29/2020 16:20:07", "content": "The pre-trained weights are already available for noisy student.
\n\nIf you are using qubvel, you can call:\n\nimport efficientnet.keras as eff\nmodel = eff.EfficientNetB0(weights='noisy-student')\n
For Tensorflow, the pre-trained checkpoints can be downloaded from Tensorflow github.
", "votes": null, "replies": [ { "id": 950805, "author_name": "hiramcho", "author_url": "", "post_date": "07/29/2020 16:24:10", "content": "That's what i wanted thank you so much.
", "votes": null, "replies": [] } ] }, { "id": 951954, "author_name": "datafan07", "author_url": "", "post_date": "07/30/2020 14:16:30", "content": "Sirish gave the answer for how to use it, about your second question I'd say it increased my score, so yes...
", "votes": null, "replies": [ { "id": 952007, "author_name": "hiramcho", "author_url": "", "post_date": "07/30/2020 14:54:39", "content": "@datafan07 Abusing of your kindness i have one question, I used \"noisy-student\" but my score was lower than using \"imagenet\" weights, did my previous model overfit the data and therefore had a higher score?
", "votes": null, "replies": [] }, { "id": 952055, "author_name": "datafan07", "author_url": "", "post_date": "07/30/2020 15:25:38", "content": "Maybe I wouldn't say if it overfits without seeing CV's but if you think it's overfitting try to reduce epochs?
", "votes": null, "replies": [] } ] }, { "id": 954860, "author_name": "fireheart7", "author_url": "", "post_date": "08/02/2020 06:37:40", "content": "They seem to have reported relatively higher accuracy with noisy-student weights rather than imagenet.
", "votes": null, "replies": [ { "id": 955204, "author_name": "hiramcho", "author_url": "", "post_date": "08/02/2020 12:50:50", "content": "Not my case but probably it's because i'm still learning the \"properties\"(?) of EfficientNet.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "950694": "When i visited the repository of EfficientNet, i realized that many models have been trained with **Noisy Student** but they don't offer an example of how to do that. Then, i'm wondering the following, anyone had successfully implemented an Efficientnet model using Noisy Student? and if the answer it's yes, did improve your score?\n\n- [The paper](https://arxiv.org/pdf/1911.04252.pdf)\n- [Video Explanation](https://www.youtube.com/watch?v=q7PjrmGNx5A)", "950799": "The pre-trained weights are already available for **noisy student**.\n\nIf you are using [qubvel](https://github.com/qubvel/efficientnet), you can call:\n```\nimport efficientnet.keras as eff\nmodel = eff.EfficientNetB0(weights='noisy-student')\n```\n\nFor Tensorflow, the pre-trained checkpoints can be downloaded from [Tensorflow github](https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet).", "950805": "That's what i wanted thank you so much.", "951954": "Sirish gave the answer for how to use it, about your second question I'd say it increased my score, so yes...", "952007": "datafan07 Abusing of your kindness i have one question, I used \"noisy-student\" but my score was lower than using \"imagenet\" weights, did my previous model overfit the data and therefore had a higher score?", "952055": "Maybe I wouldn't say if it overfits without seeing CV's but if you think it's overfitting try to reduce epochs?", "954860": "They seem to have reported relatively higher accuracy with noisy-student weights rather than imagenet.", "955204": "Not my case but probably it's because i'm still learning the \"properties\"(?) of EfficientNet." }, "source": "meta" }