![\"image\"](\"https://i.postimg.cc/X7vsbKJh/image.png\")
The second iteration of the famous EfficientNet model family has been released. The authors claim huge improvements.
\nCompared to EfficientNet and more recent works, our EfficientNetV2 trains up to 11x faster while being up to 6.8x smaller.\nThis image from the paper sums it up quite nicely (model tagged with 21K are pretrained on ImageNet21k)
\n
A detailed comparison between the V1 EfficientNetB7 model and the V2 M model show the potential improvements, especially the training time reduction is fascinating.
\n![\"image2\"](\"https://i.postimg.cc/NjGhH2dv/image2.png\")
The paper can be found here and a Pytorch implementation is on its way. A Tensorflow implementation should be released here.
\nAs this competition is hugely computationally intensive, these models could cause a breakthrough for the encoders.
", "messages": [ { "id": "1272197", "postDate": "04/13/2021 09:55:50", "content": "The second iteration of the famous EfficientNet model family has been released. The authors claim huge improvements.
\nCompared to EfficientNet and more recent works, our EfficientNetV2 trains up to 11x faster while being up to 6.8x smaller.\nThis image from the paper sums it up quite nicely (model tagged with 21K are pretrained on ImageNet21k)
\n
A detailed comparison between the V1 EfficientNetB7 model and the V2 M model show the potential improvements, especially the training time reduction is fascinating.
\n
The paper can be found here and a Pytorch implementation is on its way. A Tensorflow implementation should be released here.
\nAs this competition is hugely computationally intensive, these models could cause a breakthrough for the encoders.
", "rawMarkdown": "The second iteration of the famous EfficientNet model family has been released. The authors claim huge improvements.\n\n```\nCompared to EfficientNet and more recent works, our EfficientNetV2 trains up to 11x faster while being up to 6.8x smaller.\n```\n\nThis image from the paper sums it up quite nicely (model tagged with 21K are pretrained on ImageNet21k)\n\n\n\nA detailed comparison between the V1 EfficientNetB7 model and the V2 M model show the potential improvements, especially the training time reduction is fascinating.\n\n\n\nThe paper can be found [here](https://arxiv.org/pdf/2104.00298.pdf) and a [Pytorch implementation](https://github.com/lukemelas/EfficientNet-PyTorch) is on its way. A Tensorflow implementation should be released [here](https://github.com/google/automl/tree/master/efficientnetv2).\n\nAs this competition is hugely computationally intensive, these models could cause a breakthrough for the encoders.", "votes": null }, { "id": "1272268", "postDate": "04/13/2021 11:18:27", "content": "this is interesting, from the paper:
\n\"4.2. Progressive Learning with adaptive Regularization\"
\nThen, we gradually increase image size but also
\nmaking learning more difficult by adding stronger regularization. Our approach is built upon (Howard, 2018) that
\nprogressively changes image size, but here we adaptively
\nadjust regularization as well.
Waiting for tensorflow/keras model to come-up sooner 🌈
", "rawMarkdown": "Waiting for tensorflow/keras model to come-up sooner 🌈", "votes": null }, { "id": "1283712", "postDate": "04/25/2021 07:43:43", "content": "release early so that we can use it . appreciate.
", "rawMarkdown": "release early so that we can use it . appreciate.", "votes": null } ], "comments": [ { "id": 1272268, "author_name": "hengck23", "author_url": "", "post_date": "04/13/2021 11:18:27", "content": "this is interesting, from the paper:
\n\"4.2. Progressive Learning with adaptive Regularization\"
\nThen, we gradually increase image size but also
\nmaking learning more difficult by adding stronger regularization. Our approach is built upon (Howard, 2018) that
\nprogressively changes image size, but here we adaptively
\nadjust regularization as well.
Waiting for tensorflow/keras model to come-up sooner 🌈
", "votes": null, "replies": [] }, { "id": 1283712, "author_name": "dragonzhang", "author_url": "", "post_date": "04/25/2021 07:43:43", "content": "release early so that we can use it . appreciate.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1272197": "The second iteration of the famous EfficientNet model family has been released. The authors claim huge improvements.\n\n```\nCompared to EfficientNet and more recent works, our EfficientNetV2 trains up to 11x faster while being up to 6.8x smaller.\n```\n\nThis image from the paper sums it up quite nicely (model tagged with 21K are pretrained on ImageNet21k)\n\n\n\nA detailed comparison between the V1 EfficientNetB7 model and the V2 M model show the potential improvements, especially the training time reduction is fascinating.\n\n\n\nThe paper can be found [here](https://arxiv.org/pdf/2104.00298.pdf) and a [Pytorch implementation](https://github.com/lukemelas/EfficientNet-PyTorch) is on its way. A Tensorflow implementation should be released [here](https://github.com/google/automl/tree/master/efficientnetv2).\n\nAs this competition is hugely computationally intensive, these models could cause a breakthrough for the encoders.", "1272268": "this is interesting, from the paper:\n\n\"4.2. Progressive Learning with adaptive Regularization\"\n\nThen, we gradually increase image size but also\nmaking learning more difficult by adding stronger regularization. Our approach is built upon (Howard, 2018) that\nprogressively changes image size, but here we adaptively\nadjust regularization as well.", "1280492": "Waiting for tensorflow/keras model to come-up sooner 🌈", "1283712": "release early so that we can use it . appreciate." }, "source": "meta" }